acceptodds
Under review as a conference paper at ICLR 2027

Verification Before Commitment: Active Environmental Verification for Embodied Understanding

Abstract

Embodied agents must often infer environmental states from partial observations. However, limited evidence may already yield a plausible interpretation without being sufficient to establish its correctness, leading agents to commit prematurely. We address this gap between interpretation plausibility and evidence sufficiency through a verification-based policy, which treats a proposed answer as a provisional hypothesis and shifts subsequent interaction from answer seeking to active hypothesis testing. Using embodied question answering (EQA) as a testbed, we introduce Sure-EQA, a training-free framework that re-examines existing visual support from alternative viewpoints and probes unseen regions for answer-changing evidence. Verification findings then guide whether the candidate should be accepted, further tested, or revised. Experiments on A-EQA and EXPRESS-Bench show consistent improvements in answer correctness and path-weighted performance. Under matched additional observation, Sure-EQA also outperforms continued exploration in reducing missing evidence and correcting mistaken interpretations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied Question Answering, Embodied Understanding, Active Environmental Verification, Vision-Language Models, Active Perception

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RamKOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies premature commitment in VLM-based embodied agents: an agent forms an answer that looks plausible from partial observations and stops, even though the evidence it has gathered is not enough to establish that answer. The authors propose Sure-EQA, a training-free framework that treats the agent's answer as a hypothesis to be tested, not a final decision. With the candidate held fixed, it runs two kinds of active checks. Evidence verification re-observes the supporting regions from new viewpoints. Coverage verification visits unexplored frontiers to look for information that would change the answer. A VLM verifier labels each new observation as SUPPORTED, CHANGED, or UNCLEAR. These labels are combined into pose-weighted answer-change risk scores, which decide the next viewpoint and whether to accept the candidate. On A-EQA and EXPRESS-Bench with Qwen3-VL 8B, Sure-EQA improves over previous methods.

Critical strengths
  1. The problem is well motivated by a clear diagnostic study. Fig. 1 shows that when agents propose to stop, the annotated key evidence is still missing in 38.6–51.3% of cases.
  2. With a training-free design and an open-source 8B backbone, Sure-EQA achieves state-of-the-art results on A-EQA and EXPRESS-Bench benchmarks.
Critical weaknesses
  1. Every number comes from a single run on 184 or 409 questions, scored by a GPT-4 judge. There are no seeds, confidence intervals, or paired significance tests. Many ablation and analysis differences are 1–2 points (e.g., Table 3), which is within plausible run-to-run and judge variance for an LLM-driven agent on 184 questions.
  2. The answerer and both verifiers are the same 8B VLM, so their errors are likely correlated. The paper reports no verifier accuracy, for example how often SUPPORTED/CHANGED labels agree with the ground-truth effect on the answer. Fig. 12 shows 41.7% of verified-accepted A-EQA answers are still not fully correct, so the verification signal seems weak.
  3. Table 6 says one verification step is cheaper than one exploration step. The relevant question is how much total wall-clock time and compute Sure-EQA adds per episode compared with the baseline, including the extra exploration after rejected rounds. That number is not reported.
Questions for authors

"w/o Risk Score" (Table 3) drops to 53.9. Why does removing the risk term from selection alone erase almost the whole gain?

4 comments
AuthorsOpenReview ↗

Response to Reviewer RamK (Part 1/2)

Thank you for the detailed feedback. Below, we clarify the five-seed protocol, assess verifier reliability, and report episode-level runtime and computation. Finally, we explain how risk-guided viewpoint selection contributes to the observed performance gains.

W1. Repeated runs and statistical reliability.

Our experiments use five random seeds, each corresponding to a separate trajectory rollout. Reported scores are averaged across the five runs. Below, we provide SD, paired 95% confidence intervals for performance differences, and two-sided paired t-test p-values. The repeated-run protocol is detailed in revised Appendix A.4.2 (“Repeated runs”).

Main comparisons.

Benchmark / metricOurs mean ± SDComparator mean ± SDDifference95% paired CIp-value
A-EQA / LLM-Match58.7±0.6958.7_{\pm 0.69}53.3±0.8853.3_{\pm 0.88}+5.40[4.51, 6.29]0.000073
A-EQA / LLM-SPL52.8±0.6452.8_{\pm 0.64}48.5±0.8548.5_{\pm 0.85}+4.30[3.30, 5.30]0.000279
EXPRESS-Bench / C∗C^{*}78.19±0.7178.19_{\pm 0.71}70.54±0.7070.54_{\pm 0.70}+7.65[6.48, 8.82]0.000054
EXPRESS-Bench / EpathE_{\mathrm{path}}50.69±0.6650.69_{\pm 0.66}47.66±0.4947.66_{\pm 0.49}+3.03[2.74, 3.32]0.0000081

Ablation comparisons.

Benchmark / metricComparatorOurs mean ± SDComparator mean ± SDDifference95% paired CIp-value
A-EQA / LLM-Matchw/o View Novelty58.7±0.6958.7_{\pm 0.69}55.0±0.6155.0_{\pm 0.61}+3.70[3.11, 4.29]0.000063
A-EQA / LLM-Matchw/o Navigation Cost58.7±0.6958.7_{\pm 0.69}58.2±0.5658.2_{\pm 0.56}+0.50[0.21, 0.79]0.008858
A-EQA / LLM-Matchw/o Risk Score58.7±0.6958.7_{\pm 0.69}53.9±0.9753.9_{\pm 0.97}+4.80[3.52, 6.08]0.000483
A-EQA / LLM-Matchw/o Structural Record58.7±0.6958.7_{\pm 0.69}54.3±0.6054.3_{\pm 0.60}+4.40[3.70, 5.10]0.000064
A-EQA / LLM-Matchw/o Visual Feedback58.7±0.6958.7_{\pm 0.69}55.7±0.7655.7_{\pm 0.76}+3.00[2.10, 3.90]0.000757
EXPRESS-Bench / C∗C^{*}w/o View Novelty78.19±0.7178.19_{\pm 0.71}75.90±0.5375.90_{\pm 0.53}+2.29[1.80, 2.78]0.000210
EXPRESS-Bench / C∗C^{*}w/o Navigation Cost78.19±0.7178.19_{\pm 0.71}76.30±0.5776.30_{\pm 0.57}+1.89[1.48, 2.30]0.000220
EXPRESS-Bench / C∗C^{*}w/o Risk Score78.19±0.7178.19_{\pm 0.71}73.50±0.5273.50_{\pm 0.52}+4.69[4.10, 5.28]0.000025
EXPRESS-Bench / C∗C^{*}w/o Structural Record78.19±0.7178.19_{\pm 0.71}74.00±0.6374.00_{\pm 0.63}+4.19[3.79, 4.59]0.0000085
EXPRESS-Bench / C∗C^{*}w/o Visual Feedback78.19±0.7178.19_{\pm 0.71}75.40±0.5475.40_{\pm 0.54}+2.79[1.77, 3.81]0.001627

Notably, our method achieves statistically significant improvements over Pred-EQA across all four metrics and over the ablated variant on both benchmarks. These analyses demonstrate that, by coupling risk-guided viewpoint selection with retained verification feedback, Sure-EQA can effectively identify mistaken interpretations through targeted environmental checks and correct them before commitment.

AuthorsOpenReview ↗

Response to Reviewer RamK (Part 2/2)

W2.1. Verifier reliability

Regarding the concern that sharing a backbone may lead to correlated errors, Sure-EQA uses distinct verifier roles and newly acquired environmental evidence to challenge the answerer’s initial interpretation. The verifiers re-examine its supporting evidence and seek missing or contradictory evidence through targeted environmental checks.

We further assess individual verifier reliability through an offline analysis on A-EQA. An LLM judge (GPT-6 Astra) evaluates all 1,855 judgments, complemented by a human expert audit of 100 sampled judgments.

EvaluationSUPPORTED precision (%)CHANGED precision (%)UNCLEAR precision (%)Overall Accuracy (%)Overall Macro-F1 (%)
LLM judge88.2584.3184.5486.5284.06
Human expert audit85.4584.2184.6285.0083.02

These statistics indicate that individual verifier judgments are reasonably reliable. Since individual checks remain fallible, Sure-EQA accumulates verification findings through risk-score updates to improve robustness to occasional misjudgments. Table 5 in the appendix supports this design, showing higher answer quality and lower answer degradation than individual verification.

W2.2. Corrective effectiveness.

Verified acceptance indicates that the candidate satisfies our evidence-based verification criteria, rather than guaranteeing ground-truth correctness. Figure 12 shows that our framework corrects 18.59% of initially incorrect answers while degrading only 5.13% of initially correct answers among verified-accepted A-EQA cases, increasing the fully correct answer rate from 44.87% to 58.33%. Moreover, among the 41.67% of accepted answers that remain less than fully correct, 18.46% still achieve higher evaluation scores than their initial counterparts. These results demonstrate the effectiveness of our verification framework in correcting mistaken interpretations and improving answer quality.

W3. Episode-level runtime and computational overhead.

As shown in the table below, we report mean episode-level costs on EXPRESS-Bench, including all verification rounds and additional exploration following rejected rounds.

MethodLatency (s/episode) ↓Computation (PFLOPs/episode) ↓
Baseline5676.30
Sure-EQA7757.76
Relative increase+36.6%+23.1%

Although Sure-EQA incurs higher total episode-level costs, its incremental overhead only amounts to 36.6% of baseline runtime and 23.1% of baseline model computation. This includes verification and additional exploration following unsuccessful rounds. At this controlled additional cost, Sure-EQA can effectively enable the agent to uncover and correct mistaken interpretations.

Q. Performance drop after removing risk-guided selection.

This sharp drop highlights the effectiveness of our verification-guided viewpoint selection in directing observations toward evidence that may challenge the agent's current interpretation. In this ablation, only the risk term in Eq. (7) is removed, while the verification mechanisms and risk-based acceptance criterion in Eq. (6) remain unchanged. Without risk guidance, viewpoint selection is driven primarily by geometric novelty and travel cost, potentially directing limited observations away from unresolved concerns about the candidate answer. This weakens the evidence available for effective feedback and answer revision. The results thus support our core design of using accumulated verification findings to actively guide subsequent environmental interaction, rather than merely assessing existing evidence.

Thank you for the detailed response. The additional results address most of my concerns. I suggest moving the episode-level runtime and computation comparison into the main paper rather than leaving it to the rebuttal. I will keep my score.

AuthorsOpenReview ↗

Response to Reviewer RamK

Thank you again for your constructive and thoughtful feedback. We have described episode-level costs in Appendix Tab. 6 according to your detailed suggestions. Together with per-step cost analysis, Sure-EQA uses efficient step to correct self-consistent erroneous interpretations for embodied agent.

If any concerns remain, we would be grateful if you could share them and we may address them more fully. Should further clarification resolve these concerns, we hope you might reconsider your rating in the revised manuscript and additional experiments.

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

In this paper, the authors propose a verification-based policy that treats interactions as hypothesis testing. They introduce Sure-EQA (embodied QA), an active inference-time verification framework that probes unexplored or alternative regions to uncover missed evidence that can guide the model toward further testing, revision, or acceptance based on the agent’s proposed answer. It uses two components: an answerer (to determine if the evidence is sufficient) and a planner (to define a set of possible targets). They evaluate the framework across open-source and proprietary models based on answer correctness, grounded quality, and cost (specifically, travel cost and full trajectory).

Critical strengths
  • The analysis of the problem's systematicity on an external benchmark (EXPRESS-Bench), along with thorough experiments and ablations.
  • The use of geometric (pose) similarity to determine a finding's relevance across targets while discounting travel cost.
Critical weaknesses
  • Verification baselines: The existing results should also be compared to other verification methods, rather than solely to unverified baselines. This comparison is distinct from candidate generators (as they all use the same verification framework in Section 4.3) or the continual baselines in Section 4.4.

  • Inference-time latency: What is the inference-time latency of using this framework? The authors do not report it relative to unverified baselines or other verification methods.

Questions for authors

Missing and interesting statistics:

  • How many new pieces of information or evidence (active verification turns) are required in cases where the initial partial-evidence-based interpretation was incorrect? (This would provide additional context to the results in Figure 6.)
  • How many times does the agent switch back to the exploration-based policy (based on verification feedback)?
  • Are there statistics regarding the reasoning or rationales, such as clusters, groups, or types (as mentioned in line 294)?
  • How does performance vary across different scene complexities in Habitat?

Clarification on verification:

  • Is the verification applied at every depth level? For example, if there is an initially proposed answer that leads to FF possible new targets, is each new target actively simulated and assessed again by the planner for the next turn?

Model backbones and hyperparameters:

  • Regarding other VLM backbones (almost all of the open-source models used are Qwen-based, with one being Janus), what about evaluating LLaVA versions?
  • How are the thresholds TeT_e and TcT_c chosen?

Ablations:

  • An ablation study on the risk score is missing.

Writing and Style:

  • There is a missing line break or gap between the Figure 2 caption and the following paragraph (lines 68-71).
  • In Figure 1, which VLMs are participating? Is the metric a mean across different backbones of varying scales? This is unclear when it is first introduced in the introduction section.
4 comments
AuthorsOpenReview ↗

Response to Reviewer qSEi (Part 1/2)

We sincerely appreciate your detailed feedback. Below, we address the concerns about verification baselines and inference costs. In response to your questions, we also provide additional verification statistics and evaluations.

W1. Comparison with existing verification methods.

Directly applicable verification baselines for embodied agent remain limited. Among existing baselines, Explore-EQA offers a closely related alternative through confidence-calibrated exploration. Our main experiments in Table 1 already demonstrate substantial gains over this baseline. Moreover, Figure 4(a) evaluates self-check and continued exploration as controls we construct to represent two natural verification modes. Our framework also outperforms both, reinforcing the advantages of our risk-guided verification policy.

We further adapt Self-Refine and Pelican, two verification frameworks from other domains, to EQA. Both use the same exploration-based policy as Sure-EQA. Results on A-EQA are reported below:

MethodLLM-Match ↑LLM-SPL ↑
Self-Refine54.049.0
Pelican54.649.5
Sure-EQA58.752.8

Self-Refine performs iterative self-feedback, while Pelican decomposes and verifies claims against available visual evidence. In contrast, our framework uses verification findings to actively guide further environmental observations, seeking evidence that may challenge the current answer before accepting or revising it. These results support the benefit of integrating active evidence acquisition into the verification process, beyond reassessing existing observations alone.

W2. Inference-time latency relative to baselines.

Appendix Table 6 already reports per-step latency and computational cost for exploration and verification. We further report mean episode-level costs on EXPRESS-Bench, including all verification rounds and additional exploration following rejected rounds.

MethodLatency (s/episode) ↓Computation (PFLOPs/episode) ↓
Baseline (Pred-EQA)5676.30
Explore-EQA7047.08
Self-Refine8958.68
Pelican1,0249.71
Sure-EQA7757.76

As shown above, Sure-EQA maintains runtime and computational costs comparable to Explore-EQA. Relative to Self-Refine and Pelican sharing the same exploration-based policy, Sure-EQA enables the higher task performance with the lower cost.

Q. Missing and interesting statistics.

R1. To quantify the additional evidence for initially incorrect candidates, We conducted an offline analysis summarized as below:

Benchmarkn/N (%)Mean turns (all)Mean recordsE/C (%)S/CH/U (%)
A-EQA75/184 (40.76)2.17 (3.36)6.5143.77/56.2329.97/42.66/27.37
EXPRESS-Bench120/409 (29.34)2.85 (3.57)8.5541.74/58.2635.45/44.59/19.96

n/N: initially incorrect cases with first-round feedback / all evaluated questions. E/C: evidence/coverage distribution in records. S/CH/U: SUPPORTED/CHANGED/UNCLEAR distribution in records. Each verification turn yields three records. Among them, frequent CHANGED judgments indicate effective detection of incorrect candidates, consistent with the correction gains in Figure 6.

R2. By default, Sure-EQA requires at least one verification round before acceptance (Figure 7a). The table below reports mean switches back to the exploration-based policy per episode, with question counts in parentheses.

BenchmarkAll questionsInitially incorrect candidates
A-EQA0.83 (184)1.37 (98)
EXPRESS-Bench0.44 (409)1.05 (145)

Sure-EQA more frequently resumes exploration for initially incorrect candidates, enabling their underlying interpretations to be challenged and revised before commitment.

AuthorsOpenReview ↗

Response to Reviewer qSEi (Part 2/2)

Q. Missing and interesting statistics.

R3. We propose an offline GPT-6 Astra analysis that assigns each rationale to one of five categories. The illustrative distributions are summarized below.

Rationale categoryA-EQA (%)EXPRESS-Bench (%)
Object/attribute evidence33.0131.70
Spatial/contextual evidence27.4830.43
Coverage findings22.6823.23
Observation limitations13.7412.34
Other3.092.30

Crucially, our method accommodates diverse verification needs, examining the different ways an embodied interpretation may be mistaken or insufficiently supported. For verification-type and judgment distributions, please refer to the R1 table under “Q. Missing and interesting statistics.”

R4. Following HM3D’s official definition of scene complexity, we stratify HM3D scenes into low-, medium-, and high-complexity groups based on navigational complexity. The table below reports LLM-Match on A-EQA and C∗C^{*} on EXPRESS-Bench.

BenchmarkComplexityQuestionsBaselineSure-EQAGain
A-EQALow6361.064.5+3.5
A-EQAMedium6453.058.3+5.3
A-EQAHigh5745.052.7+7.7
EXPRESS-BenchLow14382.4086.80+4.40
EXPRESS-BenchMedium11371.6079.10+7.50
EXPRESS-BenchHigh15358.6769.47+10.80

Notably, Sure-EQA consistently improves performance across all groups, with larger gains in more complex environments.

Q. Clarification on verification.

Verification is performed selectively rather than exhaustively across all candidate targets or planning depths. The framework maintains a set of reachable verification targets and estimates their risks using accumulated verification findings, without simulating or invoking the verifier on every target. At each step, Eq. (7) selects one target based on risk, viewpoint novelty, and travel cost. Only the selected target is visited and verified, after which the risk scores are updated and the acceptance condition in Eq. (6) is re-evaluated. This process continues until the candidate is accepted, the verification round ends with feedback, or the interaction budget is exhausted.

Q. Model backbones and hyperparameters.

R1. Sure-EQA yields consistent improvements across different VLM backbones. To further assess this generalizability, we replace the backbone of both Pred-EQA and Sure-EQA with LLaVA-NeXT–Mistral-7B. The results on the A-EQA are shown as below:

BenchmarkMetric ↑Pred-EQASure-EQA
A-EQALLM-Match48.654.2
A-EQALLM-SPL44.148.2
EXPRESS-BenchCC46.1251.36
EXPRESS-BenchC∗C^{*}64.8372.18
EXPRESS-BenchEpathE_{\mathrm{path}}40.9444.27

These gains are consistent with Table 1, demonstrating that Sure-EQA’s effectiveness is not confined to a specific VLM backbone.

R2. The risk thresholds are tuned through a grid search, with (TeT_e=TcT_c) to control search costs. Figure 10 presents the comparisons. Development-set performance on LLM-SPL and C metric guides the final choice of (TeT_e=TcT_c=0.53).

Q. Ablations.

Table 3 already reports an ablation of the answer-change risk score. Removing the score from viewpoint selection while preserving the acceptance rule leads to the largest performance drop, highlighting its importance for correcting mistaken interpretations.

Q. Writing and Style.

R1. Thank you for pointing this out. We have revised the manuscript accordingly.

R2. Figure 1 reports results obtained with Qwen VLMs and we have clarified this. Appendix A.1 already details the protocol and analysis of this pilot experiment.

Thank you to the authors for the detailed rebuttal and the effort put into the additional evaluations. I specifically appreciate the new comparisons with the verification baselines (Self-Refine and Pelican) and the added statistics regarding inference latency and verification turns.

While your response successfully addresses and clarifies my initial concerns regarding baselines and latency, my overall assessment of the paper's core contribution remains a 3 (Weak acceptance). The proposed framework, while addressing an important practical gap in embodied question answering, relies heavily on standard agentic design patterns. The addition of a dual-risk verification loop is certainly a sensible and well-executed step, but it feels more like an incremental policy adjustment.

Thank you again for your hard work and for engaging constructively with the review process.

AuthorsOpenReview ↗

Response to Reviewer qSEi

Thank you for your thoughtful follow-up and for acknowledging that our additional evaluations address your initial concerns. We also appreciate your recognition of the practical significance of our work.

We would like to clarify that our contribution goes beyond adding a dual-risk verification loop to an existing agent. Our key insight is that a plausible but mistaken interpretation can itself reinforce erroneous decisions, as the agent may continue reasoning from this interpretation or stop exploring before discovering contradictory evidence. Our verification-based policy breaks this cycle by turning the proposed interpretation from a premise for subsequent decisions into a hypothesis to be actively challenged. Importantly, evidence and coverage verification are not merely additional checking modules. Their accumulated findings jointly determine what evidence to acquire next and whether the proposed answer should be accepted, establishing a different interaction objective and decision mechanism. Controlled experiments support this distinction, showing improvements from explicit verification under identical observations (58.7 vs. 56.5) and risk-guided viewpoint selection (58.7 vs. 53.9).

We hope this clarification and the additional experimental evidence strengthen your confidence in our methodological contribution. We would be grateful if you could reconsider whether a more favorable final assessment is warranted.

Reviewer_LXG5OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of premature commitment in embodied question answering, where agents generate seemingly plausible answers based on insufficient environmental evidence and terminate exploration too early. To tackle this issue, the authors propose Sure-EQA, a training-free active environmental verification framework that examines candidate answers through evidence verification and coverage verification, and enables answer revision through risk-guided viewpoint selection and verification feedback. Experimental results on A-EQA and EXPRESS-Bench demonstrate that the proposed method consistently improves answer correctness and path efficiency while effectively reducing missing evidence and mistaken interpretations.

Critical strengths
  1. Well-motivated and empirically supported problem formulation. The paper identifies an important failure mode in embodied question answering: agents may prematurely commit to plausible answers without sufficient environmental evidence. The distinction between answer plausibility and evidence sufficiency is clearly illustrated by the example in Figure 2. More importantly, the diagnostic study in Figure 1 demonstrates that missing evidence remains a common issue even with stronger VLM backbones.
  2. The experimental analysis is comprehensive. The paper provides extensive experiments to investigate the effectiveness of the proposed framework, including matched-observation comparisons against continued exploration, component ablations, evaluations with four candidate generators, and analyses of missing evidence and answer correction. Additional experiments on verification budgets and computational overhead further strengthen the evaluation.
Critical weaknesses
  1. Insufficient comparison with existing verification and active exploration paradigms. Although the paper presents verification before commitment as a new interaction objective, several core components are closely related to existing approaches in active perception, uncertainty-driven exploration, hypothesis testing, and VLM-based self-correction. Specifically, the framework combines candidate answer generation, additional viewpoint acquisition, VLM-based verification, evidence aggregation, and answer revision. Therefore, the main novelty appears to lie in organizing these components around candidate answers and distinguishing two types of verification risk. While this represents a meaningful system-level contribution, it remains unclear whether the paper introduces a fundamentally new verification principle or primarily presents a carefully engineered verification strategy for EQA. A more explicit comparison with existing verification and active exploration approaches would help clarify the novelty and significance of the proposed method.
  2. Heuristic risk estimation and insufficient analysis of verification reliability. The answer-change risk score is constructed from heuristic combinations of VLM judgments and geometric cues, without a principled calibration or a thorough assessment of its reliability. Moreover, since the answer generation and verification processes rely on similar VLM capabilities, correlated errors may cause the verifier to reinforce incorrect answers rather than correct them.
Questions for authors
  1. Could the authors provide a controlled comparison with simpler uncertainty-based or information-gain-based viewpoint selection strategies under the same candidate answers, verification prompts, and observation budgets to isolate the contribution of the proposed risk-guided verification?

  2. Is the proposed answer-change risk score intended as a heuristic ranking criterion or a calibrated probability estimate, and how robust is the framework to incorrect VLM verification judgments?

5 comments
AuthorsOpenReview ↗

Response to Reviewer LXG5 (Part 1/2)

Thank you for the constructive feedback. The responses below address your concerns about our novelty and technical contributions. We also report additional controlled experiments demonstrating Sure-EQA’s advantages.

W1. Novelty Beyond Existing Exploration and Verification Approaches.

Our contribution lies in identifying a previously underexplored failure mode in embodied understanding and developing an active verification approach to address it. As shown in Fig. 1, agents often terminate exploration once they form plausible environmental interpretations, even when critical evidence remains missing. We characterize this failure as premature commitment, revealing a gap between interpretation plausibility and evidence sufficiency that conventional answer-seeking exploration does not explicitly address. This motivates a different perspective on embodied understanding: a plausible answer should be treated as a hypothesis requiring further evidence, rather than as sufficient grounds for terminating exploration (Fig. 2).

Our methodological contribution goes beyond combining exploration and verification techniques. While conventional exploration primarily seeks information to answer questions, Sure-EQA directs subsequent observations toward testing an interpretation the agent is already prepared to accept. Specifically, we distinguish two sources of potential answer changes: misinterpretation of existing observations and omission of decisive evidence. We estimate their respective answer-change risks through evidence verification and coverage verification, and use these risks to jointly guide viewpoint selection and answer acceptance. This establishes a common verification objective for evidence acquisition and stopping decisions, rather than treating exploration and verification as separate procedures.

Our controlled experiments support this methodological distinction. Under matched observation budgets, Sure-EQA outperforms continued exploration explicitly instructed to seek counterevidence (58.7 vs. 53.9 LLM-Match) and reduces missing evidence from 39.6% to 19.1% (Fig. 4). It also outperforms uncertainty-based viewpoint selection (58.7 vs. 54.7), suggesting that targeting potential changes to the current interpretation is more effective than relying on general uncertainty in this setting. Given identical observations and trajectories, explicit verification outperforms candidate-aware factual feedback (58.7 vs. 56.5, Appendix Fig. 9). Furthermore, removing risk-guided viewpoint selection reduces LLM-Match from 58.7 to 53.9 (Table 3). These results indicate that the improvement stems not merely from additional exploration or answer checking, but from using verification to guide what evidence should be acquired before committing to an interpretation.

W2. Reliability of Risk Estimation and Verification

Regarding risk estimation, our answer-change risk is designed as a decision score rather than a calibrated probability estimate. It aggregates verification judgments across viewpoints while accounting for spatial relevance and redundant observations, providing a more informed basis for verification decisions than isolated judgments (Eqs. 4–7).

We further assess its reliability and effectiveness through the following evidence.

First, we evaluate the reliability of the underlying VLM verification judgments:

EvaluationSamplesAccuracy (%)Macro-F1 (%)
LLM judge (GPT-6 Astra)1,85586.5284.06
Human expert audit10085.0083.02

Second, we assess the effectiveness of risk-based decisions. Table 3 shows that removing risk-guided viewpoint selection reduces LLM-Match from 58.7 to 53.9. Appendix Fig. 10 further shows that our geometrically weighted risk control outperforms count-based control (58.7 vs. 56.7), while requiring fewer verification steps (3.36 vs. 3.93).

Regarding correlated VLM errors, although the answerer and verifiers share the same backbone, they perform different roles and rely on different environmental evidence. The answerer generates a candidate from available observations, while verification actively tests this candidate by re-examining its supporting evidence and seeking potentially missing or contradictory evidence. Thus, rather than simply repeating the original reasoning, our framework can identify and correct errors through targeted environmental observations. Moreover, verification findings are accumulated across viewpoints to reduce reliance on individual judgments. Empirically, Fig. 6 shows more corrections than degradations (16.38% vs. 4.65%), with consistent trends in Appendix Fig. 12.

AuthorsOpenReview ↗

Response to Reviewer LXG5 (Part 2/2)

Q1. Controlled Comparison with Uncertainty-Based Viewpoint Selection

To isolate the contribution of risk-guided viewpoint selection, we compare it with a history-based uncertainty strategy while keeping candidate answers, verification prompts, observation budgets, and other framework components unchanged. The baseline estimates uncertainty from the disagreement between SUPPORTED and CHANGED judgments at nearby verification viewpoints, assigning higher priority to locations with more balanced judgments.

Selection StrategyLLM-Match ↑LLM-SPL ↑
History-Based Uncertainty54.749.8
Risk-Guided Selection (Ours)58.752.8

Our approach improves LLM-Match and LLM-SPL by 4.0 and 3.0 points, respectively. While the uncertainty baseline prioritizes locations with conflicting verification judgments, our risk-guided strategy focuses on evidence indicating that the current candidate may need revision. This comparison supports the value of explicitly considering answer-change risk in viewpoint selection.

Q2. Risk-Score Interpretation and Robustness to Verification Errors

Regarding risk-score interpretation, our answer-change risk is a heuristic decision score, not a calibrated error probability. It guides viewpoint selection and candidate acceptance, with acceptance thresholds selected on development data and fixed before evaluation (Appendix A.4.2).

Regarding robustness, our framework is designed to tolerate occasional incorrect verification judgments rather than rely on any single judgment. The verifier can return UNCLEAR when evidence is inconclusive, while SUPPORTED and CHANGED judgments are accumulated across viewpoints with spatial relevance and redundancy considered (Eq. 5). This prevents a single potentially misleading CHANGED judgment from immediately triggering answer revision. Empirically, Appendix Table 5 shows that accumulated verification improves LLM-Match from 54.2 to 58.7 and reduces answer degradation from 11.96% to 5.43% compared with immediate feedback.

AuthorsOpenReview ↗

Response to Reviewer LXG5

Thank you again for your thoughtful and constructive feedback. We have taken your concerns very seriously and, following our detailed rebuttal, made further efforts to refine the manuscript. In particular, we have clarified our core methodological contributions in Section 3 and strengthened the supporting experimental analyses in Section 4. We sincerely hope these revisions, together with the additional evidence presented in our rebuttal, help address your concerns more convincingly.

We would greatly appreciate it if you could take another look at our revised manuscript. If you find that your concerns have been satisfactorily addressed, we would be truly grateful if you could reconsider your rating. Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank our reviewers for their constructive feedback and the time devoted to evaluating our work.

Our Core Contributions

  • Identifying premature commitment in embodied understanding. We identify a previously underexplored failure mode in which agents terminate exploration after forming self-consistent erroneous interpretations despite insufficient environmental evidence. This exposes a gap between interpretation plausibility and evidence sufficiency.
  • Shifting environmental interaction from answer-seeking exploration to hypothesis-challenging verification. Rather than treating a plausible answer as sufficient grounds to stop exploring, our approach actively challenges the agent's current interpretation before commitment. Evidence verification reexamines potentially misinterpreted observations, while coverage verification seeks missing decisive information. Both verifications jointly guide subsequent observations and answer acceptance. When verification leaves the candidate unresolved, retained feedback supports further exploration and revision of the underlying interpretation.

Rebuttal Responses and Manuscript Revisions

In response to the reviewers' feedback, we have clarified our methodological contributions, conducted additional experiments and analyses, and revised the manuscript accordingly. The key updates and supporting evidence are summarized below.

Methodological Clarifications

  • Reorganized Section 3 to explain the motivation for verification and how it governs the interaction policy, addressing Reviewer LXG5’s concerns about the contribution beyond component integration.
  • Refined the answer-change risk score as a heuristic decision criterion rather than a calibrated error probability, and claried how aggregation accounts for spatial relevance and redundant observations, addressing the risk-score interpretation and robustness questions raised by Reviewer LXG5.

Additional Experiments and Analyses

  • Added a comparison with a history-based uncertainty baseline, achieving 58.7/52.8 LLM-Match/LLM-SPL compared with 54.7/49.8, isolating the contribution of risk-guided viewpoint selection in response to Reviewer LXG5.
  • Expanded the repeated-run description for the main and ablation comparisons. Our method obtained significantly statistically improvement with all reported unadjusted p-values below 0.01,addressing Reviewers RamK and LXG5's concern about experimental variability.
  • Added an evaluation of 1,855 verifier judgments by GPT-6 Astra, complemented by a label-stratified human audit of 100 judgments. The respective accuracies are 86.52%/85.00%, with Macro-F1 scores of 84.06%/83.02%, addressing the judgment-reliability concerns raised by Reviewers LXG5 and RamK.
  • Extended the cost analysis beyond individual steps to episode-level. On EXPRESS-Bench, the baseline and our method require 567/775 seconds and 6.30/7.76 PFLOPs per episode, respectively, addressing the inference-cost questions raised by Reviewers qSEi and RamK.
  • Added comparisons with the adapted Self-Refine and Pelican baselines. Their A-EQA LLM-Match scores are 54.0 and 54.6 versus 58.7 for our method, as requested by Reviewer qSEi.
  • Added backbone and scene-complexity evaluations. With LLaVA-NeXT–Mistral-7B, our method achieves 54.2 LLM-Match on A-EQA compared with 48.6 for Pred-EQA, addressing the generalizability questions raised by Reviewer qSEi.
  • Added statistics on verification turns, feedback records, switches back to exploration, and rationale distributions. On A-EQA/EXPRESS-Bench, mean switches back to exploration are 0.83/0.44 per episode overall and 1.37/1.05 for initially incorrect candidates, as requested by Reviewer qSEi.

Other Clarifications

  • Explained how risk thresholds are selected on development data and how target visitation depends on the evidence/coverage risks and the remaining interaction budget, addressing the questions raised by Reviewer qSEi.
  • Clarified that the risk-score ablation removes risk guidance from viewpoint selection while preserving the acceptance rule, making the score’s role in evidence acquisition explicit in response to Reviewers qSEi and RamK.
  • Clarified the Qwen backbones used in the pilot experiment in Figure 1, referenced the corresponding protocol and analysis in Appendix A.1, and corrected the reported formatting issues, as requested by Reviewer qSEi.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.