acceptodds
Under review as a conference paper at ICLR 2027

Learning to Verify: Allocating Revision Effort in Embodied Agents

Abstract

Improving an embodied agent requires recognizing when stable execution fails to satisfy an instruction. A robot may remain upright while moving in the wrong direction, traveling too far, or performing actions out of order. We propose Learn to Verify, then Revise (LVR), which turns feedback into executable task checks and uses them to guide subsequent revisions. LVR selects clarification questions by their expected effect on verification decisions, then reuses the learned checks without further feedback. Our analysis characterizes when a shared check is reliable enough for additional candidate search not to reduce expected task success. On recorded revision chains for 47 simulated robot tasks, LVR achieves 87.2% success, compared with 76.6% when every task is revised, while using 58% fewer revisions. With separately generated revision chains on 39 tasks, updated checks achieve 89.7% success versus 74.4% before feedback, selecting 50 versus 40 revisions under offline stopping. Classification and transfer tests show no corresponding established gain in held-out accuracy. These results motivate evaluating learned verification through the revision decisions it enables, alongside its predictive accuracy.

Primary area:
applications to robotics, autonomy, planning
Keywords:
learned verification, embodied agents, revision allocation, active learning, language-model feedback, robot learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jgpbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Learn to Verify, then Revise (LVR), a method to learn verification checks from the feedback which is then used to guide the revisions without training the robot policy control. The paper shows that execution stability is not enough to determine task success as nearly half of the tasks with stable robot execution still fail to satisfy the given instructions.

LVR learns verification checks based on trajectory features, thresholds, and applicability conditions using feedback from a language model. The method selects clarification questions based on how much their answers are expected to improve verification decisions. Once the checks are learned, they are kept fixed and used to identify which motions need revision. When a check fails, the planner receives feedback about the failed requirement and generates a revised motion.

The method is evaluated on simulated Microduck robot tasks. LVR achieves 87.2% task success with 59 selected revisions, compared to 76.6% success with 141 revisions when all tasks are revised. The updated verifier also improves task success from 74.4% to 89.7% compared to the verifier before feedback.

Critical strengths
  1. The problem is well motivated. The paper identifies stable robot execution can fail to satisfy the given instructions.

  2. LVR does not need any updates to the robot control policy, as verification checks are learned from language-model feedback and uses them to decide which motions need revision.

  3. Experiments show that LVR achieves 41/47 successful tasks with 59 selected revisions, compared to 36/47 with 141 revisions for repair-all on recorded revision chains. The updated verifier also improves success from 29/39 to 35/39 tasks compared to the pre-feedback verifier.

Critical weaknesses
  1. In Table 2, direct labeling achieves lower classification loss than LVR, while LVR increases false rejections compared to the fixed verifier. Also in Experiment C, the learned checks also fail to improve classification on held-out tasks. It is unclear whether the proposed query selection method performs better than simpler baselines or that the learned checks generalize to new tasks.
  2. In Table 3, the updated verifier improves task success but also selects more revisions. As the pre-feedback and updates verifiers uses different number of revisions, it is not clear if the improvement comes from more revisions or learning better checks.
  3. The main revision results use recorded revision chains with stopping decisions applied offline. So, the reported reduction in revisions reflects the revisions selected by the policy rather than the actual computation used during the experiments.
  4. The evaluation is limited to one simulated robot, one model. It is unclear if the results would hold with other simulated robot, model or real robot settings.
Questions for authors
  1. Could you compare the pre-feedback and updated verifiers using the same number of revisions?
  2. Why do the learned checks fail to improve classification on held-out tasks in Experiment C?
  3. Could you compare the proposed query selection method with random queries and direct labeling in terms of final task success, using the same feedback and revision budget?
  4. Could you evaluate LVR with online stopping instead of applying stopping decisions to recorded revision chains? Since the reported reduction in revisions is based on offline stopping, how much computation does LVR actually save when feedback and verification costs are included?
  5. Could you test whether the improvements generalize to different robots or models?
1 comment
Reviewer_jgpbOpenReview ↗

Rebuttal Acknowledgement

I thank the authors for their clarification and new experiments. My concern about revision allocation is partly addressed. As random queries can also learn verification checks, I am not sure what impact that would have on the final task success. I think the benefit of their method is not established over the simpler ones. Therefore, I will maintain my original rating.

Reviewer_DFdHOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

LVR learns executable task checks from clarification feedback and uses them to guide motion revisions. Experiments evaluate verification, transfer, and revision allocation using simulated trajectories and recorded revision chains.

Critical strengths

Explicit checks distinguish task satisfaction from stable execution and provide interpretable repair diagnostics. On shared recorded chains, selective revision achieves 41/47 successes with 59 revisions, compared with 36/47 using 141 revisions for repair-all.

Critical weaknesses

Evidence for the proposed question-selection method is weak: it establishes no classification advantage, while learned checks change no held-out transfer decisions. The learning comparison uses one verifier pair, offline stopping, and unequal selected revision counts (50 versus 40). It therefore does not establish a learning advantage at matched expenditure.

Questions for authors

Clarification: Which query strategy, feedback budget, and aggregation rule produced the updated verifier in Table 3? Additional analysis: How much of Table 3’s success improvement remains when the selected revision budget is matched?

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

The work proposes to implement a feedback system for an embodied agent, in this case, the Microduck simulation. They do this by identifying primitives that verify various agent actions and then provide correction/revision feedback to improve the current action. The success is measured by whether the identified revision was necessary or useful in the first place, and how much post-revision the performance improves.

Critical strengths
  • It's a different perspective on the synergy between the planner and verifier to evaluate how valuable the verifier is. It's also interesting to see it being applied in an embodied setting
Critical weaknesses
  • The writing is not very good
    • Lots of jargon that doesn’t help understanding at all.
    • Section 3 opens by listing the experiments (execution–task disagreement (A), feedback-driven classification (B), transfer (C), revision allocation (D)) without stating what each one measures. The relevant metrics are scattered across Sections 2.3 and 3.2–3.3, and Appendix F, and some table entries are never defined, e.g., “selections /16” in Table 2.
    • Is the planner an LLM? If yes, which one, and why is it not stated anywhere in the manuscript?
  • The core idea is to learn explicit verification checks from feedback and evaluate them by how well they guide revisions. This seems like a setting that applies equally to language-modeling tasks; e.g., code that executes without errors but returns the wrong answer is a direct analog of the “silent failure” motivating this work. Is there a particular reason the embodied setting is necessary to study this, given that it requires hand-defining a feature vocabulary for the checks, and the learned checks do not transfer to held-out tasks (Experiment C)?
Questions for authors
  • Your thoughts on the points raised in the weakness section would be useful for the discussion.
  • Are there any obvious LLM or VLM as verifiers baselines (e.g. AHA by Duan et al., 2025 ) from prior work that can apply to this task? If yes, any reason why they were not evaluated or compared against?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.