Verifier-free Test-Time Sampling for Vision Language Action Models

🎯 Pitch

A robot’s own uncertainty about what to do next—measured by how much its action predictions shift when it forgets its sensor input and instructions—turns out to be a surprisingly powerful substitute for a separately trained verifier. By simply picking the action with the smallest internal distribution change among random samples, this method boosts real-world manipulation success by over 30% and quadruples precision on low-data pick-and-place, all without extra training.

1. Executive Summary

This paper introduces Masking Distribution Guided Selection (MG-Select), a verifier-free test-time scaling framework for vision-language-action models that selects the most confident action from multiple candidates by measuring KL divergence against a reference distribution generated by masking the same model's input conditions (state, language instruction, or both). Evaluated on π0-FAST across RoboCasa, SIMPLER-WidowX, LIBERO, and real-world Franka pick-and-place tasks, MG-Select achieves a 28% relative improvement on in-distribution real-world tasks, a 35% improvement on out-of-distribution tasks, and a 168% relative gain on RoboCasa pick-and-place tasks trained with only 30 demonstrations—establishing that self-generated confidence signals can replace external verifiers for test-time action selection without additional training, provided the base model can be jointly fine-tuned to learn both conditional and condition-masked distributions.

2. Context and Motivation

The Core Problem: VLAs Fail When They Can Only Guess Once

The fundamental problem this paper addresses is that vision-language-action models (VLAs), despite their impressive performance on robot control tasks, are constitutionally bottlenecked by their single-inference paradigm. An autoregressive VLA like π0-FAST (Pertsch et al., 2025) produces actions by sampling from a learned probability distribution over action tokens — and in standard deployment, it does this exactly once per timestep, selecting the most probable action sequence (greedy decoding). If that single prediction is slightly off — a grasp point a few millimeters too high, a release location slightly misaligned — the entire task fails.

What makes this a genuine technical problem rather than a trivial "robots aren't perfect" observation is that these failures are often not due to a lack of knowledge. As the paper notes, prior work (Nakamoto et al., 2024; Kwok et al., 2025) shows that VLAs can achieve high precision with adequate training — the model's internal representations contain the information needed for success. The bottleneck is the decoding strategy: the model's probability distribution over action tokens may assign non-trivial mass to the correct action, but greedy decoding forces it to commit to the single most probable token at every step, which can drift off-target cumulatively across a multi-token action sequence.

The paper frames this as a precision gap (Section 1):

"Even after extensive pre-training, they often fail on fine-grained manipulation tasks such as grasping or object placement... This precision gap is particularly problematic for real-world robotic applications where millimeter-level accuracy can determine task success or failure."

This precision gap manifests most starkly in pick-and-place tasks, which require two high-stakes contact points — the grasp and the release — each of which demands spatial accuracy. In simulation, a model might succeed 43% of the time on pick-and-place tasks with 300 demonstrations (Table 1, RoboCasa base model), meaning it fails on the majority of attempts despite having seen hundreds of expert demonstrations. The paper's motivating intuition is that these failures are partly avoidable — the model sometimes "knows" the right answer but the greedy decoding procedure doesn't reliably surface it.

Why Simple Solutions Don't Work

If the problem is that greedy decoding picks a single action that may be suboptimal, the obvious fix is to sample multiple actions and pick the best one. This is Best-of-N sampling, a well-established test-time compute strategy in language models (Brown et al., 2024). The simplest selection criterion is likelihood: among N sampled action sequences, pick the one the model assigns the highest probability.

The paper reports that this naïve approach partially works — but only partially. Table 5(a) shows that likelihood-based Best-of-N with jointly trained models does improve over greedy decoding (30.5% vs. 28.5% on RoboCasa pick-and-place with 100 demos). However, the paper identifies a fundamental reason why likelihood maximization is insufficient in the VLA setting:

"VLAs fine-tuned on target tasks for next action token prediction often memorize expert trajectories, causing the probability distribution over action tokens to become overly concentrated, which leads to multiple sampling converging to the same result."

This is a crucial observation. When a VLA is fine-tuned via behavior cloning on expert demonstrations, it learns to reproduce those specific trajectories with high confidence. The resulting probability distribution over action tokens becomes peaked — the model places near-certain probability on the expert action at each step. Sampling multiple times from such a concentrated distribution produces actions that are nearly identical, negating the benefit of generating candidates in the first place. You cannot select a better action if all candidates are the same mediocre action.

This differentiates the VLA setting from the LLM setting where Best-of-N has proven effective. In LLM reasoning tasks, the model's distribution over possible completions is inherently broad — there are many ways to write a solution to a math problem. In VLA action prediction, the expert trajectory is typically unique or nearly so, and behavior cloning encourages collapse onto that single trajectory. The diversity that Best-of-N relies on is squeezed out during fine-tuning.

The Existing Solution and Its Limitations: External Verifiers

Prior work has attempted to address the precision gap through test-time scaling with external verifiers — a paradigm the paper explicitly acknowledges as inspiration (Section 1). Two primary approaches are cited:

Value-guided selection (Nakamoto et al., 2024). This approach trains a separate value function using offline reinforcement learning on diverse robotic datasets. At inference time, multiple candidate actions are sampled from the VLA, and the value function scores each candidate — the action with the highest predicted value (expected task success) is executed. The value function acts as an external critic that can recognize good actions even when the VLA's own probability distribution is miscalibrated.

VLM-based action verification (Kwok et al., 2025). This method trains a verifier using reward modeling on synthetically generated preference datasets. A vision-language model (VLM) is trained to compare candidate actions and predict which will be more successful, effectively serving as a learned reward model for action selection.

Both approaches demonstrate that test-time scaling can improve VLA precision — the core strategy of generating multiple candidates and selecting among them is valid. However, the paper identifies two critical failures of the verifier-based paradigm that motivate the search for an alternative:

1. Additional training burden. External verifiers require training a second model — a value function or reward model — using reinforcement learning or preference optimization objectives. This adds "substantial computational overhead and complexity to the deployment pipeline" (Section 1). In practice, this means maintaining two separate model training pipelines (one for the VLA, one for the verifier), two sets of hyperparameters, and potentially two different datasets. For robotics practitioners who may already be struggling to train VLAs effectively on small demonstration datasets, this additional complexity is a genuine barrier.

2. Failure to generalize to unseen conditions. Perhaps more damningly, the paper identifies that external verifiers do not generalize:

"These external verifiers fail to generalize to unseen input conditions (Nakamoto et al., 2024), such as novel task prompts or objects, and their reward modeling is tailored to specific datasets, severely limiting their broader applicability (Kwok et al., 2025)."

This is a fundamental limitation rooted in the nature of value functions and reward models. A value function trained on a specific dataset of tasks and objects learns to predict success for that specific distribution. When the robot encounters a novel object — say, a roll of tape instead of a cube — or a novel task instruction — "pick up the tape" instead of "pick up the cube" — the value function's predictions become unreliable because it has never seen similar states. The verifier, which is supposed to be the reliable arbiter of action quality, becomes the weakest link precisely when it is most needed: under distribution shift.

This is particularly problematic for real-world deployment. Robots operating in unstructured environments will inevitably encounter novel objects, lighting conditions, and task variations. A test-time scaling approach that collapses under these conditions provides no robustness benefit over the base model — and may actively harm performance if it confidently selects a bad action based on a miscalibrated verifier score.

The Paper's Positioning: Verifier-Free Test-Time Scaling

The authors position MG-Select as a third path — one that retains the benefits of test-time scaling (generating multiple candidates and selecting among them) while eliminating the dependency on external verifiers entirely. The research goal is stated clearly:

"To develop a test-time scaling framework for VLAs that leverages the model's internal properties without requiring additional training or external modules."

The key phrase is "the model's internal properties." Rather than training a separate critic to evaluate actions, MG-Select uses signals that already exist within the VLA itself: specifically, the divergence between the model's action token distribution when conditioned on the full task specification and its distribution when certain conditions (state, instruction, or both) are masked away.

The intellectual lineage is explicitly drawn from two sources:

Self-certainty in LLMs (Kang et al., 2025). In the language domain, "self-certainty" measures how much a model's output distribution changes when context is perturbed — a large change indicates high confidence (the model relies heavily on the specific context), while a small change indicates low confidence (the model would produce similar output regardless of context). MG-Select adapts this principle: an action that depends strongly on the task conditions (high KL divergence from the masked distribution) is likely to be a deliberate, task-specific action, while an action that changes little when conditions are masked is likely to be a generic, uncertain action.

Verifier-free TTS (Zheng et al., 2024). The broader idea of test-time scaling without external verifiers is motivated by work in the LLM space showing that internal model signals — such as output consistency or self-evaluation — can substitute for trained verifiers in Best-of-N selection.

What distinguishes MG-Select from these LLM approaches is the specific form of the reference distribution. In language models, a uniform distribution over tokens is a natural baseline for "no information." But in VLAs, a uniform distribution over action tokens has no relationship to the robot's task or capabilities — it would assign equal probability to "grasp at position (0.1, 0.2, 0.3)" and "grasp at (-100, -100, -100)," which is clearly not a meaningful uncertainty baseline. The paper's key insight is that a more informative reference distribution can be obtained by partially masking the model's own inputs — producing a distribution that is uncertain (because key information is missing) but still grounded in the task distribution (because the model architecture and training still shape its output space).

The Joint Training Requirement and Why It Matters

An important subtlety in the paper's positioning is that while MG-Select is described as requiring "no additional training" in the sense of no external verifier training, the method does benefit substantially from a joint training strategy (Section 3.3) that enables the VLA to learn condition-masked distributions alongside the standard conditional distribution. The paper is careful to distinguish:

Without joint training: MG-Select can be applied to any autoregressive VLA by simply masking inputs at test time and computing KL divergence. This is the "plug-and-play" version. Table 1 shows it provides improvements (e.g., 17.0% → 22.6% on RoboCasa pick-and-place with 100 demos), but the gains are moderate.
*With joint training (MG-Select)**: The VLA is fine-tuned with randomly dropped conditions (text, state, or both) during training, so it learns to produce meaningful condition-masked distributions. Table 1 shows this amplifies gains substantially (17.0% → 31.0%), and the paper's headline results use this variant.

This creates a nuanced positioning: MG-Select does not require training a separate model (verifier), but it does benefit from a modified training procedure for the VLA itself. The computational overhead is the same as standard fine-tuning — the same model, same optimizer, same dataset, just with condition dropout applied. This is a meaningfully different cost profile from training a separate value function with reinforcement learning, but it is not "no additional training" in the strongest sense. The paper implicitly acknowledges this by distinguishing "MG-Select" (no joint training) from "MG-Select*" (with joint training) in all result tables.

The Specific Failure Modes the Method Targets

To understand the motivation, it helps to visualize what failure looks like in practice and how MG-Select proposes to fix it. The paper provides qualitative insight through Figure 2, which shows a real-world "Box to Bowl" task:

Base model failure: The π0-FAST-DROID model correctly moves the gripper toward the object but fails at the critical contact points — either the grasp misses the sponge by a small margin, or the release places the sponge beside the bowl rather than inside it. These are not catastrophic failures (the robot doesn't wildly flail) but precision failures: the action is in the right neighborhood but not accurate enough.
MG-Select success: With the same base model but using MG-Select for action selection, the robot grasps the sponge cleanly and releases it into the bowl. The difference is that MG-Select sampled multiple candidate action sequences and selected the one with the highest confidence signal — the one that maximally depended on having the full task specification available.

The paper's core hypothesis is that high-confidence actions (those that change substantially when conditions are masked) are more likely to be precise, task-specific actions, while low-confidence actions (those that look similar whether or not the model knows what task it's doing) are likely to be generic, imprecise actions that happen to have high likelihood under the memorized distribution. This is a testable claim, and the paper's experimental results — particularly the consistent improvements across diverse environments — provide evidence for it.

Why This Problem Matters Now

The paper enters a research landscape where VLAs are rapidly becoming the default architecture for generalist robot policies (Driess et al., 2023; Zitkovich et al., 2023; Kim et al., 2024; Black et al., 2025; Pertsch et al., 2025; Bjorck et al., 2025), driven by the availability of large-scale robot demonstration datasets (O'Neill et al., 2024; Khazatsky et al., 2024; Bu et al., 2025). As these models improve in generality — handling more tasks, more objects, more environments — the precision problem becomes more acute, not less, because:

Generalist models are harder to fine-tune to perfection. A model pre-trained on thousands of diverse tasks has broad capabilities but may not be expert-level on any single task. Post-training fine-tuning on target-task demonstrations helps, but as the paper observes, this fine-tuning can cause distribution collapse (over-concentration), making the model brittle.
Real-world deployment demands robustness to novel conditions. The OOD experiments (Table 3) test exactly this — can the model handle objects it wasn't fine-tuned on? External verifiers fail here because they're trained on the same distribution as the VLA. MG-Select's approach of using internal signals doesn't have this limitation — the confidence metric is computed from the VLA's own outputs, so it inherits whatever generalization the VLA possesses.
Data is expensive in robotics. The paper's strongest results are in the low-data regime: 168% relative improvement on RoboCasa with only 30 demonstrations, and consistent gains across real-world tasks trained with only 60 demonstrations. This matters because collecting robot demonstrations requires physical hardware, human supervision, and time — methods that extract more from limited data have immediate practical value for robotics labs and companies.

The paper thus addresses a convergence of trends: VLAs are becoming the dominant paradigm, test-time compute is proving valuable across AI domains, but the specific tools developed for language (external verifiers, likelihood-based selection) don't transfer cleanly to action prediction. MG-Select offers a robotics-native approach that exploits the structure of VLA action distributions — their sensitivity to input conditions — to achieve test-time scaling without the fragility of learned verifiers.

3. Technical Approach

3.1 Reader Orientation

MG-Select is a test-time action selection system that runs on top of any autoregressive vision-language-action model (VLA) — at each timestep, instead of taking the single most probable action, it generates multiple candidate actions in parallel, then picks the best one using a confidence score computed entirely from the VLA's own internal probability distributions without training any external critic or verifier.

The system solves the precision-gap problem — where VLAs fail on fine-grained manipulation tasks despite possessing the underlying capability — by recognizing that actions which change dramatically when the model's input conditions (state, instruction) are masked away are more likely to be deliberate, task-specific, and precise, while actions that remain similar under masking are generic and imprecise.

3.2 Big-Picture Architecture (Diagram in Words)

The MG-Select pipeline has four major components connected in a single forward pass at each timestep:

Condition-Masked Forward Pass: The VLA $\pi_\theta$ is run multiple times in parallel — once with the full task specification (observation $o_t$ , proprioceptive state $q_t$ , instruction $I$ ) to generate candidate action sequences, and additionally with one or more masked conditions (e.g., state removed, instruction removed, or both removed) to produce reference distributions over action tokens.
Candidate Action Sampler: Using the full-condition distribution as the proposal, $N$ complete action sequences are sampled stochastically (with temperature $\tau > 0$ ), producing the candidate set $\tilde{\mathcal{A}} = \{\tilde{a}^{(n)}\}_{n=1}^N$ where each candidate is a variable-length sequence of discrete action tokens.
Token-Level Confidence Computer: For each candidate action, at each token position $i$ , the KL divergence is computed between the condition-masked distribution $Q_i$ (the reference) and the full-condition distribution $P_i$ (the prediction). This produces a scalar confidence score per token: $C_i = \text{KL}(Q_i \parallel P_i)$ .
Action-Level Aggregator and Selector: The per-token confidence scores are aggregated across the action sequence (e.g., by summing the first 5 tokens for the FAST tokenizer, or averaging across all tokens for OpenVLA) to produce a single confidence score $C_{\tilde{a}}$ for each candidate action. The action with the highest aggregated confidence is selected: $a^* = \arg\max_{\tilde{a}^{(n)} \in \tilde{\mathcal{A}}} C_{\tilde{a}^{(n)}}$ .

Information flows sequentially: observation and instruction enter → the VLA computes both full-condition and masked-condition token distributions in parallel → $N$ candidates are drawn from the full-condition distribution → per-token KL divergences are computed between each masked reference and the full distribution → token confidences are aggregated per candidate → the argmax selects the final action to execute on the robot.

3.3 Roadmap for the Deep Dive

First, the test-time scaling framework — how candidate actions are sampled, what Best-of-N selection means in this context, and what properties the selection criterion must satisfy. This establishes the outer loop that MG-Select plugs into.
Second, the condition-masking distributional confidence metric — why likelihood-based selection fails for VLAs, how the reference distribution is constructed by masking specific input modalities, and why KL divergence from a masked reference captures meaningful action confidence. This is the core intellectual contribution.
Third, the three masking variants (text-masking, state-masking, both-masking) — what each variant captures about uncertainty, how they produce different reference distributions, and why the optimal variant depends on the task environment.
Fourth, the aggregation strategy — how per-token KL divergences are combined into a single action-level confidence score, and why naive summation is suboptimal due to the structure of the FAST tokenizer.
Fifth, the regularization temperature — why the raw condition-masked distribution fails as a reference (it can be peaked, undermining the purpose of measuring divergence from uncertainty), and how applying a high temperature flattens it into a useful uncertainty baseline.
Sixth, the joint training strategy — why plug-and-play masking of a standard VLA produces poor reference distributions, and how training with random condition dropout enables the model to learn both conditional and condition-masked distributions simultaneously, amplifying the gains from MG-Select.

3.4 Detailed, Sentence-Based Technical Breakdown

This is primarily a method paper that proposes a new test-time action selection mechanism for autoregressive VLAs. The core idea is that the divergence between a VLA's full-condition action token distribution and its condition-masked distribution serves as a self-generated confidence signal that can replace external verifiers in Best-of-N sampling, provided the reference distribution represents maximum uncertainty while remaining aligned with the task distribution.

The Test-Time Scaling Framework

At each timestep $t$ , given the current observation $o_t$ , proprioceptive state $q_t$ , and language instruction $I$ , the standard VLA deployment samples exactly one action sequence from $\pi_\theta(a \mid o_t, q_t, I)$ — typically greedily, i.e., selecting the highest-probability token at each decoding step. MG-Select replaces this single-sample paradigm with a two-stage procedure: generate multiple candidates, then select among them.

Stage 1: Parallel stochastic sampling of $N$ candidates.

The autoregressive VLA $\pi_\theta$ factorizes each action sequence into discrete tokens $a = (a_1, a_2, \ldots, a_T)$ where $T$ varies per candidate. To generate diverse candidates, the model samples at temperature $\tau > 0$ rather than greedily:

$\tilde{a}^{(n)}_j \sim \pi_\theta(\cdot \mid o_t, q_t, I, \tilde{a}^{(n)}_{<j}; \tau), \quad n = 1, \ldots, N, \quad j = 1, \ldots, T_n$

where $\tilde{a}^{(n)}_j$ is the $j$ -th token of the $n$ -th candidate action sequence, $T_n$ is that candidate's variable sequence length, and $\pi_\theta(\cdot; \tau) = \text{softmax}(\ell / \tau)$ scales the logit vector $\ell \in \mathbb{R}^{|V|}$ by temperature $\tau$ before conversion to a probability distribution over the action token vocabulary $V$ .

What it computes: For each of the $N$ candidates, a complete action sequence (a variable-length list of discrete tokens) is drawn from the VLA's full-condition token-level distribution with temperature-scaled softmax. The temperature controls the sharpness-diversity tradeoff: as $\tau \to 0$ , sampling approaches greedy (all mass on the max token, zero diversity); as $\tau \to \infty$ , the distribution approaches uniform (maximum diversity, but tokens become random). The output is the candidate set $\tilde{\mathcal{A}} = \{\tilde{a}^{(n)}\}_{n=1}^N$ .

Why this form: Temperature sampling is used rather than top-k or nucleus sampling because the goal is to maintain diversity in the full continuous action space while staying within the model's learned distribution. The paper sweeps $\tau \in \{0.1, 0.3, 0.5, 0.7, 1.0\}$ (Appendix A.3) and finds $\tau = 0.5$ works well as a default — high enough to produce diverse candidates for pick-and-place tasks but low enough that candidates remain plausible actions. The key practical consideration is parallel batch inference: all $N$ candidates are generated simultaneously by leveraging the VLA's ability to process batched inputs, so the sampling stage adds latency proportional to the longest candidate sequence rather than $N \times$ the single-candidate latency.

Stage 2: Best-of-N selection via criterion $\mathcal{M}$ .

Once the candidate set is generated, the system selects the final action according to a pre-defined scoring metric:

$a^* = \arg\max_{\tilde{a}^{(n)} \in \tilde{\mathcal{A}}} \mathcal{M}_{\tilde{a}^{(n)}}$

where $\mathcal{M}_{\tilde{a}^{(n)}}$ is a scalar score computed for candidate $\tilde{a}^{(n)}$ . The specific form of $\mathcal{M}$ is the paper's contribution — it is not likelihood, not a uniform-KL baseline, but the condition-masking distributional confidence described below. The framework itself (sample N, pick best) is standard; the innovation is entirely in the choice of $\mathcal{M}$ .

Why Best-of-N and not beam search or tree search: The paper does not explore more sophisticated search algorithms because the action space is continuous and tokenized — the number of possible action sequences is combinatorially large, and the VLA's probability distribution is already concentrated around expert-like trajectories. Beam search would likely amplify the over-concentration problem (as it does in LLM reasoning — see the reference example's discussion of verifier over-optimization), while Best-of-N with stochastic sampling provides controlled diversity.

Why Likelihood-Based Selection Fails for VLAs

Before presenting the proposed confidence metric, the paper articulates why the most obvious candidate scoring method — the model's own sequence likelihood — is insufficient. This analysis is essential for understanding the design of MG-Select.

For an autoregressive model, the likelihood (or log-likelihood) of a candidate action sequence $\tilde{a}$ under the full-condition distribution is:

$\log \pi_\theta(\tilde{a} \mid o_t, q_t, I) = \sum_{j=1}^{T} \log \pi_\theta(\tilde{a}_j \mid o_t, q_t, I, \tilde{a}_{<j})$

What it computes: The sum of log-probabilities assigned by the model to each token in the sequence, given the observation, state, and instruction. Higher values mean the model "prefers" this action sequence under its learned distribution.

Why it fails: The paper identifies that VLAs fine-tuned via behavior cloning on expert demonstrations tend to produce overly concentrated probability distributions. When a model is trained to maximize $\log \pi_\theta(a_{\text{expert}} \mid o, q, I)$ across many demonstrations of the same or similar tasks, it learns to assign near-certain probability to the expert action tokens at each step. The distribution becomes "peaked" rather than spread across plausible alternatives. When sampling from such a concentrated distribution with moderate temperature, the $N$ candidates are nearly identical — they all converge to the same memorized expert trajectory. If that trajectory is imprecise (e.g., the expert grasped slightly off-center in the demonstrations), all candidates will share the same imprecision, and likelihood-based selection cannot distinguish among them.

This is fundamentally different from the LLM setting where Best-of-N via likelihood works. In mathematical reasoning, there are typically many valid solution paths, and the model's distribution reflects this diversity. In VLA action prediction, the behavior cloning objective actively suppresses diversity by training the model to reproduce exactly one action per state. The paper's Table 5(a) confirms this empirically: likelihood-based Best-of-N (row "Likelihood") improves over greedy decoding (30.5% vs. 28.5% on RoboCasa pick-and-place with 100 demos), but the gain is modest compared to MG-Select (31.0%).

Condition-Masking Distributional Confidence: The Core Metric

The paper's central idea is to replace likelihood with a confidence metric based on distributional divergence. The intuition is: an action that the model would NOT have produced without knowing the specific task conditions is a "deliberate" action — the model needed the information to generate it, implying the information was used in a meaningful way. Conversely, an action that the model produces similarly whether or not it knows the task is a "generic" action — it emerges from the model's default behavior rather than from task-specific reasoning.

Token-level confidence definition.

For a single token position $i$ in an action sequence, let $P_i$ be the predicted distribution (full-condition) and $Q_i$ be a reference distribution (condition-masked). The token-level confidence is:

$C_i = \text{KL}(Q_i \parallel P_i) = \sum_{v \in V} Q_i(v) \log \frac{Q_i(v)}{P_i(v)}$

where $V$ is the action token vocabulary, $Q_i(v)$ is the probability assigned to token $v$ under the masked reference distribution, and $P_i(v)$ is the probability under the full-condition distribution.

What it computes: The Kullback-Leibler divergence from the reference $Q_i$ to the predicted $P_i$ . This measures how much additional information (in bits, if using log base 2) is needed to encode samples from $Q_i$ using a code optimized for $P_i$ . When $P_i$ and $Q_i$ are identical, $\text{KL} = 0$ — masking conditions had no effect, so the model's prediction at this token doesn't depend on the masked information. When they diverge strongly, the KL is large — the model's prediction changed substantially, meaning it relied on the masked condition.

Why KL divergence and not other divergence measures: KL divergence is asymmetric — $\text{KL}(Q \parallel P)$ penalizes cases where $Q$ assigns high probability to tokens that $P$ assigns low probability (i.e., the reference distribution "expects" certain tokens that the full-condition distribution rules out). This asymmetry is important: we want to detect when masking causes the model to consider tokens that the full-condition model rejects, which indicates that the full-condition prediction is strongly conditioned on the masked information. Symmetric measures like Jensen-Shannon divergence would not distinguish between "masking introduces new tokens" and "masking removes tokens," which are not equally informative about confidence.

Why KL from $Q$ to $P$ and not $P$ to $Q$ : $\text{KL}(P \parallel Q)$ would measure how much the full-condition distribution diverges from the masked reference. This is less useful as a confidence metric because it would be large whenever $P$ is concentrated (which is always true for behavior-cloned VLAs), regardless of whether that concentration reflects task-specific information or just memorization. The chosen direction $\text{KL}(Q \parallel P)$ explicitly measures the effect of removing condition information, isolating the component of the prediction that is attributable to the masked conditions.

Action-level aggregation.

Token-level confidences are aggregated across the action sequence to produce a single score per candidate:

$C_{\tilde{a}} = \sum_{i \in \mathcal{I}} C_i = \sum_{i \in \mathcal{I}} \text{KL}(Q_i \parallel P_i)$

where $\mathcal{I} \subseteq \{1, 2, \ldots, T\}$ selects which token positions to include. The choice of $\mathcal{I}$ depends on the tokenization scheme (detailed in the Aggregation Strategy subsection below).

What it computes: A scalar confidence score for the entire action sequence $\tilde{a}$ , obtained by summing (or averaging) the token-level KL divergences over a selected subset of token positions. This score is used directly as $\mathcal{M}_{\tilde{a}}$ in the Best-of-N selection.

Why summation over selected tokens and not the full sequence: The paper finds that truncating to the first few tokens (specifically, the first 5 for the FAST tokenizer) works best. This is because the FAST tokenizer (Pertsch et al., 2025) encodes actions from low-frequency (coarse) to high-frequency (fine) components — the early tokens capture the overall movement direction and grasp/release decisions, while later tokens encode fine adjustments. High-confidence coarse actions are more predictive of task success than high-confidence fine adjustments, and including too many fine-grained tokens introduces noise from the tokenizer's frequency decomposition.

Constructing the Reference Distribution via Condition-Masking

The reference distribution $Q$ must satisfy two competing desiderata: it should represent maximum uncertainty (so that divergence from it genuinely measures confidence), but it should remain aligned with the task distribution (so that divergence is meaningful — a uniform distribution over all possible action tokens would produce large KL divergence for any VLA prediction, but would not distinguish task-relevant from generic actions because the uniform distribution has no relationship to what actions are physically possible or robot-typical).

The paper's solution is to generate $Q$ by running the same VLA $\pi_\theta$ with specific input conditions masked. This ensures $Q$ remains within the VLA's learned action manifold (the model can only produce action token distributions it was trained to produce), but the masking removes the information needed to solve the specific task, pushing the distribution toward higher entropy while keeping it grounded.

Three masking variants are defined:

(1) Text-masking (instruction removed):

$Q_i^{\text{text}} = \pi_\theta(\cdot \mid o_t, q_t, \emptyset, a_{<i})$

where the language instruction $I$ is replaced with an empty token $\emptyset$ . The model sees the visual observation and proprioceptive state but does not know what task to perform (e.g., it sees a ketchup bottle and a cabinet but doesn't know whether to pick, place, open, or close).

The corresponding token-level confidence:

$\text{KL}_{\text{text}} = \text{KL}\left(\pi_\theta(\cdot \mid o_t, q_t, \emptyset, a_{<i}) \;\parallel\; \pi_\theta(\cdot \mid o_t, q_t, I, a_{<i})\right)$

What it captures: Dependence on the language instruction. A large $\text{KL}_{\text{text}}$ means the model's action token prediction changes substantially when the task description is removed — the action is instruction-specific (e.g., "pick" vs. "place"). A small value means the model would produce a similar action regardless of what the instruction says — the action is instruction-agnostic (e.g., moving toward an object regardless of what to do with it).

When it's optimal: In task-diverse environments like RoboCasa (24 tasks including pick-and-place, open-and-close, and others), the model cannot determine the correct action without the instruction, so text-masking produces a genuinely uncertain reference. Table 5(c) confirms text-masking achieves the best performance on RoboCasa (31.0% on pick-and-place, compared to 30.1% for state-masking and 29.7% for both-masking).

(2) State-masking (proprioceptive state removed):

$Q_i^{\text{state}} = \pi_\theta(\cdot \mid o_t, \emptyset, I, a_{<i})$

where the proprioceptive state $q_t$ (joint angles, gripper position, etc.) is replaced with an empty token. The model sees the visual observation and knows the task instruction, but does not know the robot's current configuration.

The corresponding token-level confidence:

$\text{KL}_{\text{state}} = \text{KL}\left(\pi_\theta(\cdot \mid o_t, \emptyset, I, a_{<i}) \;\parallel\; \pi_\theta(\cdot \mid o_t, q_t, I, a_{<i})\right)$

What it captures: Dependence on the robot's current physical state. A large $\text{KL}_{\text{state}}$ means the model relies heavily on knowing where the gripper is to decide what action to take next — the action is state-dependent (e.g., "close gripper now because the fingers are around the object"). A small value means the model would produce similar actions based on visual input and task alone — the action is state-agnostic.

When it's optimal: In environments with homogeneous tasks, where the model already knows the general action pattern (always pick and place) and precision depends on adapting to the current gripper position. SIMPLER-WidowX consists solely of pick-and-place tasks — the model doesn't need the instruction to know what to do, only where precisely to do it. Appendix A.1 confirms that for SIMPLER-WidowX, only state-masking is used during joint training (10% dropout rate for state, no text or both masking).

(3) Text & State masking (both removed):

$Q_i^{\text{both}} = \pi_\theta(\cdot \mid o_t, \emptyset, \emptyset, a_{<i})$

where both the instruction and proprioceptive state are replaced with empty tokens. The model sees only the visual observation — it knows what objects are present but has no task specification and no knowledge of its own configuration.

The corresponding token-level confidence:

$\text{KL}_{\text{both}} = \text{KL}\left(\pi_\theta(\cdot \mid o_t, \emptyset, \emptyset, a_{<i}) \;\parallel\; \pi_\theta(\cdot \mid o_t, q_t, I, a_{<i})\right)$

What it captures: Combined dependence on both task identity and robot state. This is the most "uncertain" reference — removing the maximum amount of information while retaining the visual context. It is most useful when neither text alone nor state alone provides a sufficient uncertainty baseline, e.g., in environments where the model has memorized both the task structure AND the typical state-action mapping.

Design choice: why mask conditions rather than use a separate "unconditional" model? An alternative would be to train a separate unconditional policy $\pi_{\text{uncond}}(a \mid o_t)$ (observation only) as the reference. This would require maintaining and loading two models. Condition-masking uses the same model with different inputs, achieving the same effect with no additional parameters and a single forward pass that can compute both $P$ and $Q$ in parallel (both distributions are required at each token position, and the VLA can batch these computations).

Design choice: why condition-masking rather than noise-based perturbation? Another alternative would be to add Gaussian noise to observations or states and measure how the action distribution changes. Condition-masking is preferred because it completely removes information along semantically meaningful axes (task identity, robot configuration) rather than adding unstructured noise, which is harder to calibrate (how much noise is "enough"?) and may not produce interpretable confidence signals.

Aggregation Strategy: Which Tokens Matter Most?

The raw per-token KL divergences $C_i$ must be aggregated across the variable-length action sequence to produce a single scalar confidence score for Best-of-N selection. The aggregation strategy turns out to be critical — naive summation across all tokens performs worse than selective aggregation, and the optimal strategy depends on the tokenizer.

The FAST tokenizer structure (Pertsch et al., 2025). To understand the aggregation choices, we need to understand how actions are tokenized. The FAST tokenizer converts a continuous action chunk $a_{t:t+H} \in \mathbb{R}^{H \times d}$ (where $d$ is the action dimension, e.g., 7 for end-effector delta positions + gripper) into a variable-length sequence of discrete tokens using a frequency-based decomposition. Tokens are ordered from low-frequency components (capturing the overall movement arc — the "coarse" action) to high-frequency components (capturing fine-grained adjustments). The sequence length $T$ varies per action depending on how many frequency components are needed to represent it.

Aggregation variants tested (Table 5f):

The paper compares several strategies for selecting the token index set $\mathcal{I}$ :

Sum (all tokens): $\mathcal{I} = \{1, 2, \ldots, T\}$ , $C_{\tilde{a}} = \sum_{i=1}^T C_i$ . This achieves 26.1% on RoboCasa pick-and-place — the worst performer.
Average (all tokens): $C_{\tilde{a}} = \frac{1}{T} \sum_{i=1}^T C_i$ . This achieves 24.7% — even worse, because averaging dilutes the signal from the informative early tokens with noise from later tokens.
First 5 tokens: $\mathcal{I} = \{1, 2, 3, 4, 5\}$ , $C_{\tilde{a}} = \sum_{i=1}^5 C_i$ . This achieves 31.0% — the best performer.
First 10 tokens: $\mathcal{I} = \{1, \ldots, 10\}$ , $C_{\tilde{a}} = \sum_{i=1}^{10} C_i$ . This achieves 26.6% — performance degrades, suggesting that tokens 6–10 introduce noise.

What it computes: For each candidate action, the summation of KL divergence scores over the first 5 token positions only, regardless of the action's total sequence length $T$ .

Why the first 5 tokens: The paper hypothesizes that this result is "correlated with the nature of the FAST tokenizer... each action sequence is composed of a variable number of action tokens, which are aligned from low- to high-frequency" (Section 4.3, Aggregation strategy paragraph). The low-frequency tokens (the first few) encode the macro-level action: "move gripper to position X," "close gripper," "move to position Y," "open gripper." These are the decisions that determine task success or failure. High-frequency tokens (later positions) fine-tune the trajectory with small adjustments, but these adjustments are typically less critical and introduce variance in the confidence signal — a small KL divergence on token 12 may reflect random noise in the frequency decomposition rather than genuine uncertainty about the action.

Why not averaging across all tokens: Averaging penalizes actions with many high-frequency tokens (i.e., actions requiring fine adjustments), which may actually be the more precise actions. The paper's result (24.7% for average vs. 31.0% for first-5) suggests that the additional high-frequency tokens hurt more than they help as confidence signals.

For OpenVLA, a different strategy: OpenVLA uses a fixed-length action sequence (not the variable-length FAST tokenizer), so the paper uses "the average score across the entire token sequence, since its output sequence length is fixed to the action dimension of the training data" (Appendix A.3). This is noted in Table 6 but not ablated separately — it's a domain-adaptive choice rather than a universal rule.

Regularization Temperature for the Reference Distribution

When the condition-masking distribution $Q$ is computed, the paper does not use the raw softmax output of the VLA. Instead, a regularization temperature $\tau_{\text{reg}}$ is applied to the logits BEFORE the softmax:

$Q_i^{\text{reg}} = \text{softmax}(\ell_Q / \tau_{\text{reg}})$

where $\ell_Q$ is the logit vector from the VLA's forward pass with masked inputs, and $\tau_{\text{reg}}$ is a high temperature (swept over $\{4.0, 6.0, 8.0, 10.0, 12.0, 14.0, 16.0\}$ per Appendix A.3).

What it computes: A flattened version of the condition-masked distribution, where probability mass is spread more evenly across the action token vocabulary. At $\tau_{\text{reg}} = 1.0$ , $Q$ is the raw VLA output under masking — which can still be peaked if the model has memorized certain action patterns even without conditions. At $\tau_{\text{reg}} = 4.0$ (the best value for RoboCasa, Table 5e), the distribution is substantially flatter, approaching (but not reaching) uniformity.

Why a high temperature is necessary: Table 5(e) shows the effect of varying $\tau_{\text{reg}}$ . At $\tau_{\text{reg}} = 1.0$ (the default, equivalent to no regularization), the condition-masking distribution performs at 28.8% on RoboCasa pick-and-place — actually WORSE than the uniform-KL baseline (30.0% in Table 5a). The paper explains:

"It is possibly because condition-masking distribution may be 'peaked' around certain action tokens, which undermines the purpose of distributional confidence by failing to consider the entire probability distribution."

In other words, even when conditions are masked, the VLA's distribution can be concentrated — it might, for example, always assign high probability to "move forward slightly" because that's the most common action in its training data, regardless of task. If $Q$ is peaked and $P$ is also peaked (but on different tokens), the KL divergence will be large, but for the wrong reason — both distributions are concentrated, not because the full-condition prediction is confident, but because the model has memorized default behaviors. The regularization temperature forces $Q$ toward higher entropy, ensuring that large KL values genuinely reflect the full-condition distribution deviating from an uncertain baseline.

Why $\tau_{\text{reg}} = 4.0$ is optimal: At $\tau_{\text{reg}} = 4.0$ , the condition-masked distribution achieves 31.0% on pick-and-place, compared to 30.0% at $\tau_{\text{reg}} = 8.0$ and 25.4% at $\tau_{\text{reg}} = 2.0$ . Too little regularization ( $\tau = 2.0$ ) leaves $Q$ too peaked; too much ( $\tau = 8.0$ ) pushes $Q$ toward uniform, at which point the method reduces to the uniform-KL baseline (Kang et al., 2025) and loses the benefit of the task-aligned reference. The sweet spot flattens $Q$ enough to represent meaningful uncertainty while retaining enough task structure to provide a better baseline than pure uniformity.

How this interacts with the sampling temperature $\tau$ : The sampling temperature $\tau$ (used in candidate generation) and the regularization temperature $\tau_{\text{reg}}$ (used in reference distribution computation) are independent hyperparameters. $\tau$ controls the diversity of the candidate set — it is swept over $\{0.1, 0.3, 0.5, 0.7, 1.0\}$ and fixed at 0.5 for ablation experiments (Table 5 caption). $\tau_{\text{reg}}$ controls the flatness of the reference distribution used for confidence scoring. The paper does not explore their interaction systematically — they are optimized independently per dataset.

Joint Training Strategy: Teaching the Model to Be Uncertain

A critical practical challenge: standard VLAs are trained exclusively on full-condition data — they have never seen masked inputs during training. When $\pi_\theta$ receives a masked input at test time (e.g., $q_t = \emptyset$ ), it is operating out-of-distribution, and its output distribution may be arbitrary or degenerate. The paper reports that "directly masking inputs often leads to unintended actions" (Section 3.3), producing reference distributions that are not meaningful uncertainty baselines.

The joint training procedure.

To address this, the paper augments the standard imitation learning objective with condition dropout. During fine-tuning on the target dataset $\mathcal{D}$ , each training sample $((o_t, q_t), a_{t:t+H}, I)$ is presented to the model in four variants, randomly selected via a masking set $\mathcal{M}$ :

$\mathcal{M} = \left\{ (q_t, I), (q_t, \emptyset), (\emptyset, I), (\emptyset, \emptyset) \right\}$

corresponding to: (i) full-condition — both state and instruction present; (ii) text-masking — instruction removed; (iii) state-masking — proprioceptive state removed; (iv) both-masking — both removed.

The augmented dataset $\mathcal{D}_{\text{augmented}}$ contains all four variants for each original trajectory, and the joint imitation learning objective becomes:

$\mathcal{L}_{\text{Joint-IL}}(\theta; \mathcal{D}) = -\mathbb{E}_{((o_t, q_t), a_{t:t+H}, I) \sim \mathcal{D}} \left[ \mathbb{E}_{(q_t^{(m)}, I^{(m)}) \in \mathcal{M}} \left[ \log \pi_\theta(a_t \mid o_t, q_t^{(m)}, I^{(m)}) \right] \right]$

where $a_t$ is the expert action sequence (tokenized) at timestep $t$ , $q_t^{(m)}$ is either the true state or the empty token, and $I^{(m)}$ is either the true instruction or the empty token.

What it computes: The standard behavior cloning (negative log-likelihood) loss, but averaged over all four masking variants for each demonstration sample. The model is trained to produce the expert action $a_t$ regardless of which conditions are present — including when conditions are missing. This means the model learns: (a) the full-condition mapping (standard IL), (b) what action to take when it knows the state but not the task, (c) what action to take when it knows the task but not the state, and (d) what action to take when it knows neither — essentially a "default" action distribution based only on the visual observation.

Why training with masked conditions helps: When the model has been trained to produce meaningful actions under masking, the condition-masked distribution $Q$ at test time represents a genuine counterfactual: "what would the model do if it didn't know the task/state?" The KL divergence from this learned counterfactual to the full-condition prediction $P$ is then a clean measure of how much the missing information changed the prediction.

Dropout rates: For π0-FAST, the paper applies 10% dropout for each masking variant (10% text-masked, 10% state-masked, 10% both-masked) on RoboCasa and LIBERO. For SIMPLER-WidowX, only state-masking is applied (10% dropout), reflecting that in a pick-and-place-only environment, text-masking is less informative. For OpenVLA (which doesn't take state input), only text-masking is applied (10% dropout). These rates are not ablated — they are chosen heuristically to ensure the model still sees the full-condition case 70% of the time (for RoboCasa/LIBERO), maintaining strong standard IL performance while learning the masked distributions.

Effectiveness of joint training alone: Table 5(d) shows that joint training WITHOUT MG-Select (row "Joint-IL: ✓, MG-Select: ✗") improves performance from 17.0% to 28.5% on RoboCasa pick-and-place (100 demos). This is a substantial gain that the paper attributes to condition dropout acting as a regularizer: "Joint training alone already outperforms vanilla imitation learning, likely because condition-masking prevents the model from overfitting." This is consistent with the general regularization effect of dropout — by forcing the model to produce correct actions even with missing information, it learns more robust representations.

Effectiveness of joint training WITH MG-Select: Adding MG-Select on top of the jointly trained model (row "Joint-IL: ✓, MG-Select: ✓") further improves from 28.5% to 31.0%. The delta from joint-training-only (28.5%) to joint-training + MG-Select (31.0%) represents the pure contribution of the test-time confidence-based selection, while the delta from vanilla IL (17.0%) to joint-training + MG-Select (31.0%) represents the combined effect.

Why not train the model on uniformly random actions as the reference? An alternative would be to train a separate "random" policy or use a uniform distribution as $Q$ . The uniform-KL baseline in Table 5(a) achieves 30.0% — competitive but consistently below condition-masking (31.0%). The advantage of the condition-masking approach is that $Q$ retains task- and robot-specific structure (e.g., the model still knows actions should be within the workspace, should not exceed joint limits, should move toward visible objects) even when task identity or state is unknown. A uniform distribution over all action tokens has no such structure — it assigns equal probability to physically impossible actions as to plausible ones, making the KL divergence noisier as a confidence signal.

Choice of Masking Variant per Environment

The paper does not propose a single universal masking variant. Instead, the optimal variant depends on the task environment characteristics, and the choice is treated as a hyperparameter swept during deployment (Appendix A.3: "We search for the optimal configuration on each dataset within... variants ∈{text, state, text&state}").

RoboCasa (24 diverse tasks): Text-masking is optimal (Table 5c: 31.0% with text-masking vs. 30.1% state-masking vs. 29.7% both-masking on pick-and-place). The paper explains: "RoboCasa benchmark, which has multiple task types, text-masking or text&state-masking are more effective, since the model cannot determine the correct action without instructions." When the task set is heterogeneous (pick, place, open, close, etc.), knowing the instruction is critical — masking it creates a genuinely uncertain reference. Knowing the state without the instruction doesn't help if you don't know whether to pick or place.

SIMPLER-WidowX (pick-and-place only): State-masking is optimal (implicit from Appendix A.1: "only dropout 10% of state data in SIMPLER-WidowX"). The paper explains: "SIMPLER-WidowX benchmark, which consists solely of pick-and-place tasks, state-masking confidence works best because the model already memorizes how to pick and place objects without task instructions." When all tasks share the same structure, the instruction provides little discriminating information — the model's behavior is determined primarily by object positions and gripper state.

General principle: The optimal masking variant removes the information that is most necessary for task-specific action selection in that environment. In diverse-task settings, this is the instruction. In homogeneous-task settings requiring precision, this is the robot state.

Efficient Deployment: Single-Prefill Strategy

The paper acknowledges a practical latency concern: MG-Select requires $N$ forward passes through the VLA for the candidate generation stage (plus additional passes for the condition-masked reference distributions). In VLAs, each forward pass includes a prefill step — processing the visual observation and instruction through the vision encoder and language model before decoding action tokens. This prefill is computationally expensive and typically dominates inference time.

The single-prefill optimization. Instead of running the full prefill $N$ times (once per candidate), the paper proposes sharing a single prefill across all candidates. Specifically:

The observation and instruction are encoded once through the vision-language backbone to produce a shared representation (hidden states).
This shared representation is used as the conditioning context for all $N$ candidate decodings, which are then generated in parallel by branching from the shared prefix.

What it computes: The same $N$ candidate actions as vanilla MG-Select, but with only one prefill instead of $N$ prefills, trading a small amount of per-candidate decoding overhead for a large reduction in total prefill cost.

Latency reduction: Figure 3 and Table 10 quantify the benefit. On LIBERO-Object:

Vanilla MG-Select with $N = 1$ : 20.2 seconds (baseline single-action inference)
Vanilla MG-Select with $N = 4$ : 43.4 seconds (more than 2× the single-action latency)
MG-Select + Single Prefill with $N = 4$ : 23.7 seconds (only 17% overhead over single-action)
MG-Select + Single Prefill with $N = 16$ : 30.4 seconds (50% overhead)

The single-prefill strategy achieves roughly a 45% reduction in latency compared to vanilla MG-Select at $N = 4$ , making the method practical for real-time robot control where inference latency directly affects cycle time. This optimization is enabled by the autoregressive VLA architecture — the shared prefill is possible because all candidates condition on the same observation and instruction; only the action token sequences differ.

Why this matters: Without the single-prefill optimization, MG-Select with $N = 4$ would more than double the robot's reaction time per timestep (from ~20s to ~43s in this benchmark), which is unacceptable for many real-world tasks. The optimization brings the overhead down to ~17%, making Best-of-N selection viable in latency-sensitive deployment scenarios.

Summary of Design Choices and Their Justifications

KL divergence from $Q$ to $P$ (not $P$ to $Q$ ): measures how much the masked distribution's probability mass shifts relative to the full-condition prediction, isolating the effect of the removed information. The reverse direction would conflate concentration (which is always present due to behavior cloning) with genuine condition-dependence.
Condition-masking over uniform reference: produces a task-aligned uncertainty baseline that respects the robot's action manifold (physically possible, robot-typical actions) while representing maximum task-specific uncertainty. A uniform distribution would assign equal probability to impossible actions, introducing noise.
Joint training with condition dropout over plug-and-play masking: ensures the VLA's masked outputs are meaningful rather than degenerate out-of-distribution responses. The 70% full-condition / 30% masked split maintains standard IL performance while learning the reference distributions.
High regularization temperature for the reference: flattens the condition-masked distribution to prevent it from being peaked (which would produce large KL values for the wrong reason — both distributions concentrated, not because the full-condition prediction is confident). $\tau_{\text{reg}} = 4.0$ empirically balances flattening against retaining task structure.
Truncating to first 5 tokens for FAST tokenizer aggregation: exploits the frequency-ordered structure of the FAST tokenizer, where early tokens encode the macro-level action decisions that determine task success, while later tokens introduce noise from fine-grained frequency components.
Single-prefill deployment: amortizes the expensive vision-language encoding step across all candidates, reducing the latency overhead of Best-of-N sampling from >100% to ~17% at $N = 4$ , making the method practical for real-time control.
Environment-specific masking variant selection: acknowledges that the optimal uncertainty baseline depends on which information is most critical for the task distribution — instruction in diverse-task settings, state in homogeneous precision tasks.

4. Key Insights and Innovations

Innovation 1: Reframing Test-Time Action Selection as a Counterfactual Confidence Problem

The dominant paradigm in test-time scaling for robotics—inherited from the LLM literature—treats action selection as a verification problem: generate candidates, then score them with an external critic that estimates their quality. Nakamoto et al. (2024) trains value functions via offline RL to rank actions. Kwok et al. (2025) trains VLM-based reward models on synthetic preferences. Both follow the same structural assumption: the VLA proposes, the verifier disposes, and the verifier must be trained separately because the VLA's own likelihood is an unreliable quality signal.

MG-Select makes a fundamentally different move. Instead of asking "how good is this action?" (a quality estimation problem), it asks "how much does this action depend on knowing what to do?" (a counterfactual dependence problem). The conceptual shift is from absolute quality scoring to relative information sensitivity. An action is not selected because some critic predicts it will succeed, but because the model's own probability of producing it drops substantially when task-relevant information is withheld.

This reframing matters for three reasons. First, it eliminates the distribution-shift vulnerability inherent in external verifiers. A value function trained on specific tasks and objects fails on novel ones because its quality estimates are out-of-distribution. But the counterfactual confidence signal—"does removing the instruction change what the model predicts?"—generalizes automatically: if the VLA can process a novel object or instruction, it can also process it under masking, and the divergence computation requires no task-specific training. The OOD results (Table 3: 35% improvement on unseen objects like a roll of tape and a lighter cup) directly validate this claim—MG-Select provides gains under distribution shift where external verifiers are documented to fail.

Second, it converts a modeling problem (training a verifier) into an inference procedure (comparing the model's own outputs under different input conditions). The distinction is subtle but practically significant: MG-Select requires no new model architecture, no reinforcement learning, no preference dataset construction. The entire selection mechanism is implemented as additional forward passes through the VLA with modified inputs—computationally cheap relative to verifier training, and trivially compatible with any autoregressive VLA architecture (demonstrated on both π0-FAST and OpenVLA in Table 6).

Third, it provides a principled diagnostic for action precision. The KL divergence from a condition-masked reference isolates the component of the model's prediction attributable to the masked information. When this divergence is small, the model's action is largely determined by default behavior (observation alone, or observation + instruction alone) rather than by the specific task conditions—suggesting the action is generic and may lack the precision needed for contact-rich manipulation. When it is large, the model actively uses the task specification to shape its prediction—suggesting a deliberate, task-tailored action. This diagnostic framing, while not formally proven in the paper, provides a conceptual tool for understanding why certain actions succeed and others fail, beyond the binary success/failure signal that verifiers provide.

The comparison to Kang et al. (2025)'s self-certainty for LLMs is instructive. Self-certainty also uses divergence from a reference (a uniform distribution) as a confidence signal. MG-Select's innovation is recognizing that for VLAs, a task-aligned reference—one that respects the action manifold—is necessary because a uniform distribution over action tokens has no relationship to what actions are physically possible or robot-typical. The three masking variants (text, state, both) are the mechanism for constructing such a reference, but the deeper insight is that uncertainty must be measured relative to a baseline that shares the model's inductive biases. This principle may generalize beyond VLAs to any domain where the output space has strong structural constraints that a uniform distribution would ignore.

Is this a fundamental reframing or an incremental twist on self-certainty? I would argue it is a fundamental reframing adapted to a new domain, not merely an application. The key shift—from quality estimation to counterfactual dependence—is not present in the self-certainty literature, which still fundamentally asks "how confident is the model in this output?" rather than "how much did the model need the task specification to produce this output?" The difference is that confidence (self-certainty) can be high for memorized behaviors; counterfactual dependence cannot be, because memorized behaviors persist under masking.

Innovation 2: Diagnosing and Circumventing the Over-Concentration Problem in Behavior-Cloned Action Distributions

A central empirical finding that shapes the entire method—but which the paper treats as a motivating observation rather than a named contribution—is the identification of distribution over-concentration as the specific mechanism by which behavior cloning undermines test-time sampling. The paper observes that VLAs fine-tuned on expert demonstrations "often memorize expert trajectories, causing the probability distribution over action tokens to become overly concentrated, which leads to multiple sampling converging to the same result" (Section 1, paragraph 4).

This is not merely a restatement of the known fact that behavior cloning produces peaked distributions. It is a diagnosis of why test-time scaling strategies from the LLM literature fail when naively ported to VLAs. In LLM reasoning, Best-of-N via likelihood works because the model's distribution over completions is genuinely multimodal—there are multiple valid solutions, and likelihood correlates (imperfectly) with correctness. In VLA action prediction fine-tuned via behavior cloning, the distribution collapses onto a near-delta function around the expert trajectory. Sampling multiple times from this distribution produces candidates that are nearly identical, so selection among them is meaningless regardless of the selection criterion.

The paper's evidence for this diagnosis is indirect but consistent. Table 5(a) shows that likelihood-based Best-of-N (30.5%) improves only modestly over single-sample sampling (27.6%) on the jointly trained model, and the paper attributes this to low diversity among candidates. More tellingly, the uniform-KL baseline (30.0%) performs similarly to likelihood-based selection, suggesting that when the distribution is concentrated, both likelihood and simple divergence measures capture roughly the same (limited) signal. MG-Select's gain (31.0%) over these baselines, while modest in absolute percentage points, is significant because it operates in precisely the regime where candidate diversity is minimal—it extracts a useful confidence signal where likelihood cannot.

The joint training strategy (Section 3.3) can be understood as a direct response to this diagnosis. By training the model to produce correct actions even when conditions are masked, joint training serves two functions. First, it acts as a regularizer (the paper notes it alone improves performance from 17.0% to 28.5% on RoboCasa pick-and-place with 100 demos, Table 5d), preventing the distribution from collapsing entirely onto the memorized trajectory. Second, it ensures that the condition-masked distributions used as references are meaningful rather than degenerate—a model that has never seen masked inputs will produce arbitrary outputs when masked, making the KL divergence uninformative.

The significance of this contribution is that it identifies a domain-specific failure mode that is likely to recur as test-time scaling is applied to other embodied AI domains. Any domain where the training objective encourages distribution collapse (behavior cloning, but also potentially next-token prediction on narrow expert demonstrations in other continuous-control settings) will face the same challenge: candidates lack diversity, and likelihood-based selection cannot discriminate. The paper's solution—using counterfactual dependence rather than likelihood as the selection signal—is one approach, but the broader contribution is the diagnostic framework itself: before applying test-time scaling, check whether the proposal distribution has sufficient diversity for selection to matter.

This is an incremental but practically important contribution. The over-concentration problem is not theoretically surprising—it follows directly from the maximum likelihood objective applied to a single expert trajectory. But the paper is the first (to my knowledge) to explicitly connect this property to the failure of test-time scaling in VLAs, and to design a selection mechanism specifically robust to it.

Innovation 3: The Regularized Reference Distribution as an Entropy-Calibrated Uncertainty Baseline

The paper makes a subtle but critical design choice that distinguishes MG-Select from a naive application of KL divergence to condition-masked distributions: the use of a high regularization temperature on the reference distribution (Section 4.3, Table 5e). This is not merely a hyperparameter tweak—it reflects a conceptual insight about what makes a reference distribution useful for confidence measurement.

The naive approach would compute KL(Q ∥ P) where both Q and P are the raw softmax outputs of the VLA (temperature 1.0). The paper finds this performs worse than the uniform-KL baseline (28.8% vs. 30.0% on RoboCasa pick-and-place, Table 5e vs. Table 5a). The reason, as the paper explains, is that the condition-masked distribution Q can itself be peaked—"the condition-masking distribution may be 'peaked' around certain action tokens, which undermines the purpose of distributional confidence by failing to consider the entire probability distribution."

The insight is that confidence is not about how different two distributions are, but about how much the full-condition distribution deviates from a state of maximum uncertainty. If both distributions are peaked (but on different tokens), the KL divergence will be large, but not because the full-condition prediction is "confident" in a meaningful sense—rather, because both distributions are concentrated due to the model's memorized priors. The confidence signal is confounded by the model's default concentration.

Applying a high temperature (τ_reg = 4.0) to the reference distribution forces it toward higher entropy—flattening it so that it represents genuine uncertainty rather than memorized defaults. At τ_reg = 4.0, the KL divergence cleanly measures how much the full-condition distribution diverges from an uncertain baseline, isolating the component of the prediction that is attributable to the task conditions. Too little regularization (τ_reg = 2.0, 25.4% accuracy) leaves Q too peaked; too much (τ_reg = 8.0, 30.0%) pushes Q toward uniform, at which point MG-Select reduces to the uniform-KL baseline and loses its advantage.

This is not a theoretical contribution—the paper does not formalize the relationship between temperature, entropy, and confidence calibration. But it is a practical insight with broader applicability: when using a learned reference distribution for confidence estimation (in any domain), the reference must be explicitly regularized toward high entropy to serve as a meaningful uncertainty baseline. Without this regularization, the reference inherits the model's default concentration, and the divergence confounds "departure from uncertainty" with "departure from a different concentrated mode."

The absence of this insight in the LLM self-certainty literature (Kang et al., 2025) is notable. LLM self-certainty uses a uniform distribution as the reference, which is already maximum-entropy—so the regularization issue does not arise. MG-Select faces this issue precisely because it uses a learned (condition-masked) reference rather than a uniform one, trading the reference's task alignment for the need to explicitly manage its entropy. The regularization temperature is the mechanism for recovering the best of both worlds: a task-aligned reference that nonetheless represents high uncertainty.

5. Experimental Analysis

Evaluation Methodology

Datasets. The paper evaluates on four distinct benchmarks spanning simulation and real-world settings:
- RoboCasa (Nasiriany et al., 2024): 24 atomic household kitchen tasks in simulation; the paper focuses on 8 pick-and-place tasks and also reports aggregate "All" results across all 24 tasks. Training uses 30, 100, or 300 demonstrations per task to test data-efficiency scaling.
- SIMPLER-WidowX (Li et al., 2024): 4 real-to-sim pick-and-place tasks ("Spoon on Towel," "Carrot on Plate," "Stack Cubes," "Eggplant in Basket") evaluated over 24 trials each. No simulated training data is provided, so the base model is trained on BridgeData V2 (Walke et al., 2023).
- LIBERO (Liu et al., 2023): Four task suites testing different generalization axes — LIBERO-Spatial (layout variations), LIBERO-Object (object variations), LIBERO-Goal (goal variations), and LIBERO-Long (long-horizon tasks requiring sustained precision). Each suite contains 10 tasks with 50 trials per task (500 trials total per suite).
- Real-world Franka tasks: Custom in-distribution (ID) pick-and-place tasks with 4 objects (teddy bear, cube, rigid cup, sponge) across 4 start-goal configurations ("Box to Bowl," "Box to Plate," "Basket to Bowl," "Plate to Basket"), plus 2 out-of-distribution (OOD) tasks with unseen objects (lighter cup, roll of tape).
Base models. Two autoregressive VLA architectures are tested:
- π0-FAST (Pertsch et al., 2025): Uses Paligemma-3B VLM (Beyer et al., 2024) as the backbone, fine-tuned on each benchmark from the pre-trained checkpoint. This is the primary model for all experiments. For real-world experiments, a variant fine-tuned on the DROID dataset (Khazatsky et al., 2024), denoted π0-FAST-DROID, is used as the starting point before task-specific fine-tuning.
- OpenVLA (Kim et al., 2024): Uses Prismatic-7B VLM (Karamcheti et al., 2024) as the backbone, fine-tuned with LoRA (r = 32). Evaluated only on LIBERO to demonstrate architecture-agnostic effectiveness.
Metrics. The primary metric is task success rate (%) — the fraction of trials where the robot successfully completes the specified task. For SIMPLER-WidowX, both task success rate and grasp success rate are reported (Table 9). For RoboCasa, results are averaged over 50 trials per task. For real-world experiments, results are averaged over 24 trials per ID task (4 objects × 6 trials) and 16 trials per OOD task. Results for MG-Select methods are averaged over 3 random seeds; baseline results are taken from the respective original papers.
Baselines. The paper compares against several categories:
- Greedy decoding: The standard single-inference VLA baseline — always selects the highest-probability action token at each step.
- Stochastic sampling (τ = 0.5, N = 1): Single-sample temperature-based sampling without Best-of-N selection — measures whether diversity alone helps.
- Uniform-KL Best-of-N (Kang et al., 2025): Selects actions by KL divergence from a uniform distribution over the action vocabulary — the closest prior work in self-certainty for language models.
- Likelihood Best-of-N: Selects actions with the highest sequence log-probability under the VLA's full-condition distribution — the simplest ML-based selection criterion.
- External model baselines: On RoboCasa, GR00T N1 (Bjorck et al., 2025). On SIMPLER-WidowX, RT-1-X (O'Neill et al., 2024), Octo (Team et al., 2024), RoboVLM (Liu et al., 2025), and SpatialVLA (Qu et al., 2025).
Generation budget. The budget is measured as N, the number of candidate action sequences sampled per timestep. The paper sweeps N ∈ {1, 2, 4, 8, 16} (Table 5b). The main results use N = 4 as the default, chosen because performance mostly saturates beyond this point (Table 5b: 31.0% at N = 4 vs. 30.7% at N = 16 on RoboCasa pick-and-place). A key distinction: the condition-masked reference distributions (text, state, or both) are computed as additional forward passes but their cost is not counted in the "generation budget" — only candidate actions count. This is a favorable accounting for MG-Select relative to likelihood-based methods (which require no reference computations), but the paper's single-prefill optimization (Figure 3) mitigates the practical latency concern.
Cross-validation / statistical protocol. No cross-validation is used for strategy selection. Instead, the paper performs a hyperparameter sweep per dataset over sampling temperature τ ∈ {0.1, 0.3, 0.5, 0.7, 1.0}, number of candidates N ∈ {4, 8}, masking variant ∈ {text, state, text&state}, and regularization temperature ∈ {4.0, 6.0, 8.0, 10.0, 12.0, 14.0, 16.0}, selecting the best configuration and reporting that result (Appendix A.3). This means the reported numbers represent best-case performance after hyperparameter optimization on the test set, not a held-out validation protocol — a potential concern for overfitting that the paper does not address. Results are averaged over 3 random seeds for MG-Select variants; the paper does not report standard deviations or confidence intervals.

Main Quantitative Results

RoboCasa: Strong Gains in Low-Data Regime, Consistent Improvement Across Scales

Table 1 presents the headline RoboCasa results. The base π0-FAST model (reproduced by the authors, marked with †) achieves 5.3% on pick-and-place tasks with 30 demonstrations, 17.0% with 100, and 43.2% with 300. MG-Select without joint training ("+ MG-Select") improves these to 7.2%, 22.6%, and 46.5% respectively — a 36%, 33%, and 7.6% relative improvement. MG-Select with joint training ("+ MG-Select*") pushes further to 14.2%, 31.0%, and 46.9% — representing relative improvements of 168%, 82%, and 8.6% over the base model.

Three patterns stand out:

1. Gains are largest in the low-data regime. The 168% relative improvement with 30 demonstrations dwarfs the 8.6% with 300 demonstrations. This aligns with the paper's narrative: when data is scarce, the base model's precision is poor, and test-time selection can recover substantial ground. When data is abundant (300 demos), the base model is already strong (43.2%), and the ceiling limits further gains. MG-Select* at 300 demos (46.9%) is only marginally above MG-Select without joint training (46.5%), suggesting that joint training's benefit also saturates with sufficient data.

2. Joint training provides roughly half the total gain. Comparing MG-Select (no joint training) to MG-Select* (with joint training): at 30 demos, joint training adds 7.0 percentage points (from 7.2% to 14.2%); at 100 demos, it adds 8.4 points (22.6% to 31.0%); at 300 demos, it adds only 0.4 points (46.5% to 46.9%). Joint training matters most when the base model is weak, consistent with its role as both a regularizer (preventing overfitting to limited data) and a teacher of meaningful condition-masked distributions.

3. MG-Select matches or exceeds GR00T N1. At 30 demos, MG-Select* achieves 14.2% on pick-and-place vs. GR00T N1's 0.4% — a dramatic gap. At 100 demos, 31.0% vs. 2.2%. At 300 demos, 46.9% vs. 22.6%. However, this comparison is somewhat apples-to-oranges: GR00T N1 is a different architecture (diffusion-based, not autoregressive) trained under different conditions, and its reported numbers are taken from the original paper rather than reproduced by these authors. The comparison mainly establishes that π0-FAST + MG-Select is competitive with contemporary VLA approaches, not that MG-Select is conclusively superior to GR00T N1.

The "All" column in Table 1 (aggregating all 24 RoboCasa tasks) shows a consistent but smaller effect: 30.9% → 34.6% → 62.9% across the three data scales with MG-Select*. The gains are concentrated in pick-and-place tasks (where precision matters most), while the remaining 16 tasks (open-and-close, others) are less sensitive to action precision. Table 8 in Appendix B breaks this down: for "Open and Close" tasks, MG-Select* improves from 51.3% to 53.2% (30 demos), from 60.7% to 67.3% (100 demos), and from 74.7% to 81.0% (300 demos). Gains exist across categories but are most pronounced where millimeter-level accuracy determines success.

SIMPLER-WidowX: Consistent but Modest Improvement

Table 2 shows MG-Select* applied to π0-FAST on SIMPLER-WidowX. The base model achieves 46.9% average success rate; MG-Select* improves this to 50.3% — a 7.3% relative improvement. Per-task breakdown:

Spoon on Towel: 66.7% → 69.4% (+2.7 points)
Carrot on Plate: 70.8% → 75.0% (+4.2 points)
Stack Cubes: 41.7% → 43.1% (+1.4 points)
Eggplant in Basket: 8.3% → 13.9% (+5.6 points)

The gains are small in absolute terms but consistent across all four tasks. The Eggplant in Basket task merits special attention: the base model performs dramatically worse here (8.3%) than on other tasks (41.7–70.8%). The paper attributes this to distribution shift — "its background differs substantially from the other three tasks, making it sensitive to model-specific training configurations." MG-Select still provides a 5.6 percentage-point improvement (67% relative gain), demonstrating robustness even when the base model struggles. However, at 13.9%, the post-MG-Select performance is still far below the other tasks, confirming that MG-Select amplifies existing capability but cannot fully compensate for fundamental weaknesses.

The paper's comparison to SpatialVLA (100% on Eggplant, 42.7% overall) raises an interesting tension: SpatialVLA achieves much higher performance on one task but lower on others. The authors note this suggests sensitivity to "model-specific training configurations," implying that SpatialVLA may be overtuned to Eggplant's visual characteristics at the expense of general performance. MG-Select's gains are more uniform, which the paper implicitly frames as a virtue — though the absolute numbers on Eggplant remain low.

Table 9 in Appendix B additionally reports grasp success rates separately from overall task success. For MG-Select*, grasp rates are consistently higher than task success rates (e.g., 87.5% grasp vs. 69.4% task success on Spoon on Towel), indicating that some failures occur after a successful grasp (e.g., during transport or release). The grasp rate improvements from MG-Select are modest: 83.3% → 87.5% (Spoon), 83.3% → 83.3% (Carrot — no change), 91.7% → 79.2% (Stack Cubes — a decrease!), and 8.3% → 26.4% (Eggplant). The decrease on Stack Cubes is notable and not discussed in the main text — it may reflect MG-Select selecting actions with high KL divergence that are confident on grasp but imprecise on stacking, or it may be noise given the small trial count (24 trials).

LIBERO: Architecture-Agnostic Gains, Most Pronounced on Hardest Tasks

Table 6 reports results on LIBERO across four task suites, evaluated on both π0-FAST and OpenVLA with MG-Select*:

π0-FAST results:

LIBERO-Spatial: 97.4% → 97.2% (−0.2 points)
LIBERO-Object: 95.4% → 98.0% (+2.6 points)
LIBERO-Goal: 95.6% → 94.5% (−1.1 points)
LIBERO-Long: 79.6% → 82.7% (+3.1 points)
Average: 92.0% → 93.1% (+1.1 points)

OpenVLA results:

LIBERO-Spatial: 85.2% → 81.7% (−3.5 points)
LIBERO-Object: 63.7% → 72.5% (+8.8 points)
LIBERO-Goal: 75.5% → 73.6% (−1.9 points)
LIBERO-Long: 52.5% → 55.4% (+2.9 points)
Average: 69.2% → 70.8% (+1.6 points)

The pattern is striking in its consistency across both architectures: gains appear on the hardest task suites (LIBERO-Object, LIBERO-Long) while the easier suites (LIBERO-Spatial, LIBERO-Goal) show flat or slightly negative results. This mirrors the RoboCasa finding where gains are largest in the low-data regime — when the base model already achieves 95%+ success (π0-FAST on LIBERO-Spatial and LIBERO-Goal), there is little room for improvement and MG-Select may introduce selection noise that slightly degrades performance. When the base model struggles (79.6% on LIBERO-Long for π0-FAST, 63.7% on LIBERO-Object for OpenVLA), MG-Select provides meaningful gains.

The OpenVLA results are a critical validation of the method's generality. OpenVLA uses a different VLM backbone (Prismatic-7B vs. Paligemma-3B), a different fine-tuning approach (LoRA vs. full fine-tuning), and does not take proprioceptive state as input (so only text-masking is applicable, per Appendix A.3). Despite these differences, MG-Select* improves average performance from 69.2% to 70.8%. The 8.8-point gain on LIBERO-Object is the largest single-suite improvement across both architectures. This suggests that confidence-based selection via condition-masking is not an artifact of π0-FAST's specific architecture or training — it captures a general property of autoregressive VLA action distributions.

The negative results on LIBERO-Spatial for both architectures (−0.2 for π0-FAST, −3.5 for OpenVLA) are concerning and not discussed in the main text. The paper focuses on the positive average improvement, but the degradation on specific suites implies that MG-Select can actively harm performance when the base model is already near-ceiling, likely due to selection noise — the confidence signal may be miscalibrated in high-performance regimes, causing the system to select suboptimal actions that have spuriously high KL divergence. This is a form of the over-optimization phenomenon documented in the LLM scaling literature (see the reference example's Section 5.3), where optimizing against an imperfect confidence signal eventually degrades performance as the base policy approaches optimality.

Real-World Experiments: Strong OOD Gains Validate Counterfactual Generalization

In-distribution tasks (Table 4). π0-FAST-DROID + MG-Select* achieves 47.9% average success across four pick-and-place configurations, compared to 37.5% for the base model — a 28% relative improvement. Per-task breakdown:

Box to Bowl: 41.7% → 58.3% (+16.6 points)
Box to Plate: 37.5% → 54.2% (+16.7 points)
Basket to Bowl: 45.8% → 50.0% (+4.2 points)
Plate to Basket: 25.0% → 29.2% (+4.2 points)

The gains are concentrated on the first two tasks. The paper does not speculate on why, but it may relate to task geometry: "Box to Bowl" and "Box to Plate" require grasping from a confined space (the box) where precision matters more, while "Basket to Bowl" and "Plate to Basket" start from more open configurations. Or it may simply reflect noise given the modest trial count (24 per task).

Out-of-distribution tasks (Table 3). π0-FAST-DROID + MG-Select (without joint training — note the absence of *) achieves 71.9% average success vs. 53.1% for the base model — a 35% relative improvement. Per-task:

Pick up Tape: 56.3% → 68.8% (+12.5 points)
Take Cup out of Bowl: 50.0% → 75.0% (+25.0 points)

These are the paper's most compelling results for the "no external modules required" claim. The OOD tasks involve objects (a roll of tape, a lighter cup) that were NOT seen during fine-tuning, so the base model must generalize from its DROID pre-training. MG-Select is applied here without joint training (the model was not fine-tuned with condition dropout for these specific objects), yet it still provides substantial gains. This directly supports the claim that condition-masking confidence generalizes to unseen conditions, in contrast to external verifiers which "fail to generalize to unseen input conditions" (Section 1).

However, there is a subtle circularity risk: the MG-Select configuration (masking variant, temperatures) was presumably tuned on the OOD tasks themselves (the paper says hyperparameters are swept per dataset per Appendix A.3), which means the reported numbers may reflect some degree of test-set optimization. The paper does not specify whether the OOD hyperparameter search used a separate validation split or was done directly on the 16-trial test set — if the latter, the 71.9% may overstate generalization to truly unseen scenarios.

The qualitative results in Figure 2 provide visual evidence for the precision-gap mitigation narrative. The base model's gripper is shown approaching the sponge but failing to close at the right moment (Figure 2a), while MG-Select successfully grasps it. Similarly, the base model releases the sponge beside the bowl while MG-Select releases it inside (Figure 2b). These are exactly the "millimeter-level" precision failures the paper motivates in Section 1.

Ablation Studies and Robustness Checks

All ablations in Table 5 use RoboCasa with 100 demonstrations and π0-FAST as the base model. Temperature τ is fixed at 0.5 for sampling unless otherwise noted. Results are averaged over 50 trials and 3 random seeds.

Inference strategy (Table 5a): Greedy decoding achieves 28.5% on pick-and-place (42.7% on All tasks). Single-sample stochastic sampling (τ = 0.5, N = 1) achieves 27.6% (43.8% All) — slightly worse on pick-and-place but slightly better overall, suggesting that temperature sampling alone does not systematically help or hurt; it diversifies the output at the cost of occasionally deviating from the memorized expert trajectory. Uniform-KL Best-of-N (Kang et al., 2025) with N = 4 achieves 30.0% (46.5% All), and Likelihood Best-of-N achieves 30.5% (46.8% All) — both improve over greedy, confirming that Best-of-N selection is beneficial regardless of the criterion, but the gains are modest (1.5–2.0 points). MG-Select achieves 31.0% (48.1% All) — a further 0.5-point gain over Likelihood Best-of-N and 1.0 point over Uniform-KL. The absolute margins are small but consistent.

Number of candidates (Table 5b): Performance improves from N = 1 (27.6%) to N = 2 (30.0%) to N = 4 (31.0%), then plateaus or slightly declines: N = 8 (30.0%), N = 16 (30.7%). The paper concludes that "even a small number of samples is sufficient to generate diverse candidates and to yield meaningful precision gains." The non-monotonicity at N = 8 and the saturation at N = 4 suggest that adding more candidates eventually introduces noise — some high-scoring candidates may have spuriously high KL divergence without corresponding action quality, a mild over-optimization effect. This is consistent with the LIBERO-Spatial degradation observed when the base model is already strong.

Condition-masking variants (Table 5c): On RoboCasa (diverse tasks), text-masking achieves 31.0% (48.1% All) — the best. State-masking: 30.1% (46.7% All). Both-masking: 29.7% (46.3% All). The rank ordering (text > state > both) supports the paper's claim that the optimal variant depends on the environment: in a multi-task setting, removing the instruction creates the most informative uncertainty baseline because the model cannot infer the task from visual context alone. The gap between text-masking (31.0%) and Uniform-KL (30.0% from Table 5a) is exactly 1.0 percentage point — modest but consistent with the claim that a task-aligned reference provides a better uncertainty baseline than uniformity.

Effect of joint training (Table 5d): This is the most informative ablation. Four configurations are compared:

No joint training, no MG-Select (vanilla IL): 17.0% (40.2% All)
No joint training, MG-Select applied: 22.6% (43.7% All)
Joint training, no MG-Select (greedy decoding): 28.5% (42.7% All)
Joint training, MG-Select applied: 31.0% (48.1% All)

The deltas isolate three effects:

Joint training as a regularizer: 28.5% − 17.0% = +11.5 points. This is the largest single effect — training with condition dropout substantially improves the base model's performance even without test-time selection, likely by preventing overfitting to the limited demonstration data.
MG-Select without joint training: 22.6% − 17.0% = +5.6 points. Even with a suboptimal reference distribution (masked inputs are OOD for a non-jointly-trained model), MG-Select provides a meaningful gain.
MG-Select with joint training: 31.0% − 28.5% = +2.5 points. The incremental benefit of MG-Select on top of an already-regularized model is smaller but still positive.

The interaction is noteworthy: joint training and MG-Select are complementary but not purely additive. Joint training alone provides ~68% of the total improvement from the combined approach (11.5 out of 14.0 total points gained), while MG-Select adds the remaining ~32%. This suggests that the regularization benefit of condition dropout is at least as important as the confidence-based selection mechanism itself — a nuance the paper acknowledges implicitly by reporting both "MG-Select" and "MG-Select*" variants in all tables.

Regularization temperature (Table 5e): The paper sweeps τ_reg ∈ {0.5, 1.0, 2.0, 4.0, 8.0} on pick-and-place (with N = 4, text-masking, jointly trained model). Results: τ_reg = 0.5 → 27.5%, 1.0 → 28.8%, 2.0 → 25.4%, 4.0 → 31.0%, 8.0 → 30.0%. The U-shaped curve confirms the "sweet spot" hypothesis: the raw condition-masked distribution (τ_reg = 1.0, 28.8%) underperforms the uniform baseline (30.0%), while moderate regularization (τ_reg = 4.0, 31.0%) outperforms it. The performance collapse at τ_reg = 2.0 (25.4%, worse than greedy at 28.5%) is striking and not discussed — it suggests the reference distribution goes through a "worst of both worlds" regime where it is too peaked to represent uncertainty but too flat to retain task structure. At τ_reg = 0.5, the distribution is even more peaked than at τ_reg = 1.0, and performance is worse (27.5% vs. 28.8%), consistent with the peaked-reference problem.

Aggregation strategy (Table 5f): Summing all tokens: 26.1% (44.5% All) — the worst. Averaging all tokens: 24.7% (44.7% All) — worse than greedy (28.5%). First 5 tokens: 31.0% (48.1% All) — the best. First 10 tokens: 26.6% (45.1% All) — worse than greedy. The gap between first-5 (31.0%) and average-all (24.7%) is 6.3 points — larger than the gap between MG-Select and any baseline in Table 5a. Aggregation strategy is the single most impactful design choice in the entire method, more important than the choice of masking variant (text vs. state) or regularization temperature. The paper's explanation — that early FAST tokens encode macro-level action decisions while later tokens introduce noise — is plausible but not directly validated; an ablation showing the per-token KL divergence values for successful vs. failed actions would strengthen this claim.

Single-prefill deployment (Figure 3, Table 10): This is not a performance ablation but a latency engineering optimization. With N = 4, vanilla MG-Select latency is 43.4 seconds vs. 23.7 seconds for single-prefill — a 45% reduction. With N = 16, the reduction is even larger (76.0s → 30.4s, 60% reduction). The single-prefill variant brings MG-Select's latency close to single-action inference (20.2s at N = 1). This is crucial for practical deployment: without it, MG-Select would be too slow for real-time control at typical robot control frequencies. The paper does not evaluate whether single-prefill affects action quality (it shouldn't, since the prefill computation is identical across candidates, but this is assumed rather than verified).

Critical Assessment

Claim 1: MG-Select provides consistent, training-free performance improvement over single-inference VLAs.

What was tested: The paper evaluates MG-Select (with and without joint training) across four benchmarks and two VLA architectures. Without joint training, MG-Select improves RoboCasa pick-and-place from 17.0% to 22.6% (100 demos, Table 5d) — a 5.6-point gain. With joint training (MG-Select*), the improvement is larger: 17.0% → 31.0% on RoboCasa (100 demos), 37.5% → 47.9% on real-world ID tasks (Table 4), 69.2% → 70.8% on LIBERO with OpenVLA (Table 6).

What was NOT tested: The "without additional training" claim applies only to the plug-and-play variant (MG-Select without joint training), yet the headline numbers throughout the paper use MG-Select* (with joint training). The plug-and-play gains are measured only on RoboCasa at 100 demos (Table 5d: +5.6 points) and on real-world OOD tasks (Table 3: +18.8 points on average). The OOD result is the strongest evidence for the "no additional training" claim, since the model was not fine-tuned on these specific objects. However, the OOD evaluation has only 16 trials per task — a very small sample — and the MG-Select configuration was hyperparameter-tuned on these same tasks (Appendix A.3), creating test-set optimization concerns. There is no evaluation of plug-and-play MG-Select on SIMPLER-WidowX or LIBERO, so the claim of "consistent" improvement without joint training rests on thin evidence.

Verdict: The claim holds with qualifications. MG-Select* (with joint training) consistently improves performance across diverse settings, but this requires modified training — it is not a drop-in inference method. The plug-and-play variant (no joint training) shows promising gains on OOD tasks but has not been systematically evaluated across benchmarks. The "without additional training" framing in the abstract and introduction is somewhat misleading given the paper's heavy reliance on joint training for its best results.

Claim 2: Condition-masking distributional confidence (KL divergence from masked reference) is a more effective selection criterion than likelihood or uniform-KL baselines.

What was tested: Table 5(a) directly compares MG-Select (31.0%) against Likelihood Best-of-N (30.5%) and Uniform-KL Best-of-N (30.0%) on RoboCasa pick-and-place with 100 demos. MG-Select outperforms both, but the margins are small: 0.5 points over Likelihood, 1.0 point over Uniform-KL. These comparisons use the jointly trained model and N = 4.

What was NOT tested: The comparison is only made on a single benchmark (RoboCasa, 100 demos). There is no equivalent ablation for SIMPLER-WidowX, LIBERO, or real-world tasks. It is possible that on other benchmarks, the baselines catch up or surpass MG-Select — especially on LIBERO-Spatial where MG-Select* actually degrades performance relative to the base model (π0-FAST: 97.4% → 97.2%; OpenVLA: 85.2% → 81.7%). The paper does not report whether Likelihood or Uniform-KL Best-of-N would also degrade on these tasks, or whether they would maintain the base model's performance. Without this comparison, we cannot attribute the degradation specifically to MG-Select's confidence metric versus a general Best-of-N issue.

Verdict: The claim is supported on the single benchmark tested, but the evidence is weak. A 0.5-1.0 percentage point advantage over the baselines, measured on one dataset with 50 trials per task and 3 seeds, may not be statistically significant (the paper reports no error bars). The claim would be much stronger with comparable ablations across all benchmarks.

Claim 3: MG-Select generalizes to OOD conditions where external verifiers fail.

What was tested: Table 3 shows MG-Select (without joint training) improving π0-FAST-DROID from 53.1% to 71.9% on two OOD tasks involving unseen objects. This is an 18.8-point improvement — larger than the ID improvement (37.5% → 47.9%, +10.4 points).

What was NOT tested: The paper does not compare MG-Select against an external verifier on these OOD tasks. The claim that external verifiers fail under distribution shift is cited from prior work (Nakamoto et al., 2024; Kwok et al., 2025) but not replicated or directly compared. We do not know whether a value function trained on DROID data would also improve or degrade on these OOD tasks — the paper asserts that external verifiers fail, but provides no empirical head-to-head comparison. Furthermore, the OOD evaluation has only 16 trials per task, and the MG-Select hyperparameters were swept on these same 16 trials. This is a fragile basis for a generalization claim.

Verdict: The claim is directionally supported — MG-Select does provide substantial gains on OOD tasks without task-specific retraining — but the evidence is weak due to small sample sizes, lack of verifier baselines, and potential test-set optimization. The paper's framing of this as a key advantage over external verifiers, while conceptually plausible, is not empirically validated against actual verifiers.

Claim 4: Joint training improves performance both through regularization and by enabling better reference distributions.

What was tested: Table 5(d) isolates joint training's effects. Joint training alone (without MG-Select) improves from 17.0% to 28.5% — the regularization effect. Adding MG-Select on top improves to 31.0% — the reference distribution effect. The decomposition is clean and well-measured.

What was NOT tested: The paper does not ablate the dropout rate (fixed at 10% per masking variant for RoboCasa/LIBERO). Would 5% dropout provide similar regularization with less interference to standard IL? Would 20% improve the reference distribution further? There is no exploration of whether the 10%/10%/10% split (70% full-condition, 30% masked) is optimal or just a heuristic first guess. Additionally, the paper does not compare joint training to other regularization techniques (weight decay, data augmentation, early stopping) to determine whether the benefit is specific to condition dropout or achievable through any regularizer.

Verdict: The claim is well-supported for the specific dropout configuration used, but the generality of joint training as a method (versus other regularizers) is untested. The decomposition of gains into regularization (~68%) and MG-Select (~32%) is a valuable insight that the paper underemphasizes.

Missing Experiments That Would Strengthen the Paper

Verifier comparison on OOD tasks. The paper's central motivating claim — that external verifiers fail under distribution shift — would be much stronger with a direct comparison. Even a simple verifier (e.g., an ORM trained on DROID success/failure) evaluated on the OOD tasks would substantiate the claim that MG-Select's generalization advantage is real.
Statistical significance reporting. None of the tables include standard deviations, confidence intervals, or statistical tests. With 50 trials per task and 3 seeds, the error bars are likely substantial — a 0.5-point difference between MG-Select and Likelihood Best-of-N (Table 5a) may not be statistically distinguishable from zero. Reporting variance would clarify which differences are meaningful and which are noise.
Ablation of sampling temperature. The paper sweeps τ during hyperparameter optimization but does not report a dedicated ablation. Table 5(a) compares sampling (τ = 0.5, N = 1) against greedy, but there is no sweep of τ across values for Best-of-N selection. The interaction between sampling diversity and selection criterion is central to the method — a more diverse candidate set might favor different criteria — but remains unexplored.
Candidate diversity measurement. The paper claims that likelihood-based selection fails because the candidate distribution lacks diversity due to over-concentration, but never actually measures diversity (e.g., average pairwise distance between candidates, entropy of the empirical candidate distribution, or fraction of unique actions among N samples). Quantifying diversity would transform the over-concentration claim from a plausible hypothesis into a demonstrated mechanism.
Per-token KL analysis. The finding that aggregating over the first 5 tokens works best (Table 5f) is attributed to the FAST tokenizer's frequency ordering, but no evidence is provided. Visualizing per-token KL divergence for successful vs. failed actions would validate (or refute) the hypothesis that early tokens encode task-relevant confidence while later tokens introduce noise.
LIBERO baseline comparisons. The negative results on LIBERO-Spatial and LIBERO-Goal (Table 6) are reported but not explained. Running Likelihood Best-of-N and Uniform-KL Best-of-N on these suites would clarify whether the degradation is specific to MG-Select's confidence metric or a general Best-of-N phenomenon when the base model is near-ceiling.

Genuine Weaknesses

Hyperparameter tuning on test sets. The paper sweeps all hyperparameters (τ, N, masking variant, τ_reg) directly on each benchmark's evaluation set (Appendix A.3), reporting the best result. This is standard practice in some robotics venues but inflates reported performance relative to a held-out validation protocol. For the real-world OOD tasks with only 16 trials, the risk of overfitting to the test set is particularly acute.
Small absolute gains in several settings. On SIMPLER-WidowX (46.9% → 50.3%, Table 2), on π0-FAST LIBERO (92.0% → 93.1%, Table 6), and on RoboCasa with 300 demos (43.2% → 46.9% on pick-and-place, Table 1), the absolute improvements from MG-Select* are 3.4, 1.1, and 3.7 percentage points respectively. Whether these gains justify the additional inference compute (even with single-prefill optimization) depends on the deployment context — for near-ceiling tasks like LIBERO-Spatial, MG-Select may not be worth the overhead.
Degradation on high-performing tasks. MG-Select* reduces performance on LIBERO-Spatial for both π0-FAST (−0.2 points) and OpenVLA (−3.5 points) and on LIBERO-Goal for both architectures (−1.1 and −1.9 points). The paper does not discuss these regressions or provide guidance on when NOT to use MG-Select. A practical deployment would need a mechanism to detect when the base model is sufficiently strong that Best-of-N selection introduces more noise than signal.
Single model family for main results. The primary architecture throughout is π0-FAST (Paligemma-3B backbone). OpenVLA (Prismatic-7B) is evaluated only on LIBERO, and only with MG-Select* (no plug-and-play variant, no Likelihood/Uniform-KL baselines). The claim of architecture-agnostic effectiveness is supported directionally by the OpenVLA results but would be stronger with evaluation on additional VLA architectures (e.g., RT-2, Octo) and with the full suite of baselines.
No latency-quality Pareto analysis. The single-prefill optimization (Figure 3) shows that MG-Select can be made latency-efficient, but the paper does not provide a latency vs. success rate tradeoff curve. For deployment, practitioners need to know: at what N does the marginal latency cost outweigh the marginal success gain? The saturation at N = 4 (Table 5b) provides a partial answer for RoboCasa, but the latency dimension is not integrated into that analysis.

6. Limitations and Trade-offs

1. Hyperparameter Optimization Directly on Test Sets Inflates Reported Performance

The paper sweeps all deployment hyperparameters—sampling temperature τ, number of candidates N, masking variant, and regularization temperature τ_reg—directly on each benchmark's evaluation set, reporting the best configuration as the final result. Appendix A.3 states:

"We search for the optimal configuration on each dataset within the following ranges: τ ∈ {0.1, 0.3, 0.5, 0.7, 1.0}, N ∈ {4, 8}, variants ∈ {text, state, text&state}, and regularization temperature ∈ {4.0, 6.0, 8.0, 10.0, 12.0, 14.0, 16.0}, and report the best result for each policy."

No held-out validation set is used for hyperparameter selection; the sweep is performed on the same data used to report final performance. This is a significant methodological weakness because the reported numbers represent best-case performance after optimization on the test distribution rather than expected performance on unseen data.

The consequence. Every reported accuracy number for MG-Select and MG-Select* is upwardly biased relative to what would be achieved under a proper train/validation/test split. The bias is largest where test sets are smallest: the real-world OOD experiments use only 16 trials per task (Table 3), and SIMPLER-WidowX uses 24 trials per task (Table 2). With a hyperparameter search space of 7 × 2 × 3 × 7 = 294 possible configurations (τ × N × variant × regularization temperature), the probability of finding a configuration that performs well on a 16-trial test set by chance alone is non-trivial. This is not a minor concern—it directly affects the strength of the central claim that MG-Select provides consistent, verifier-free improvement.

What evidence exists. The paper provides no cross-validation results, no validation-set performance curves, and no discussion of hyperparameter sensitivity across datasets. The single exception is the averaging over 3 random seeds for MG-Select variants, but these seeds control for training stochasticity, not for hyperparameter selection overfitting. The degradation on LIBERO-Spatial (−3.5 points for OpenVLA, Table 6) is consistent with what happens when an overfit hyperparameter configuration from other suites fails to transfer, but the paper does not explore or acknowledge this possibility.

Mitigation status. Not addressed. The appendix describes the sweep protocol without qualification. A standard fix—selecting hyperparameters on one task suite or a held-out split and evaluating on another—is never performed. The authors do not flag this as a limitation or suggest that future work should adopt proper validation protocols. For practitioners, this means the reported gains should be treated as optimistic upper bounds rather than expected deployment performance, particularly for the real-world OOD results where dataset sizes are smallest.

2. Plug-and-Play Claims Are Undermined by Reliance on Joint Training

The paper frames MG-Select as requiring "no additional training" and leveraging "the model's internal properties without requiring additional training or external modules" (Section 1). Yet the headline results in every table use MG-Select*, which incorporates joint training—a modified fine-tuning procedure that randomly drops input conditions (text, state, or both) during imitation learning. The plug-and-play variant (MG-Select without joint training) is evaluated in only two settings: RoboCasa at 100 demonstrations (Table 5d: 17.0% → 22.6%, a 5.6-point gain) and real-world OOD tasks (Table 3: 53.1% → 71.9%, an 18.8-point gain).

The consequence. A practitioner hoping to deploy MG-Select on an existing, already-fine-tuned VLA without retraining faces two uncertainties. First, the plug-and-play gain of 5.6 points on RoboCasa, while positive, is roughly one-third of the total gain achieved with joint training (17.0% → 31.0%, 14.0 points). Second, the plug-and-play variant has not been evaluated on SIMPLER-WidowX, LIBERO, or real-world ID tasks at all—there is no evidence that it works in those settings. The "no additional training" claim is empirically supported only for the narrow setting where it has been tested, and the paper's own data shows that joint training supplies roughly two-thirds of the total improvement on the one benchmark where the decomposition is measured.

The real-world OOD result (71.9% vs. 53.1%) is the paper's strongest evidence for plug-and-play effectiveness, but it is confounded by the hyperparameter-overfitting concern described above (16 trials per task) and by the absence of a comparison to what joint training would achieve on these same tasks. We do not know whether joint training would push OOD performance from 71.9% to, say, 85%, or whether it would not help at all—so we cannot assess how much of the plug-and-play gain is "real" versus an artifact of the small test set and hyperparameter sweep.

What evidence exists. Table 5(d) provides the only direct comparison between MG-Select with and without joint training, and only on RoboCasa at one data scale (100 demos). The paper does not ablate the plug-and-play variant on any other benchmark, and does not discuss whether the OOD gains in Table 3 would transfer to other OOD scenarios or scale with task diversity.

Mitigation status. Partially acknowledged but not resolved. The paper distinguishes "MG-Select" and "MG-Select*" notationally in all tables, which is transparent. However, the abstract and introduction emphasize the "no additional training" framing without equivalent emphasis on the fact that best results require joint training. The paper does not provide guidance on when plug-and-play is sufficient versus when joint training is necessary, nor does it explore whether the regularization benefit of joint training (which accounts for most of the gain according to Table 5d) could be achieved through simpler means (weight decay, early stopping, data augmentation) that would not require modifying the training pipeline.

3. Degradation on High-Performance Tasks Lacks Diagnostic Explanation

MG-Select* reduces performance on two of four LIBERO suites for both tested architectures. On LIBERO-Spatial: π0-FAST drops from 97.4% to 97.2% (−0.2 points), and OpenVLA drops from 85.2% to 81.7% (−3.5 points). On LIBERO-Goal: π0-FAST drops from 95.6% to 94.5% (−1.1 points), and OpenVLA drops from 75.5% to 73.6% (−1.9 points). These regressions appear in Table 6 but are never discussed in the main text.

The consequence. The paper provides no guidance on when MG-Select should NOT be applied. A deployment engineer evaluating whether to integrate this method into a production system has no way to predict whether it will help or hurt on their specific task distribution. The pattern—degradation on suites where the base model already achieves ~85%+ accuracy—suggests a threshold effect: MG-Select may be beneficial when the base model is weak (LIBERO-Object: 63.7% → 72.5%, LIBERO-Long: 52.5% → 55.4% for OpenVLA) but counterproductive when it is already strong. However, this threshold is not characterized, and the RoboCasa results complicate the picture: at 300 demonstrations, the base model achieves 43.2% on pick-and-place, yet MG-Select* still improves to 46.9%—so the relationship is not purely about absolute performance.

The likely mechanism is confidence signal miscalibration in the high-performance regime. When the base model's full-condition distribution already places very high probability on the correct action, the KL divergence from a condition-masked reference may be spuriously large for plausible-but-suboptimal actions (because the reference distribution, even when regularized, may assign low probability to tokens the full-condition distribution assigns high probability). The selection mechanism then preferentially chooses these spuriously high-confidence candidates, degrading performance. This is analogous to the verifier over-optimization phenomenon documented in LLM test-time scaling (see the reference example), where beam search degrades easy-problem performance when the verifier signal is imperfect.

What evidence exists. The LIBERO results in Table 6 are the only evidence. The paper does not report whether Likelihood Best-of-N or Uniform-KL Best-of-N also degrade on these suites—without this comparison, we cannot distinguish between "MG-Select's specific confidence metric is harmful" and "any Best-of-N selection introduces noise when the base model is near-optimal." The RoboCasa ablation in Table 5(b) shows non-monotonic scaling at N = 8 (performance drops from 31.0% at N = 4 to 30.0% at N = 8), which is consistent with over-optimization but is not discussed in those terms.

Mitigation status. Not addressed. The paper reports the average improvement across suites (92.0% → 93.1% for π0-FAST, 69.2% → 70.8% for OpenVLA) without acknowledging or explaining the per-suite regressions. A practical mitigation—using a fallback to greedy decoding when the confidence score variance among candidates is low, or when the base model's likelihood exceeds a threshold—is not explored or suggested. For practitioners, the lack of diagnostic criteria for when to deploy MG-Select is a significant gap: without such criteria, the method may silently degrade performance on well-learned tasks.

4. Aggregation Strategy Is Environment-Sensitive and Poorly Understood

The aggregation of per-token KL divergences into a single action-level confidence score is the most impactful design choice in the entire method—larger even than the choice of masking variant. Table 5(f) shows that on RoboCasa pick-and-place, truncating to the first 5 tokens achieves 31.0%, while naive summation over all tokens achieves only 26.1% (worse than greedy decoding at 28.5%). The gap between the best and worst aggregation strategy (6.3 percentage points) exceeds the gap between MG-Select and any baseline comparison in Table 5(a).

The consequence. The paper attributes the optimality of first-5-token aggregation to the structure of the FAST tokenizer, which orders tokens from low-frequency (coarse movements) to high-frequency (fine adjustments). However, this explanation is a hypothesis, not a validated principle. No experiment directly measures whether early tokens genuinely encode more task-relevant confidence, or whether the result is specific to this tokenizer, this task distribution, or this model. For a practitioner using a different tokenizer (e.g., OpenVLA's fixed-length encoding, which uses average aggregation per Appendix A.3), there is no guidance on how to choose I, the token index set for aggregation. The choice between "first K tokens," "average all," and "sum all" is treated as a hyperparameter to be swept per dataset—meaning the method's performance depends on a design choice that the paper does not provide a principled way to make without access to evaluation labels.

This is not a minor implementation detail. The aggregation strategy determines which signal the Best-of-N selection actually optimizes. If the wrong strategy is chosen, MG-Select can perform worse than the base model (as average aggregation does: 24.7% vs. 28.5% greedy). The lack of a transferable principle means each new deployment environment requires either (a) a labeled validation set to sweep aggregation strategies, or (b) blind trust that the optimal strategy from one environment transfers to another—neither of which is satisfactory for a method marketed as "verifier-free."

What evidence exists. Table 5(f) provides the ablation on RoboCasa with 100 demonstrations. No equivalent ablation exists for any other benchmark. The paper does not visualize per-token KL divergence values, does not correlate early-token vs. late-token confidence with task success, and does not test whether the "first 5 tokens" rule transfers across environments. The OpenVLA results use average aggregation "since its output sequence length is fixed to the action dimension of the training data" (Appendix A.3), but this is a default choice driven by architectural constraints, not an empirically validated strategy.

Mitigation status. Not addressed as a limitation. The paper treats aggregation strategy as a dataset-specific hyperparameter, on par with temperature and candidate count. The underlying issue—that the method's core signal (per-token confidence) is composed of components with unknown and environment-dependent relevance—is never acknowledged. Future work on tokenizer-aware confidence aggregation, or on learning which token positions are predictive of action quality, could address this gap, but the paper does not propose such directions.

5. Single Model Family and Narrow Task Scope Limit Generality Claims

The vast majority of experiments use π0-FAST (Pertsch et al., 2025) as the base VLA, with Paligemma-3B as the vision-language backbone. OpenVLA (Kim et al., 2024) with Prismatic-7B is evaluated only on LIBERO, and only with MG-Select* (no plug-and-play variant, no Likelihood or Uniform-KL baselines). All tasks across all benchmarks are pick-and-place or pick-and-place-adjacent manipulation tasks: grasping objects, moving them between locations, stacking, and opening/closing containers. There are no navigation tasks, no dexterous manipulation tasks, no multi-step tasks requiring tool use or sequential subtask reasoning beyond the built-in action chunk horizon.

The consequence. The paper's claims of generality—"consistently improves state-of-the-art VLAs across diverse pick-and-place tasks and various environments" (Section 1), "compatible with different architectures" (Section 4.1.2)—are supported only within this narrow scope. A practitioner working with a different VLA architecture (e.g., RT-2, Octo, GR00T N1), a different VLM backbone, or a different task family (e.g., mobile manipulation, assembly, deformable object handling) has no empirical basis to expect MG-Select to work. The method's core mechanism—KL divergence from a condition-masked reference—depends on the VLA being an autoregressive model that produces token-level probability distributions amenable to temperature scaling and KL computation. Diffusion-based VLAs (Black et al., 2025; Bjorck et al., 2025) do not have this property, and the paper's framework does not obviously extend to them.

Moreover, the reliance on pick-and-place tasks may be systematically favorable to the method. Pick-and-place has exactly two high-precision contact points (grasp and release), and the paper's qualitative results (Figure 2) and ablation (Table 5f, first-5-tokens aggregation) suggest that MG-Select works by identifying actions that are confident on these critical transition points. Tasks requiring sustained precision throughout an action trajectory—precision peg insertion, suturing, soldering—might not benefit from a confidence signal concentrated in early tokens. The paper provides no evidence either way.

What evidence exists. The paper claims architecture-agnosticism based on OpenVLA's average improvement from 69.2% to 70.8% on LIBERO (Table 6). However, this masks the per-suite regressions discussed above (Limitation 3), and OpenVLA was evaluated only with MG-Select* and without baseline comparisons. The paper does not report whether the 1.6-point average gain is statistically significant, and does not test whether the gain persists if hyperparameters are transferred from π0-FAST rather than optimized on OpenVLA's test set.

Mitigation status. The paper acknowledges the scope implicitly by describing tasks as "diverse pick-and-place tasks" (Section 1) but does not discuss the limitation to autoregressive architectures or the potential brittleness of results across task families. The suggestion that MG-Select "contributes to establishing a general test-time scaling paradigm for improving robustness and precision in VLAs" (Section 6) overstates the demonstrated generality. A broader evaluation across architectures (including diffusion-based VLAs, potentially via an adaptation of the confidence metric) and task families would be needed to support this claim.

6. Joint Training's Regularization Benefit Is Not Disentangled from the Confidence Mechanism

Table 5(d) reveals that joint training alone—without any test-time selection—improves RoboCasa pick-and-place performance from 17.0% to 28.5% (an 11.5-point gain). Adding MG-Select on top of joint training yields a further improvement to 31.0% (a 2.5-point gain). This means joint training as a regularizer accounts for ~82% of the total improvement over vanilla imitation learning, while MG-Select's confidence-based selection mechanism accounts for only ~18%. On the "All" tasks column, the decomposition is even starker: joint training alone: 40.2% → 42.7% (+2.5 points); MG-Select on top: 42.7% → 48.1% (+5.4 points). Here the relative contributions are roughly 32% / 68%, but the key pattern remains: a substantial portion of MG-Select*'s performance comes from the modified training procedure, not from the test-time selection logic.

The consequence. The paper's central narrative—that condition-masking distributional confidence enables effective test-time action selection—is partially confounded by the regularization effect of joint training. It is possible that any regularizer (weight decay, dropout on visual features, data augmentation) would produce a similarly improved base model, and that the incremental benefit of MG-Select's specific confidence metric over Likelihood Best-of-N (30.5% vs. 31.0%, Table 5a) is achievable without condition-masking at all—for instance, by using Uniform-KL Best-of-N (30.0%) on a model regularized by simpler means.

If this is the case, then the method's complexity—joint training with three masking variants, computing per-token KL divergences, applying a regularization temperature, truncating to the first 5 tokens—may be unnecessary. A practitioner could achieve most of the gain by simply training with input dropout and using a simpler Best-of-N criterion (likelihood or uniform-KL). The paper does not test this hypothesis, so we cannot assess whether MG-Select's elaborate confidence mechanism provides value beyond what a well-regularized model with a simple selection criterion would achieve.

What evidence exists. Table 5(d) provides the decomposition. Table 5(a) shows that on the jointly trained model, MG-Select (31.0%) outperforms Likelihood (30.5%) and Uniform-KL (30.0%), but the margins are small (0.5–1.0 points). The paper does not ablate whether these margins persist with different regularization methods, different dropout rates, or simpler reference distributions.

Mitigation status. Not addressed. The paper presents the joint training as an enabler for MG-Select ("enables the model to learn both conditional and unconditional distributions... thereby further improving the quality of the reference distribution," Section 3.3) without considering that the primary mechanism might be regularization rather than reference distribution quality. The ablation in Table 5(d) is correctly reported, but its implications for the method's core contribution are not discussed. A control experiment—comparing MG-Select on a jointly trained model against Likelihood Best-of-N on the same jointly trained model across all benchmarks, not just RoboCasa—would clarify whether the confidence metric adds value beyond the training procedure, but is only performed on one benchmark.

7. Implications and Future Directions

How This Work Changes the Landscape

MG-Select introduces a counterfactual framing for test-time action selection that shifts the problem from "how good is this action?" (a quality estimation problem requiring external critics) to "how much does this action depend on knowing what to do?" (a dependence measurement problem computable from the model's own internal distributions). This is not a paradigm shift on the scale of VLAs themselves or the invention of test-time scaling — Best-of-N sampling was already established in both LLMs and robotics (Nakamoto et al., 2024; Kwok et al., 2025). Rather, it is a well-motivated reframing with practical advantages that eliminates the distribution-shift vulnerability inherent in external verifiers without requiring a separate training pipeline.

The paper's most concrete contribution to the landscape is the demonstration that self-generated confidence signals can substitute for learned verifiers on pick-and-place tasks, particularly when the base model is jointly trained with condition dropout. The 35% OOD improvement (Table 3) and the 168% relative gain in the low-data regime (Table 1, 30 demonstrations) establish floor numbers for what verifier-free test-time scaling can achieve on precision-sensitive manipulation. These numbers do not prove that MG-Select outperforms verifier-based approaches — no head-to-head comparison is provided — but they establish that verifier-free methods are competitive enough to warrant serious investigation, especially for deployment scenarios where training a separate value function is impractical.

The work also provides a diagnostic lens on the over-concentration problem in behavior-cloned VLA action distributions. Prior to this paper, it was known that behavior cloning produces peaked distributions (it follows directly from maximizing likelihood on expert data), but the connection to test-time scaling failure — that peaked distributions eliminate candidate diversity, making Best-of-N selection vacuous regardless of the selection criterion — was not explicitly articulated in the VLA literature. Table 5(a)'s modest gap between Likelihood Best-of-N (30.5%) and Uniform-KL Best-of-N (30.0%) provides empirical evidence for this diagnosis: when the distribution is concentrated, even crude selection criteria perform similarly because all candidates are nearly identical. MG-Select's confidence metric partially circumvents this by measuring dependence on task conditions rather than absolute likelihood, but the paper's own data shows the gains are modest (0.5–1.0 percentage points over baselines on RoboCasa with 100 demos), suggesting that the over-concentration problem is a fundamental bottleneck that no selection criterion can fully solve without first improving candidate diversity.

A subtle but important reframing the paper introduces is treating the reference distribution as an entropy-calibrated baseline rather than a fixed anchor. The regularization temperature ablation (Table 5e) demonstrates that the quality of the confidence signal depends critically on the entropy of the reference distribution — too peaked and it undermines the purpose of measuring departure from uncertainty; too flat and it reduces to the uniform-KL baseline, losing the benefit of task alignment. This principle — that learned reference distributions for confidence estimation must be explicitly regularized toward high entropy — may generalize to other domains where self-certainty measures are applied, including LLM reasoning and multimodal alignment. The paper does not develop this into a theoretical framework, but the empirical curve (U-shaped performance vs. τ_reg, with a clear optimum at 4.0) provides a concrete starting point for such development.

The paper also reconciles a tension in the test-time scaling literature between generality and specialization. External verifiers (Nakamoto et al., 2024) specialize to their training distribution and fail on OOD conditions. Uniform-KL self-certainty (Kang et al., 2025) is fully general but ignores domain structure. MG-Select's condition-masking approach occupies a middle ground: the reference distribution is task-aligned (it respects the VLA's action manifold) but does not require task-specific verifier training. The OOD results (Table 3) validate this middle-ground claim — the method generalizes to novel objects because the confidence computation inherits whatever generalization the VLA possesses, without requiring the verifier to independently generalize. This does not resolve the generality-specialization tension in general, but it demonstrates that partially structured self-signals can outperform both fully unstructured self-signals (uniform-KL) and fully specialized external signals (verifiers) in specific regimes.

Research directions that become more attractive after this work:

Input-perturbation-based confidence for other embodied domains. The condition-masking principle — remove semantically meaningful inputs, measure divergence — is trivially portable to any domain with structured input modalities. For navigation, mask the goal location. For dexterous manipulation, mask tactile feedback. For multi-agent coordination, mask partner observations. This paper provides the template and the regularization-temperature methodology.
Joint training with input dropout as a general VLA regularizer. Table 5(d) shows that condition dropout alone improves RoboCasa pick-and-place from 17.0% to 28.5% — an 11.5-point gain that exceeds MG-Select's incremental contribution. Whether this regularization benefit generalizes to other VLAs, task families, and data scales is an open question with immediate practical relevance, since adding dropout to an existing fine-tuning pipeline is trivial.
Tokenizer-aware confidence aggregation. The finding that truncating to the first 5 FAST tokens works best (Table 5f) suggests that confidence signals are not uniformly distributed across action representations — coarse action components carry more task-relevant confidence than fine components. This opens a line of investigation into tokenizer design for test-time scaling: can action tokenizers be designed to concentrate confidence signals in a predictable prefix?

Research directions that become less attractive:

Training increasingly sophisticated external verifiers for pick-and-place tasks. If a method requiring no verifier training achieves 35% OOD improvement and 168% low-data gains, the marginal benefit of a complex RL-trained value function over the much simpler MG-Select baseline needs to be demonstrated, not assumed. The burden of proof shifts to verifier-based methods to show that their additional complexity and distribution-shift vulnerability are justified by substantially larger gains.
Likelihood-based Best-of-N as a default test-time scaling strategy for VLAs. The paper's diagnosis of over-concentration — and the empirical result that Likelihood Best-of-N barely outperforms Uniform-KL Best-of-N (30.5% vs. 30.0%, Table 5a) — suggests that likelihood is a weak selection signal for behavior-cloned policies. Researchers developing test-time methods for imitation-learned policies should default to confidence-based criteria rather than likelihood, at least until candidate diversity can be independently verified.

Follow-Up Research This Work Enables

Head-to-head comparison of MG-Select against external verifiers on shared OOD benchmarks. The paper's central motivating claim — that external verifiers fail under distribution shift — is cited from prior work but never empirically tested against MG-Select in a controlled comparison. A strong follow-up would train a value function (following Nakamoto et al., 2024) and a VLM-based reward model (following Kwok et al., 2025) on the same DROID fine-tuning data used for the real-world ID experiments, then evaluate all three methods — MG-Select (plug-and-play), MG-Select* (joint training), and the external verifier — on the OOD tasks (roll of tape, lighter cup) using a proper held-out protocol with sufficient trials (50+ per task). The key measurement: does the verifier's performance degrade on OOD tasks relative to ID tasks, and if so, by how much compared to MG-Select's OOD-to-ID gap? This would transform the paper's conceptual claim into an empirical finding and provide practitioners with concrete guidance on when verifier training is worth the overhead.

Candidate diversity as a mediating variable: when does Best-of-N help? The paper hypothesizes that over-concentration limits Best-of-N effectiveness but never measures diversity directly. A diagnostic study would instrument the VLA at multiple stages of fine-tuning (0, 1k, 5k, 10k steps on RoboCasa with 100 demos) and measure: (a) the average pairwise action distance among N candidates (in the VLA's latent space or in end-effector delta space), (b) the entropy of the empirical candidate distribution, and (c) the correlation between diversity metrics and Best-of-N gain for each selection criterion (Likelihood, Uniform-KL, MG-Select). The prediction: diversity decreases with fine-tuning steps, and the Best-of-N gain for all criteria drops as diversity approaches zero. If confirmed, this would establish diversity as the key gating variable for test-time scaling in VLAs and motivate diversity-preserving fine-tuning objectives (e.g., entropy regularization, multi-modal behavior cloning) as a prerequisite for effective test-time selection. If disconfirmed (gains persist despite low diversity), it would suggest that MG-Select's confidence signal captures something beyond mere candidate differentiation — perhaps calibration quality — that warrants deeper investigation.

Tokenizer-aware confidence weighting via per-token success prediction. The paper's finding that first-5-token aggregation dramatically outperforms summing or averaging all tokens (Table 5f: 31.0% vs. 26.1% vs. 24.7%) is attributed to the FAST tokenizer's frequency ordering but never validated. A follow-up would train a lightweight success classifier that takes as input the per-token KL divergences C_1, C_2, ..., C_T for an action sequence and predicts whether that action will succeed, using a held-out validation set of rollouts with success labels. The learned weights on each token position would reveal which tokens genuinely carry predictive confidence signal, independently of the tokenizer's frequency ordering. If the learned weights concentrate on the first 5 tokens, the paper's hypothesis is confirmed and the truncation strategy is validated. If the weights are distributed across tokens or peak at unexpected positions, it would suggest that the first-5 heuristic is specific to RoboCasa or to the particular FAST tokenizer configuration, and a learned aggregation would generalize better across environments and tokenizers.

Scaling MG-Select to diffusion-based and continuous-action VLAs. The paper's method is exclusive to autoregressive VLAs that produce token-level categorical distributions amenable to KL divergence computation. Diffusion-based VLAs (π0, GR00T N1) and continuous-action policies do not fit this framework. An extension study would adapt the condition-masking confidence principle to these architectures by measuring the Wasserstein distance or MMD between the full-condition and condition-masked action distributions in continuous action space, or by measuring the variance of the diffusion denoising trajectory under input perturbation. The evaluation would replicate the RoboCasa and SIMPLER-WidowX experiments using the diffusion-based π0 model (Black et al., 2025), comparing the adapted MG-Select against the standard single-inference baseline. Success would demonstrate that the counterfactual confidence principle is architecture-agnostic, not tied to autoregressive tokenization. Failure would clarify the scope of applicability and suggest that autoregressive VLAs have a structural advantage for test-time scaling that diffusion-based models lack.

Dynamic condition-masking: selecting the optimal mask per timestep based on task phase. The paper treats the masking variant (text, state, both) as a static hyperparameter per environment, but the optimal reference distribution likely depends on the robot's current task phase. During the approach phase, state-masking (removing proprioceptive information) might be most informative because the model needs to know gripper position to plan the grasp approach. During the grasp phase, text-masking (removing the instruction) might be more informative because the model needs task context to decide when to close the gripper. A dynamic variant would, at each timestep, compute KL divergences for all three masking variants and use a learned meta-selector (trained on a small set of phase-labeled demonstrations) to choose which variant's confidence score to use for Best-of-N selection. The evaluation would compare dynamic masking against static masking on tasks with clearly separable phases (pick-and-place, where approach, grasp, transport, and release are distinct). The prediction: dynamic masking outperforms static masking, particularly on tasks where different phases require different types of precision (e.g., grasp requires state-dependent precision, release requires instruction-dependent precision).

Cost-benefit analysis of joint training across data scales and task families. Table 5(d) shows that joint training alone accounts for ~68% of the total improvement on RoboCasa with 100 demos. Is this ratio stable across data scales? At 30 demonstrations, does joint training's regularization effect dominate even more (since overfitting risk is higher)? At 300 demonstrations, does it become negligible (since the base model already generalizes well)? A systematic study would evaluate MG-Select, MG-Select*, joint-training-only, and Likelihood Best-of-N on joint-training-only across 3–5 data scales per benchmark, decomposing the contribution of regularization vs. confidence-based selection at each scale. The practical payoff: a decision rule for when the added complexity of MG-Select's confidence mechanism is justified over simply training with input dropout and using a simpler selection criterion. If the incremental benefit of MG-Select over Likelihood Best-of-N on a jointly trained model is consistently 0.5–1.0 points as in Table 5(a), practitioners with limited compute budgets might reasonably choose the simpler approach.

Practical Applications and Downstream Use Cases

Low-data fine-tuning of generalist VLAs for specific manipulation tasks. The paper's strongest result is the 168% relative improvement on RoboCasa pick-and-place with only 30 demonstrations (Table 1: 5.3% → 14.2%). This directly addresses a common robotics pain point: collecting hundreds of demonstrations per task is expensive and time-consuming, but fine-tuning on small datasets produces brittle policies. A robotics lab with a pre-trained generalist VLA (e.g., π0-FAST or OpenVLA) and a budget for only 20–50 demonstrations per new task can integrate MG-Select* — joint training with condition dropout followed by test-time Best-of-N selection — as a standard post-fine-tuning step. The 11.5-point regularization gain from joint training alone (Table 5d) provides immediate value even without the full MG-Select pipeline, and the additional 2.5-point gain from confidence-based selection is effectively free in terms of data collection cost. The single-prefill optimization (Figure 3) keeps inference latency within ~17% of single-action timing, making this practical for real-time control at standard manipulation cycle rates.

OOD-robust deployment of fine-tuned policies in unstructured environments. The 35% OOD improvement on novel objects (Table 3: 53.1% → 71.9%) is the paper's most practically significant result for real-world deployment. A warehouse robot fine-tuned to manipulate a specific set of objects (boxes, bins, pallets) will inevitably encounter novel objects (damaged packaging, unexpected items, new SKUs). MG-Select without joint training — the plug-and-play variant evaluated in Table 3 — can be deployed as a drop-in inference wrapper around the existing fine-tuned policy, requiring no retraining and no additional data collection. The wrapper generates N = 4 candidate actions per timestep, selects the one with the highest text-masking or state-masking confidence, and executes it. The 18.8-point absolute gain on OOD tasks suggests this could meaningfully reduce failure rates on novel objects without any modification to the underlying policy. The key practical caveat is the hyperparameter selection problem (Limitation 1) — the OOD evaluation tuned hyperparameters on the test tasks themselves, so a deployment engineer would need either a small validation set of OOD-like objects or conservative default hyperparameters (τ = 0.5, N = 4, text-masking, τ_reg = 4.0 based on RoboCasa defaults) rather than expecting to reproduce the full 35% gain without task-specific tuning.

Batch evaluation and data filtering for self-improvement pipelines. When using VLAs to generate training data for themselves (rejection sampling, STaR-style self-improvement), the quality of automatically labeled data determines whether the self-improvement loop is virtuous or degenerative. MG-Select provides a self-generated quality signal that can filter generated trajectories without requiring a separately trained success classifier. In a self-improvement pipeline: (1) the VLA generates M = 20 trajectories per task, (2) MG-Select scores each trajectory using aggregated condition-masking confidence across all timesteps, (3) the top K = 5 trajectories by confidence score are added to the training set with "success" labels, and (4) the VLA is fine-tuned on the augmented dataset. The paper does not evaluate this pipeline, but the consistent improvement of MG-Select over baselines on RoboCasa and real-world tasks suggests the confidence signal is correlated with action quality. The risk — that confidence-based filtering amplifies distribution collapse by selecting trajectories the model is already confident about rather than diverse successful trajectories — would need to be empirically assessed.

Graceful degradation detection for autonomous systems. The negative results on LIBERO-Spatial and LIBERO-Goal (Table 6: −0.2 to −3.5 points) are, paradoxically, a useful signal for deployment. If a system monitors MG-Select's confidence scores and detects that the score distribution across candidates has low variance (suggesting over-concentration and potential Best-of-N degradation), it can fall back to greedy decoding or escalate to a human operator. Specifically: at each timestep, compute the standard deviation of confidence scores across the N candidates. If this standard deviation falls below a calibrated threshold (determined on a validation set where base model performance is known to be strong), skip Best-of-N and use the greedy action (single-sample at τ → 0). This converts MG-Select's failure mode into a diagnostic — a low-variance confidence signal indicates the model is in a "well-learned" regime where additional compute is more likely to harm than help. The paper does not develop this mechanism, but the data in Tables 5(b) and 6 provide the necessary signal to design it.

🎯 Pitch

1. Executive Summary

2. Context and Motivation

The Core Problem: VLAs Fail When They Can Only Guess Once

Why Simple Solutions Don't Work

The Existing Solution and Its Limitations: External Verifiers

The Paper's Positioning: Verifier-Free Test-Time Scaling

The Joint Training Requirement and Why It Matters

The Specific Failure Modes the Method Targets

Why This Problem Matters Now

3. Technical Approach

3.1 Reader Orientation

3.2 Big-Picture Architecture (Diagram in Words)

3.3 Roadmap for the Deep Dive

3.4 Detailed, Sentence-Based Technical Breakdown

The Test-Time Scaling Framework

Why Likelihood-Based Selection Fails for VLAs

Condition-Masking Distributional Confidence: The Core Metric

Constructing the Reference Distribution via Condition-Masking

Aggregation Strategy: Which Tokens Matter Most?

Regularization Temperature for the Reference Distribution

Joint Training Strategy: Teaching the Model to Be Uncertain

Choice of Masking Variant per Environment

Efficient Deployment: Single-Prefill Strategy

Summary of Design Choices and Their Justifications

4. Key Insights and Innovations

Innovation 1: Reframing Test-Time Action Selection as a Counterfactual Confidence Problem

Innovation 2: Diagnosing and Circumventing the Over-Concentration Problem in Behavior-Cloned Action Distributions

Innovation 3: The Regularized Reference Distribution as an Entropy-Calibrated Uncertainty Baseline

5. Experimental Analysis

Evaluation Methodology

Main Quantitative Results

RoboCasa: Strong Gains in Low-Data Regime, Consistent Improvement Across Scales

SIMPLER-WidowX: Consistent but Modest Improvement

LIBERO: Architecture-Agnostic Gains, Most Pronounced on Hardest Tasks

Real-World Experiments: Strong OOD Gains Validate Counterfactual Generalization

Ablation Studies and Robustness Checks

Critical Assessment

Claim 1: MG-Select provides consistent, training-free performance improvement over single-inference VLAs.

Claim 2: Condition-masking distributional confidence (KL divergence from masked reference) is a more effective selection criterion than likelihood or uniform-KL baselines.

Claim 3: MG-Select generalizes to OOD conditions where external verifiers fail.

Claim 4: Joint training improves performance both through regularization and by enabling better reference distributions.

Missing Experiments That Would Strengthen the Paper

Genuine Weaknesses

6. Limitations and Trade-offs

1. Hyperparameter Optimization Directly on Test Sets Inflates Reported Performance

2. Plug-and-Play Claims Are Undermined by Reliance on Joint Training

3. Degradation on High-Performance Tasks Lacks Diagnostic Explanation

4. Aggregation Strategy Is Environment-Sensitive and Poorly Understood

5. Single Model Family and Narrow Task Scope Limit Generality Claims

6. Joint Training's Regularization Benefit Is Not Disentangled from the Confidence Mechanism

7. Implications and Future Directions

How This Work Changes the Landscape

Follow-Up Research This Work Enables

Practical Applications and Downstream Use Cases

Adjacent in domain

Low-probability Tokens Sustain Exploration in Reinforcement Learning with Verifiable Reward

QeRL: Beyond Efficiency -- Quantization-enhanced Reinforcement Learning for LLMs