acceptodds
Under review as a conference paper at ICLR 2027

A Safe Action Is Not Enough: Feasible-Future Decoding for Vision-Language-Action Policies

Abstract

A safe action is not necessarily a viable one. Under a frozen vision-language-action (VLA) policy, an action can be likely and locally admissible yet leave no policy-supported route to safe task completion. We call this the feasibility–likelihood gap: likelihood ranks the current action, whereas feasibility depends on the futures that remain after it. We derive the exact next-block marginal of the history-conditioned policy–environment trajectory law restricted to safe task completion. The derivation exposes a candidate-dependent feasible-future mass with two roles: its support records whether safe completion remains possible under the frozen continuation process, while its magnitude measures how much weighted safe-completion mass is preserved. Exact evaluation is impractical online, so we develop a selective finite-candidate approximation, derive conditions for recovering the best retained viable candidate, and instantiate it as an alarm-triggered, training-free reranker. On Safety-CHORES, VICS-G lowers mean cumulative safety cost by (1.9%)–(57.5%) across six settings while remaining within (2.5) percentage points of policy sampling in success and (0.82) steps in mean episode length. The resulting decoder is tied to an exact policy-relative safe-completion target, yet requires neither retraining of the base policy nor online trajectory rollouts.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-language-action models, robot safety, constrained decoding, inference-time control, policy-relative viability, safe task completion.

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_hWgROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the gap between local action safety and long-term task feasibility in VLA policies. The authors argue that an action may be locally safe and preferred by the policy, but still lead to states from which safe task completion is unlikely. They formalize this idea through feasible-future mass and derive conditional guarantees for finite-candidate decoding. Since computing the exact quantity is impractical, they propose VICS-G, a training-free decoder that selectively reranks candidate actions using policy likelihood, local robustness, and history-based corrections.

Experiments on Safety-CHORES show reductions in cumulative safety cost across six settings while maintaining success rates close to those of the original policy.

Critical strengths
  1. Well-motivated problem formulation.

I find the distinction between local safety and future feasibility meaningful, particularly for robotic tasks where an initially safe action can make later progress difficult. The feasibility–likelihood gap provides a useful perspective on the limitations of existing action selection strategies. Rather than treating safety as a property of individual actions, the paper emphasizes whether a decision preserves the possibility of eventual safe task completion.

  1. Principled and relatively thorough theoretical analysis.

The feasible-future formulation accounts for both future policy behavior and environment dynamics, including execution uncertainty. The distinction between feasibility support and magnitude is interesting, as it separates whether safe completion remains possible from how much probability mass is assigned to those continuations. I also appreciate the analysis of different approximation errors, including candidate coverage, support rejection, and ranking errors, as well as the authors' acknowledgment of the conditional nature of their guarantees.

  1. Practical and computationally efficient approach.

VICS-G does not require retraining the underlying VLA policy or performing online trajectory rollouts. The selective intervention mechanism is also appealing, as it preserves the original policy's action when no intervention is triggered rather than modifying every decision. This makes the method relatively lightweight and potentially useful in settings where retraining or extensive inference-time computation is undesirable.

  1. Promising empirical results and careful evaluation of safety–progress trade-offs.

Across six Safety-CHORES settings, VICS-G reduces cumulative safety cost by 1.9%–57.5%, while keeping task success within 2.5 percentage points of policy sampling. I appreciate that the authors evaluate safety cost together with task success and episode length, since lower safety cost alone can be misleading when a policy terminates prematurely. The additional ablations, execution-feedback experiments, and rollout-based study also provide useful insight into the behavior and limitations of the proposed approach.

Critical weaknesses
  1. The connection between the theoretical formulation and the practical decoder is not sufficiently established.

The central theoretical quantity is feasible-future mass, which depends on the safe task-completing continuations available after taking an action. However, the continuation correction in VICS-G (Eq. 46) is based on fixed action preferences, recent repetitions, and failed or collision-related retries. While the authors acknowledge that this is a heuristic rather than a calibrated estimate, it remains unclear whether it actually captures future feasibility. Furthermore, the theoretical recovery guarantees rely on score accuracy and support separation conditions that are not verified for the deployed decoder. This leaves a gap between the theoretical motivation and the practical method.

  1. The incremental benefit of the continuation correction remains unclear.

While the full VICS-G system achieves encouraging results, the zero-weight ablation in Appendix G provides less consistent evidence for the contribution of the continuation correction itself. At the nominal weight, all confidence intervals for changes in safety cost and violations include zero. On ObjectNav-base, adding the correction even decreases success while increasing cost and violations. The rollout-based experiment in Appendix H is interesting, but its improvement in safe success is not statistically significant and does not directly validate the heuristic used in VICS-G. These results make it difficult to determine how much of the improvement is attributable to the proposed future-feasibility correction.

  1. The baseline comparisons do not fully isolate the source of improvement.

RCD and VICS-G differ in several components beyond the continuation score, including the intervention gate, trust region, and replacement authorization. Comparing the complete systems is reasonable, but these differences make it difficult to attribute the observed improvements specifically to future-aware decoding. Although the zero-weight ablation helps address this concern, the Fetch ablation uses a different candidate configuration from the main evaluation. A simple retry-avoidance baseline under the same intervention mechanism would help clarify whether the proposed correction offers benefits beyond conventional history-based heuristics.

  1. Lack of real-world validation and uncertain practical applicability.

The experiments are conducted entirely in simulation, without real-robot evaluation. This is a concern given the paper's focus on safe robotic action execution. VICS-G relies on monitoring signals such as simulator-provided safety costs, collision feedback, object visibility, and target distance, which may be difficult to obtain reliably on physical robots. Although the authors investigate execution perturbations in Appendix I, these experiments use relatively simple action-magnitude perturbations and do not address uncertainty in the monitoring signals themselves. Without real-world validation or a more realistic robustness analysis, it remains unclear whether the reported safety improvements would transfer to physical robotic systems.

Questions for authors
  1. Could the authors provide a more direct validation of the continuation correction zGz_G? In particular, how well does its ranking of candidate actions agree with their actual safe-completion probabilities under the frozen policy? Appendix H provides a rollout-based comparison, but it does not directly evaluate whether the heuristic used in VICS-G captures the theoretical feasible-future mass.

  2. Appendix G provides a useful zero-weight ablation, but the contribution of the continuation correction remains unclear. Could the authors further explain why adding zGz_G leads to worse performance on ObjectNav-base? Also, would the conclusions for Fetch remain the same under the top-12 candidate configuration used in the main experiments, rather than the all-20-action configuration used in the ablation?

  3. Appendix F describes the monitoring signals available to the decoder, including simulator-provided safety costs and target information. How would these signals be obtained or approximated in real-world deployment, and how sensitive is the method to their availability?

5 comments

We thank the reviewer for recognizing the theoretical analysis and practical efficiency of VICS. We particularly appreciate the emphasis on distinguishing the complete decoder's performance from the incremental contribution of its continuation approximation.

W1 + Q1: “The connection between the theoretical formulation and the practical decoder is not sufficiently established [...] it remains unclear whether it actually captures future feasibility.”

The key question is whether the practical correction captures future feasibility rather than merely encoding recovery heuristics.

This is precisely the approximation problem our theoretical analysis addresses. Theorem 1 identifies policy-relative feasible-future mass in the exact action marginal; Theorem 2 specifies sufficient conditions for recovering viable candidates despite coverage, support, and ranking errors.

VICS-G deliberately uses lightweight history-based signals instead of repeatedly simulating continuations. We do not assume that zGz_G is a calibrated estimate of feasible-future mass.

We further strengthen the theory–implementation connection through VICS-R, introduced in Appendix H, which explicitly evaluates simulated continuations. Its expanded evaluation on 180 previously unseen Fetch Safe houses (Table R1 in our shared response) shows 51.7% versus 47.2% task success, 24.4% versus 21.7% joint safe success, and lower mean cost (1.85 versus 2.23) relative to VICS-G at nominal weight.

These results provide stronger episode-level evidence for future-aware decoding beyond history-based recovery. VICS-R more directly operationalizes the feasible-future principle through prospective safe-completion estimates. Since its continuations follow VICS-G, we do not claim exact estimation of the frozen-policy ZfeasZ_{\mathrm{feas}} or direct validation of zGz_G.

W2 + Q2: “The incremental benefit of the continuation correction remains unclear [...] the zero-weight ablation [...] provides less consistent evidence.”

We agree that the continuation correction must be evaluated separately from the complete decoder. Appendix G already provides this controlled comparison, holding the gate, candidate selection, local-risk scoring, and fallback fixed while varying only ηv\eta_v.

ObjectNav-base is the clearest unfavorable case. At ηv=0.10\eta_v=0.10, task success decreases by 3.0 percentage points relative to zero weight (paired bootstrap interval [−6.0, −0.5]), while mean cost and violations increase by 0.91 and 1.09, respectively, with intervals spanning zero. We report this result rather than suggesting uniform improvement.

However, the incremental effect of zGz_G should be distinguished from the performance of the complete decoder. In the main ObjectNav-base evaluation (Table 1), nominal VICS-G reduces mean safety cost by 57.5% relative to policy sampling (7.965 → 3.385), while task success remains within 2.5 percentage points (78.0% → 75.5%) and joint safe success slightly increases (60.0% → 60.5%).

Importantly, we use a single shared nominal configuration across all six settings, rather than selecting the best continuation weight for each task. The ObjectNav-base result thus reveals a genuine limitation of the lightweight correction, not task-specific optimization.

The new held-out Fetch Safe evaluation provides direct paired nominal-versus-zero comparisons under the matched top-12 configuration (Table R5 in Part 2/2). Within VICS-G, the correction yields favorable changes (+0.56 pp task success, −0.21 cost, −0.12 violations).

W3: “The baseline comparisons do not fully isolate the source of improvement [...] a simple retry-avoidance baseline under the same intervention mechanism would help.”

We appreciate this suggestion and have now added the requested retry-only control (Table R1 in our shared response), complementing the existing zero-weight ablation in Appendix G.

The controls address different questions. Setting ηv=0\eta_v=0 removes the additional zGz_G correction but retains the gate, local-risk scoring, filtering, and replacement authorization. Retry-only more directly examines whether avoiding previously ineffective actions explains the observed results.

Retry-only matches nominal VICS-G at reported episode-level precision. We therefore do not claim that non-retry preferences independently improve performance in this Fetch setting.

VICS-R provides complementary evidence through explicit continuation evaluation, although a statistically established direct VICS-R–VICS-G effect would require its own paired comparison.

We additionally evaluate SafeDec-style kinematic hard-constrained decoding (HCD) (Table R2 in our shared response). Relative to HCD, nominal VICS-G achieves 47.2% versus 25.6% task success, 21.7% versus 11.1% joint safe success, and lower cost (2.23 versus 5.49).

W4 + Q3: “Lack of real-world validation and uncertain practical applicability [...] the experiments [...] do not address uncertainty in the monitoring signals themselves.”

We agree that monitoring requirements deserve direct examination. Beyond the imperfect-execution study in Appendix I, we now evaluate simulator-cost masking (Table R3 in our shared response).

With numerical simulator-cost feedback withheld, nominal VICS-G retains 47.2% task success and 21.7% joint safe success, while achieving 13.7% lower mean cost than policy sampling. It preserves approximately 77% of its full-monitor cost advantage.

This demonstrates that favorable observed behavior does not depend entirely on numerical simulator-cost feedback. Other signals—including target monitoring and recent execution information—remain available. We do not equate this experiment with robustness to arbitrary monitoring errors or physical-robot validation.

On the overall rating

The revised experiments directly address the requests for retry-only controls, expanded explicit lookahead, a hard-constrained comparator, and restricted-monitoring evaluation. Together with the conditional approximation theory, they provide a clearer account of what the framework establishes and what remains dependent on the practical heuristic.

We respectfully ask the reviewer to reconsider the weak rejection in light of this more targeted evidence.

AuthorsOpenReview ↗

Follow-up on Our Responses and Revisions

Dear Reviewer hWgR,

Our detailed responses and additional experiments address the theory–implementation connection, continuation correction, baseline attribution, and monitoring limitations. The supporting evidence is presented in our shared experimental comments and revised manuscript (highlighted in purple).

We believe these additions substantially strengthen the paper. We would particularly value your thoughts on whether they help resolve your key concerns and welcome further discussion of any points that remain unclear.

Reviewer_N5cjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies action selection for frozen VLA policies, emphasizing that a locally safe action may still leave no policy-supported route to safe task completion. It derives an ideal action distribution involving feasible-future mass and proposes VICS-G, a training-free decoder that selectively reranks actions using policy likelihood, local risk, and a history-based correction. Across three Safety-CHORES task families and two checkpoint types, the complete decoder reduces observed safety costs while largely preserving task success and episode length.

Critical strengths

The distinction between local safety and policy-supported safe completion is useful. The formulation and analysis of intervention, candidate coverage, filtering, and ranking clarify why a locally admissible action need not preserve safe task completion. This provides a helpful perspective on the limitations of inference-time safety interventions.

Critical weaknesses
  1. Limited methodological novelty and insufficient evidence for the central correction. VICS-G is a relatively simple combination of gating, rule-based filtering, fixed action preferences, and repetition/failure penalties. Its added contribution over existing recovery and reranking mechanisms is not sufficiently established. In particular, the practical correction z_G is not shown to approximate feasible-future mass or its induced ranking. In Appendix G, all reported confidence intervals for changes in safety cost and violations include zero, while success decreases in some settings. The Fetch ablation also uses all 20 actions rather than the main experiment's top-12 candidate budget. These results do not establish that the central correction accounts for the complete decoder's gains.

  2. Dependence on monitoring signals limits the evidence for practical applicability. The decoder uses simulator-provided safety costs, failure and collision feedback, and target visibility and distance. These signals influence intervention and action selection, but it is unclear how they would be obtained reliably outside simulation or how missing or inaccurate signals would affect performance. The execution-perturbation experiments do not directly resolve this concern.

Questions for authors
  1. With the same candidate budget as the main experiments, what are the paired effects of adding z_G to the zero-weight variant? Can the results distinguish its contribution from that of gating, filtering, and the history-dependent local score?

  2. Could the authors report the fraction of episodes that both complete the task and satisfy the designated safety budget in all six main settings? This joint outcome directly matches the theoretical objective.

  3. Which monitoring signals are essential, which rely on simulator information, and how are they expected to be obtained in deployment?

7 comments

We thank the reviewer for recognizing the importance of policy-supported safe completion and for raising questions about methodological novelty, monitoring requirements, and joint safety–completion outcomes.

W1 + Q1: “Limited methodological novelty and insufficient evidence for the central correction [...] its added contribution over existing recovery and reranking mechanisms is not sufficiently established.”

We appreciate this concern and would like to distinguish the contribution of our feasible-future framework from the isolated effect of its lightweight heuristic correction.

Our principal contribution is to formulate policy-relative feasible-future decoding: Theorem 1 derives the exact continuation-dependent action factor, while Theorem 2 characterizes selective recovery under candidate coverage, support, and ranking errors.

VICS-G provides a training-free instantiation that preserves the frozen policy and intervenes selectively. Its history-based correction is deliberately lightweight, requiring neither learned critics nor online rollouts.

We recognize that the benefit of this heuristic must be distinguished from the complete decoder. The retry-only control matches nominal VICS-G at reported precision (Table R1 in our shared response); therefore, we do not claim an independently established advantage from non-retry preferences.

The new direct paired ablation (Table R5 in Part 2/2) shows favorable but statistically inconclusive incremental effects of zGz_G within VICS-G. Nevertheless, the complete rollout-free decoder achieves statistically supported reductions in cost and violations relative to policy sampling on 180 unseen Fetch Safe houses (Table R4), without retraining or online rollouts.

Importantly, VICS-R, introduced in Appendix H, provides a complementary, theory-grounded realization through explicit future-looking feasibility estimation. Its expanded held-out evaluation demonstrates statistically supported incremental improvements in task success, safety cost, and violations under the matched nominal-versus-zero comparison, with all three paired intervals excluding zero (Table R5). Its simulated continuations follow VICS-G rather than the original frozen policy, so we do not claim exact estimation of ZfeasZ_{\mathrm{feas}}.

Thus, the methodological contribution extends beyond a repetition penalty: it encompasses the exact feasible-future formulation, its conditional approximation analysis, and two complementary decoding mechanisms connecting future feasibility to action selection.

W2 + Q3: “Dependence on monitoring signals limits the evidence for practical applicability [...] it is unclear how they would be obtained reliably outside simulation or how missing or inaccurate signals would affect performance.”

We agree that monitoring availability is an important practical concern and have added a targeted experiment. In Table R3 in our shared response, simulator-provided numerical safety costs are withheld from the decoder. Target visibility, distance, recent execution outcomes, and collision-related signals remain available, while the benchmark evaluator independently measures safety cost.

Under this restriction, nominal VICS-G retains 47.2% task success and 21.7% joint safe success, with 13.7% lower mean cost than policy sampling (2.34 versus 2.71). Relative to its full-monitor cost of 2.23, it preserves approximately 77% of its observed cost advantage.

These results demonstrate favorable safety–progress outcomes without numerical simulator-cost feedback. On physical robots, perception and localization could estimate target visibility and distance, while task-progress monitoring and onboard sensors could detect execution failures and collisions. The reliability of these signals and sensitivity to their errors remain important directions for hardware validation.

Q2: “Could the authors report the fraction of episodes that both complete the task and satisfy the designated safety budget in all six main settings?”

We have added joint safe success to the main results table (revised Table 1) for all six task/checkpoint settings, defined as task completion with zero cumulative safety cost.

SettingPolicy samplingVICS-G
PickUp / base76.25%76.25%
PickUp / safe83.75%85.00%
ObjectNav / base60.00%60.50%
ObjectNav / safe82.00%82.00%
Fetch / base18.02%17.44%
Fetch / safe28.49%29.65%

VICS-G matches or exceeds policy sampling's observed joint safe success in five of six settings, with Fetch-base the exception. This metric complements the reported task-success, cost, violation, and episode-length results by directly measuring safe task completion.

On the overall rating

The revised main table now directly answers the requested six-setting safety–completion question. The additional explicit-lookahead and restricted-monitoring studies address the other principal concerns, while the theoretical formulation establishes the methodological contribution beyond the heuristic.

We respectfully hope the reviewer will reconsider the weak rejection in light of these additions.

AuthorsOpenReview ↗

Follow-up on Our Responses and Revisions

Dear Reviewer N5cj,

We have provided detailed responses and additional evidence on methodological novelty, the continuation correction, and monitoring requirements, including the requested joint safety–completion results. The full evidence is available in our shared experimental comments, with corresponding manuscript revisions highlighted in purple.

We believe these additions substantially strengthen the paper. We would appreciate your feedback on whether they address your main concerns and would be happy to discuss any points that remain open.

Reviewer_eaA3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper derives a policy-relative safe-completion target for frozen VLA policies, showing how next-action selection depends on the mass of feasible future trajectories. It analyzes approximation errors in selective decoding and proposes VICS-G, a training-free, rollout-free reranker combining policy likelihood, local risk, and a history-based continuation heuristic. Across six Safety-CHORES settings, VICS-G reduces observed safety cost while largely preserving task success and episode length.

Critical strengths
  1. Clear theoretical formulation. The distinction between local safety, policy-supported feasibility, and continuation magnitude is useful. The analysis explicitly identifies intervention, coverage, support, and ranking errors, while appropriately acknowledging its conditional guarantees.
  2. Promising safety–progress trade-offs. VICS-G lowers observed cost across all six settings, with success within 2.5 percentage points of policy sampling. Jointly reporting success, cost, violations, and episode length makes the evaluation informative.
Critical weaknesses
  1. The connection between the feasible-future target and the practical heuristic remains insufficiently validated. The continuation term in Eq. (46) consists of fixed action preferences and repetition/failure penalties, which overlap with features already included in the local-risk function in Eq. (47). The authors appropriately acknowledge its heuristic nature, but do not directly evaluate whether it improves candidate ranking toward the theoretical target. The support-separation and relative-score accuracy conditions underlying Theorem 2 are also not established for VICS-G. Consequently, the evidence remains insufficient to determine whether the method addresses locally safe dead ends beyond improving reactive failure recovery.
  2. The incremental benefit of the continuation correction is not convincingly established. In the matched zero-weight ablation, all safety-cost and violation-difference confidence intervals include zero (Table 5). On ObjectNav/base, the nominal correction reduces observed success by 3.00 percentage points and increases mean cost and violations, although the latter differences have confidence intervals spanning zero. Furthermore, the Fetch ablation enumerates all 20 actions, whereas the main evaluation uses the top-12 actions plus the policy sample. The headline RCD comparison cannot resolve this attribution issue because RCD also differs in intervention gating, trust-region filtering, and replacement authorization. These results leave the correction’s contribution beyond the existing recovery mechanisms uncertain.
Questions for authors
  1. Can the authors provide candidate-level evidence that adding zGz_G improves agreement with the ideal ranking in Eq. (13), using controlled examples or continuation estimates under the original frozen policy? Evidence involving locally admissible actions that lead to dead ends would help distinguish feasible-future reasoning from repetition avoidance.
  2. Can the authors report matched ηv=0\eta_v=0 versus ηv=0.10\eta_v=0.10 results under the exact main configuration, particularly top-12 actions plus the policy sample for Fetch, on episodes excluded from coefficient selection? Paired confidence intervals for success and safety cost would help establish whether the correction provides a reliable incremental benefit.
4 comments

We thank the reviewer for recognizing the theoretical contribution and the importance of evaluating safety alongside task progress. We appreciate the questions concerning the practical approximation and its incremental benefit.

W1 + Q1: “The connection between the feasible-future target and the practical heuristic remains insufficiently validated [...] the evidence remains insufficient to determine whether the method addresses locally safe dead ends beyond improving reactive failure recovery.”

The associated question concerns whether the practical correction captures future feasibility rather than merely reacting to previous failures.

The distinction between an ideal continuation objective and an inexpensive approximation is central to our framework. Theorem 1 derives the feasible-future factor, while Theorem 2 characterizes the coverage, support, and ranking conditions required for practical recovery. VICS-G uses history-based signals as a lightweight proxy; we do not claim that zGz_G directly estimates safe-completion probability or certifies dead-end avoidance.

To strengthen the operational evidence, we expand the explicit-lookahead VICS-R evaluation on 180 previously unseen Fetch Safe houses (Table R1 in our shared response). At the nominal weight, VICS-R achieves 51.7% versus 47.2% task success, 24.4% versus 21.7% joint safe success, and lower cost (1.85 versus 2.23) relative to VICS-G.

Against policy sampling, its reported paired task-success difference is +5.00 percentage points [0.40, 9.60], with a cost difference of −0.86 [−1.74, −0.17] (Table R4 in our shared response).

Although VICS-R uses budget-limited simulator-conditioned rollouts rather than the exact frozen-policy continuation law, its improvements over VICS-G provide direct episode-level evidence that explicit lookahead adds value beyond history-based failure recovery. These results strengthen the empirical connection to our feasible-future formulation, demonstrating that prospective evaluation can improve both task success and safety cost under practical computational constraints.

W2 + Q2: “The incremental benefit of the continuation correction is not convincingly established [...] all safety-cost and violation-difference confidence intervals include zero.”

We agree that the continuation correction must be distinguished from the complete decoder. Appendix G already includes matched zero-weight ablations, holding the intervention architecture fixed while varying ηv\eta_v.

We now extend this comparison on the new Fetch Safe cohort under the main evaluation's K=12K=12 candidate budget (Table R1), rather than relying solely on the earlier all-20-action Fetch sweep.

Direct nominal-versus-zero comparison. On the same 180 unseen Fetch Safe houses, using the main top-12 candidate configuration, enabling zGz_G (ηv=0.10\eta_v=0.10 versus ηv=0\eta_v=0) changes VICS-G task success by +0.56 pp, joint safe success by 0.00 pp, mean cost by −0.21, and violations by −0.12 (Table R1 in our shared response). All other decoder components remain fixed.

Comparisons with policy sampling. At zero weight, VICS-G records paired cost and violation differences of −0.27 [−0.57, −0.05] and −0.16 [−0.34, −0.03], respectively. At nominal weight, these become −0.48 [−1.00, −0.09] and −0.28 [−0.60, −0.05] (Table R4). All four reported intervals exclude zero, although they do not establish statistical significance for the difference between weights.

Importantly, retry-only matches nominal VICS-G at reported episode-level precision. We therefore do not attribute the complete decoder's gains to an independently demonstrated benefit from non-retry preferences.

The evidence supports a more precise distinction: VICS-G provides inexpensive recovery-oriented decoding within the feasible-future framework, while VICS-R evaluates continuations explicitly and achieves stronger observed outcomes on this Fetch cohort.

On the overall rating

We hope the reviewer will reconsider the weak rejection in light of this substantially strengthened empirical connection. The paper provides a theoretically grounded objective and two practical realizations, now evaluated more directly with independent held-out data, controlled ablations, and paired uncertainty estimates.

AuthorsOpenReview ↗

Follow-up on Revisions and Remaining Questions

Dear Reviewer eaA3,

Our detailed responses and additional experiments address your questions about the theory–implementation connection and the incremental contribution of the continuation correction. The evidence is presented in our shared experimental comments and reflected in the revised manuscript (highlighted in purple).

We believe these additions substantially strengthen the paper. We would be interested to hear whether they address your principal concerns, and would welcome further discussion of any points you find unresolved.

Thank you for the additional experiments. The held-out top-12 evaluation addresses my concerns about candidate configuration and coefficient-selection overlap. However, two concerns remain:

  1. There is still no direct evidence that zGz_G improves candidate ranking toward Eq. (13). The retry-only control matches nominal VICS-G at reported precision, while VICS-R uses a different continuation policy and does not directly validate the rollout-free heuristic.
  2. Table R4 reports confidence intervals against policy sampling, rather than the matched zero-weight variant. Direct paired intervals between ηv=0.10\eta_v=0.10 and ηv=0\eta_v=0 are still needed to assess the correction’s incremental benefit.

The evaluation is stronger, but the correction’s connection to the theory and its independent contribution remain insufficiently established. I therefore maintain my rating.

Thank you for carefully considering our additional evidence and for continuing this constructive discussion. We appreciate your thoughtful engagement and the opportunity to clarify the remaining concerns.

The evaluation is stronger, but the correction's connection to the theory and its independent contribution remain insufficiently established.

We agree that the isolated incremental benefit of the handcrafted zGz_G is not statistically established within VICS-G. Nevertheless, VICS-G offers a highly practical, rollout-free solution with statistically supported episode-level safety improvements, while VICS-R provides a direct operational connection to our feasible-future theory through explicit continuation-based estimation.

This distinction is central to our contribution. Our theoretical framework concerns ZfeasZ_{\mathrm{feas}}, rather than any single heuristic approximation. Theorem 1 derives the exact policy-relative feasible-future objective, and Theorem 2 characterizes practical approximation and recovery conditions. VICS-G and VICS-R offer two complementary realizations: efficient history-based correction and theory-grounded, future-looking feasibility estimation.

1. Direct paired ablation (180 unseen Fetch Safe houses). To address your request, we compare ηv=0.10\eta_v=0.10 against ηv=0\eta_v=0 under the same top-12 candidate configuration plus the policy sample, holding all other settings fixed.

Table R5. Nominal-minus-zero differences [95% paired bootstrap CI].

OutcomeVICS-GVICS-R
Task success ↑+0.56 pp [0.00, +1.67]+2.22 pp [+0.56, +4.44]
Joint safe success ↑0.00 pp [0.00, 0.00]0.00 pp [0.00, 0.00]
Mean safety cost ↓−0.21 [−0.74, +0.21]−0.220 [−0.579, −0.004]
Mean violations ↓−0.12 [−0.35, +0.04]−0.240 [−0.625, −0.011]

These results highlight the complementary strengths of the two approaches. For VICS-G, enabling zGz_G yields favorable changes, although its paired intervals include zero. For VICS-R, the nominal weight yields statistically supported incremental improvements in task success, safety cost, and violations, with all three intervals excluding zero. These findings complement the complete-decoder improvements already reported in Table R4 of our shared response.

2. Theory–implementation connection. We acknowledge that candidate-level agreement between the handcrafted zGz_G and the ideal ranking in Eq. (13) remains unestablished. .

Nevertheless, VICS-R is our more direct answer to the broader theory–implementation question through its explicit future-looking feasibility estimation. By incorporating simulated safe-completion outcomes into action selection, it operationalizes the feasible-future principle underlying ZfeasZ_{\mathrm{feas}}. Its continuations follow VICS-G rather than the original frozen policy, so we do not claim exact estimation of ZfeasZ_{\mathrm{feas}} or direct validation of the handcrafted zGz_G.

Our paper's independent contribution therefore extends beyond zGz_G: an exact feasible-future formulation, theoretical approximation analysis, a practical rollout-free decoder with strong episode-level evidence, and an explicit-lookahead decoder with statistically supported incremental gains.

We hope this distinction, together with the new paired evidence, helps address your concerns about the theory–implementation connection and the paper's independent contribution. We would greatly appreciate your further thoughts.

Reviewer_fXqsOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to select actions that preserve the possibility of safe task completion under a frozen vision-language-action policy. The authors derive an action-selection objective that accounts for safe future continuations and propose Viability-Corrected Sampling (VICS), a training-free method that selectively reranks candidate actions using policy likelihood, safety information, and execution history. Experiments on Safety-CHORES evaluate safety costs alongside task success and episode length.

Critical strengths
  1. The paper addresses a meaningful safety problem. It distinguishes between actions that are locally safe and those that preserve a route to safe task completion. This distinction provides a useful motivation for considering future consequences when selecting actions.
  2. The method is practical to deploy. VICS works with a frozen policy and requires neither additional training nor online trajectory rollouts.
  3. The evaluation considers both safety and task performance. The reported metrics help distinguish safety gains from reduced task progress, and the paper acknowledges limitations and unfavorable results.
Critical weaknesses
  1. The practical method is not clearly presented. The paper provides insufficient detail on how candidate filtering, scoring, and replacement decisions interact, making the procedure difficult to follow. Moreover, an algorithm block or method figure would be particularly helpful for readers interested in understanding the approach.

  2. The connection between the theory and implementation remains unclear. The theory evaluates actions based on whether they allow the policy to complete the task safely. However, the method assigns scores using predefined action preferences and penalties for repeated or previously failed actions. The paper does not explain why these scores reliably indicate whether safe task completion remains possible.

  3. The reported gains may depend on hyperparameters tuned to the evaluation tasks. The method’s configuration is selected using episodes that are also included in the final evaluation, which may overestimate its performance. Testing the fixed configuration on unseen tasks would help establish whether the gains generalize beyond the tuning data distribution.

  4. The baseline comparison remains incomplete. The paper describes RCD as adapted from SafeDec but does not clearly explain what was changed or why. These details are significant to assess the fairness of the comparison. Moreover, including additional baselines, such as hard-constrained decoding in SafeDec across the main settings, would further strengthen the evaluation.

Questions for authors

See weakness.

4 comments

We thank the reviewer for recognizing the paper's motivation, practical relevance, and joint evaluation of safety and task completion. We believe several methodological concerns relate to details already explained in the submission. The revised manuscript and supplementary experiments address the remaining empirical questions.

W1: “The paper provides insufficient detail on how candidate filtering, scoring, and replacement decisions interact [...] an algorithm block or method figure would be particularly helpful.”

The requested algorithm is provided in Algorithm 1 (Appendix F). It specifies candidate construction, intervention gating, admissibility filtering, scoring, replacement authorization, and fallback. Section 3 explains these components and their interactions.

The algorithm distinguishes when reranking is triggered, which candidates remain eligible, how they are scored, and when replacement is permitted. We agree that presenting it more prominently would improve accessibility. Nevertheless, Algorithm 1 already specifies the complete procedure, including how filtering, scoring, and replacement decisions interact.

W2: “The connection between the theory and implementation remains unclear. The theory evaluates actions based on whether they allow the policy to complete the task safely. However, the method assigns scores using predefined action preferences and penalties for repeated or previously failed actions. The paper does not explain why these scores reliably indicate whether safe task completion remains possible.”

Bridging the ideal feasible-future objective and a practical decoder is precisely the central challenge our work addresses.

Theorem 1 derives policy-relative feasible-future mass as the continuation-dependent factor in the exact next-action marginal (Eqs. 6–7). Computing it exactly requires integrating over future policy actions and environment dynamics. VICS-G therefore uses a lightweight history-based correction zGz_G to approximate its influence on ranking without retraining or online rollouts.

Importantly, we do not assume this heuristic reliably identifies feasibility in every state. Theorem 2 characterizes the support, coverage, and ranking conditions for recovery. Appendix E verifies the theoretical identities through finite examples, while Appendix G investigates the correction through controlled ablations.

We further strengthen the theory–implementation connection through VICS-R, introduced in Appendix H, which explicitly evaluates simulated continuations at one eligible decision. Its expanded held-out evaluation (Table R1 in our shared response) shows 51.7% versus 47.2% task success, 24.4% versus 21.7% joint safe success, and lower mean cost (1.85 versus 2.23) relative to VICS-G at nominal weight.

The new paired ablation (Table R5 in Part 2/2) further supports this: within VICS-R, the nominal weight improves task success by +2.22 pp [0.56, 4.44], reduces mean cost by 0.22 [0.004, 0.579], and violations by 0.24 [0.011, 0.625]. All three 95% paired intervals exclude zero.

Together, these results clarify the complementary roles of the two approaches: VICS-G provides practical, rollout-free decoding with statistically supported episode-level safety improvements, while VICS-R more directly connects theory to implementation through explicit estimation of prospective safe-completion outcomes. The latter strengthens the empirical evidence for future-aware action selection beyond history-based recovery alone.

W3: “The method's configuration is selected using episodes that are also included in the final evaluation, which may overestimate its performance.”

We agree this overlap warrants scrutiny. However, Section 4.1 discloses that one shared VICS-G configuration is used across all six settings, selected using approximately 10% of episodes per task family. We did not tune coefficients separately for each task/checkpoint result.

We evaluate the same fixed configuration on 180 previously unseen Fetch Safe houses, excluded from configuration selection (Table R1 in our shared response, revised Appendix H.1). VICS-G achieves 47.2% versus 46.7% task success, with 17.7% lower mean cost (2.71 to 2.23) than policy sampling. The paired cost difference is −0.48 [−1.00, −0.09] (Table R4 in our shared response).

These results demonstrate reduced safety cost with near-policy task success on previously unseen environments, using the same fixed configuration without retuning or evaluation–selection overlap.

W4: “The paper describes RCD as adapted from SafeDec but does not clearly explain what was changed or why [...] including additional baselines, such as hard-constrained decoding in SafeDec [...] would further strengthen the evaluation.”

We respectfully disagree that the adaptation is unspecified. Section 3 (Eq. 24) and Appendix F (Eqs. 53–54) already document its rationale, scoring function, and implementation. Unlike SafeDec’s original robustness evaluation of manually specified STL constraints over predicted successors, our RCD adaptation uses the available Safety-CHORES monitor-based local-risk signals under the same frozen-policy interface. We consider this the most appropriate interface-matched RCD comparator for our setting, as it enables a meaningful comparison using the same available safety information, without introducing additional STL specifications or predictive dynamics models not required by VICS-G.

We additionally implement the requested SafeDec-style kinematic hard-constrained decoder (HCD), filtering actions predicted to contact protected objects (Table R2 in our shared response). HCD achieves 25.6% task success, 11.1% joint safe success, and mean cost 5.49, compared with 47.2%, 21.7%, and 2.23 for nominal VICS-G. We explicitly distinguish this contact-based implementation from SafeDec's complete STL formulation.

On the overall rating

We appreciate the opportunity to clarify the decoding procedure, the RCD adaptation, and the role of approximation in our theory. The revised evaluation adds further results from the lookahead, ablation, and baseline experiments.

Given the reviewer's assessment that the soundness concerns are minor or fixable, and the recognized strengths of the paper, we respectfully hope this additional evidence will support a more favorable assessment of our work.

AuthorsOpenReview ↗

Follow-up on Our Responses and Manuscript Revisions

Dear Reviewer fXqs,

We have provided detailed responses and additional evidence addressing your concerns about the decoding procedure, theory–implementation connection, generalization, and baselines. The full results are in our shared experimental comments, with manuscript changes highlighted in purple.

We believe these additions substantially strengthen the paper. We would particularly value your feedback on whether your main concerns are now addressed. If any remain, we would welcome the opportunity to discuss them further.

Reviewer_fXqsOpenReview ↗

Official Comment by Reviewer fXqs

Thank you for the detailed rebuttal. Although some of my concerns have been alleviated, the gap between the theoretical objective and the practical VICS-G implementation remains insufficiently addressed. Besides, I encourage the authors to include an algorithm block in the main paper to improve clarity. I would maintain my score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Additional Held-Out Experiments: Connecting Feasible-Future Theory to Practical Decoding (Part 1/2)

Additional Held-Out Experiments: Connecting Feasible-Future Theory to Practical Decoding (Part 1/2)

We thank all reviewers for their constructive feedback. A common question is: How does the theoretical feasible-future objective translate into a practical decoder, and what evidence supports this connection?

This is the central challenge our work addresses. Exact feasible-future mass is generally intractable. VICS-G uses a lightweight, rollout-free correction zGz_G, while VICS-R provides complementary explicit lookahead. Our theory characterizes approximation errors and conditions for reliable ranking without assuming the heuristic is exact.

We additionally evaluate the frozen SafeVLA Safe checkpoint on 180 entirely unseen Fetch Safe houses, separate from the original 172 episodes, examining (1) continuation reasoning, (2) hard-constrained decoding, and (3) restricted simulator-cost feedback. Results are discussed below.

We have updated the manuscript accordingly, with revisions highlighted in purple.

1. Continuation Correction, Explicit Lookahead, and Paired Evaluation

Explicit continuation reasoning (VICS-R). Introduced in Appendix H, VICS-R augments VICS-G with Monte Carlo lookahead (MC4): at one eligible early decision, it compares the VICS-G action with up to two admissible alternatives using four simulated continuations per candidate. It executes the candidate with the highest estimated safe-completion probability, then resumes VICS-G. The underlying policy and downstream decoder remain unchanged.

We compare both decoders at ηv∈{0,0.10}\eta_v \in \{0,0.10\} with policy sampling, RCD, and retry-only VICS-G, without retuning.

Table R1. Held-out Fetch Safe: continuation correction and explicit lookahead. Success metrics are percentages with 95% Wilson score intervals (n=180n=180); other outcomes are mean ± SD.

MethodTask success ↑Joint safe success ↑Cost ↓Violations ↓Steps
Policy sampling46.7% [39.5, 53.9]21.7% [16.3, 28.2]2.71 ± 5.322.31 ± 4.49132.37 ± 65.80
RCD32.2% [25.8, 39.4]19.4% [14.3, 25.8]1.48 ± 2.271.19 ± 1.76123.33 ± 66.69
VICS-G, ηv=0\eta_v=046.7% [39.5, 53.9]21.7% [16.3, 28.2]2.44 ± 5.172.15 ± 4.74133.53 ± 65.96
VICS-G, ηv=0.10\eta_v=0.1047.2% [40.1, 54.5]21.7% [16.3, 28.2]2.23 ± 5.182.03 ± 4.23133.61 ± 66.06
VICS-R, ηv=0\eta_v=049.4% [42.2, 56.7]24.4% [18.7, 31.2]2.07 ± 5.171.85 ± 5.08131.14 ± 65.05
VICS-R, ηv=0.10\eta_v=0.1051.7% [44.4, 58.9]24.4% [18.7, 31.2]1.85 ± 3.971.61 ± 3.45130.88 ± 65.10
Retry-only VICS-G47.2% [40.1, 54.5]21.7% [16.3, 28.2]2.23 ± 5.182.03 ± 4.23133.61 ± 66.06

Direct nominal-versus-zero comparison. On the same 180 unseen Fetch Safe houses, using the top-12 actions plus the policy sample, VICS-G at ηv=0.10\eta_v=0.10 versus ηv=0\eta_v=0 yielded +0.56 pp in task success, 0.00 pp in joint safe success, −0.21 in mean safety cost, and −0.12 in mean violations (Table R1). These are point estimates for the incremental effect of enabling zGz_G.

We also report paired differences on matched episodes rather than relying on marginal intervals alone.

Table R4. Paired differences relative to policy sampling. Policy reference: 46.7% task success, 21.7% joint safe success, cost 2.71, violations 2.31. Success differences are percentage points. Brackets denote 95% paired bootstrap confidence intervals over matched episodes.

MethodΔ Task success (pp) ↑Δ Joint safe (pp) ↑Δ Cost ↓Δ Violations ↓
VICS-G, ηv=0\eta_v=00.00 [−3.45, 2.30]0.00 [0, 0]−0.27 [−0.57, −0.05]−0.16 [−0.34, −0.03]
VICS-G, ηv=0.10\eta_v=0.10+0.56 [−4.10, 3.95]0.00 [0, 0]−0.48 [−1.00, −0.09]−0.28 [−0.60, −0.05]
VICS-R, ηv=0\eta_v=0+2.78 [−1.90, 7.30]+2.78 [0, 7.80]−0.64 [−1.31, −0.13]−0.46 [−0.89, −0.11]
VICS-R, ηv=0.10\eta_v=0.10+5.00 [0.40, 9.60]+2.78 [0, 8.40]−0.86 [−1.74, −0.17]−0.70 [−1.40, −0.17]
Retry-only VICS-G+0.56 [−4.10, 3.95]0.00 [0, 0]−0.48 [−1.00, −0.09]−0.28 [−0.60, −0.05]
RCD−14.44 [−23.70, −5.31]−2.22 [−13.80, 9.20]−1.23 [−2.57, −0.19]−1.12 [−2.35, −0.21]

VICS-R versus policy sampling. At nominal weight, VICS-R achieves 51.7% versus 46.7% task success (+5.00 pp [0.40, 9.60]), with 31.7% lower mean cost and 30.3% fewer violations. All three reported paired intervals exclude zero. Joint safe success rises from 21.7% to 24.4%, with a paired interval that includes zero.

VICS-G versus policy sampling. The reported cost and violation intervals lie below zero at both weights. Nominal weight yields larger observed reductions in cost (−0.48 vs. −0.27) and violations (−0.28 vs. −0.16), with 47.2% vs. 46.7% task success relative to policy.

Retry-only matches nominal VICS-G at reported precision. The independent benefit of non-retry preferences remains unestablished. VICS-R tests explicit lookahead separately.

AuthorsOpenReview ↗

Additional Held-Out Experiments: Connecting Feasible-Future Theory to Practical Decoding (Part 2/2)

2. Comparison with Hard-Constrained Decoding

We evaluate a SafeDec-style kinematic hard-constrained decoder (HCD), using simulator geometry to filter actions predicted to contact protected objects.

Table R2. Held-out Fetch Safe: kinematic hard-constrained decoding. Success percentages include 95% Wilson intervals; other outcomes are mean ± SD.

MethodTask success ↑Joint safe success ↑Cost ↓Violations ↓Steps
Policy sampling46.7% [39.5, 53.9]21.7% [16.3, 28.2]2.71 ± 5.322.31 ± 4.49132.37 ± 65.80
RCD32.2% [25.8, 39.4]19.4% [14.3, 25.8]1.48 ± 2.271.19 ± 1.76123.33 ± 66.69
VICS-G, ηv=0\eta_v=046.7% [39.5, 53.9]21.7% [16.3, 28.2]2.44 ± 5.172.15 ± 4.74133.53 ± 65.96
VICS-G, ηv=0.10\eta_v=0.1047.2% [40.1, 54.5]21.7% [16.3, 28.2]2.23 ± 5.182.03 ± 4.23133.61 ± 66.06
Kinematic HCD25.6% [19.7, 32.4]11.1% [7.3, 16.5]5.49 ± 13.314.07 ± 9.57164.14 ± 51.99

Nominal VICS-G achieves higher task success (47.2% vs. 25.6%) and joint safe success (21.7% vs. 11.1%) than HCD, with lower mean cost (2.23 vs. 5.49).

This illustrates that local action constraints alone need not preserve safe task completion. These findings concern the evaluated kinematic contact filter, not all hard-constrained decoders.

3. Practicality Without Simulator-Provided Safety Costs

We assess dependence on privileged feedback by withholding simulator-provided numerical safety costs from the decoder while retaining target monitoring and independent evaluation.

Table R3. Held-out Fetch Safe: simulator-cost masking. Each condition has 180 episodes. Success percentages include 95% Wilson intervals; other outcomes are mean ± SD.

MethodMonitorTask success ↑Joint safe success ↑Cost ↓Violations ↓Steps
Policy samplingFull46.7% [39.5, 53.9]21.7% [16.3, 28.2]2.71 ± 5.322.31 ± 4.49132.37 ± 65.80
RCDFull32.2% [25.8, 39.4]19.4% [14.3, 25.8]1.48 ± 2.271.19 ± 1.76123.33 ± 66.69
RCDMasked44.4% [37.4, 51.7]20.6% [15.3, 27.0]3.53 ± 10.252.64 ± 6.54134.07 ± 66.75
VICS-G, ηv=0\eta_v=0Full46.7% [39.5, 53.9]21.7% [16.3, 28.2]2.44 ± 5.172.15 ± 4.74133.53 ± 65.96
VICS-G, ηv=0\eta_v=0Masked46.1% [39.0, 53.4]21.7% [16.3, 28.2]2.63 ± 5.542.29 ± 4.95133.83 ± 66.14
VICS-G, ηv=0.10\eta_v=0.10Full47.2% [40.1, 54.5]21.7% [16.3, 28.2]2.23 ± 5.182.03 ± 4.23133.61 ± 66.06
VICS-G, ηv=0.10\eta_v=0.10Masked47.2% [40.1, 54.5]21.7% [16.3, 28.2]2.34 ± 5.262.12 ± 4.24133.23 ± 65.94

VICS-G retains favorable safety–progress outcomes without numerical simulator-cost feedback. Nominal VICS-G preserves 47.2% task success and 21.7% joint safe success, with 13.7% lower mean cost than policy sampling (2.34 vs. 2.71).

Masking raises mean cost from 2.23 to 2.34, but VICS-G retains approximately 77% of its original cost reduction without sacrificing observed task success. RCD's mean cost rises from 1.48 to 3.53.

4. Direct Paired Ablation of the Continuation Correction

We compare ηv=0.10\eta_v=0.10 against ηv=0\eta_v=0 on 180 matched unseen Fetch Safe houses, holding other settings fixed.

Table R5. Nominal-minus-zero differences [95% paired bootstrap CI].

OutcomeVICS-GVICS-R
Task success ↑+0.56 pp [0.00, +1.67]+2.22 pp [+0.56, +4.44]
Joint safe success ↑0.00 pp [0.00, 0.00]0.00 pp [0.00, 0.00]
Mean safety cost ↓−0.21 [−0.74, +0.21]−0.220 [−0.579, −0.004]
Mean violations ↓−0.12 [−0.35, +0.04]−0.240 [−0.625, −0.011]

For VICS-G, the paired intervals for zGz_G include zero. VICS-R achieves statistically supported incremental improvements in task success, cost, and violations, with all three intervals excluding zero.

Overall Implications

The experiments reinforce two complementary realizations of our feasible-future framework: VICS-G provides a practical, rollout-free history-based correction, while VICS-R offers theory-grounded, future-looking feasibility estimation through simulated continuations.

Theorem 1 derives the exact policy-relative feasible-future objective ZfeasZ_{\mathrm{feas}}, while Theorem 2 characterizes approximation and recovery conditions. VICS-G demonstrates an efficient implementation with statistically supported episode-level safety improvements, including under restricted cost feedback. VICS-R more directly operationalizes future feasibility through prospective safe-completion estimates. Its continuations follow VICS-G rather than the frozen policy, so we do not claim exact estimation of ZfeasZ_{\mathrm{feas}}.

Together, these contributions extend beyond the independent effectiveness of zGz_G, connecting the theoretical feasible-future objective to two complementary, training-free decoders.

We hope these results clarify the theory–implementation connection and address concerns about the paper's independent contribution.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.