- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes RevMMSearch, a framework for addressing target misidentification in multimodal search agents. It maintains an entity–evidence graph to track competing identity hypotheses and update the applicability of retrieved facts as identities change. The framework combines discriminative search with evidence attribution, while a revision-aware policy optimization objective suppresses outcome credit that conflicts with admitted local evidence judgments. Experiments across five benchmarks report an average accuracy improvement of approximately 3.5 percentage points over the strongest compared specialized search agent.
Critical strengths
Relevant motivation: The paper identifies a meaningful failure mode: factually correct information can still produce incorrect answers when it concerns a misidentified visual target.
Coherent framework: Identity revision is explicitly coupled with evidence applicability, addressing both search lock-in and the continued use of facts associated with rejected identities.
Targeted empirical analysis: Beyond the main benchmark comparisons, the paper includes component ablations and diagnostic analyses examining identity recovery, stale-evidence handling, and alternative training objectives.
Critical weaknesses
Weaknesses and Questions for the Authors
-
Insufficient validation of audit reliability. The method prioritizes admitted local verdicts over conflicting outcome credit, but their reliability is not directly validated. Citation and abstention rules alone do not establish accuracy, particularly since the same audit interface produces both the gate and verdict. Please report human agreement, admitted-verdict error rates, coverage/abstention rates, and sensitivity to noisy feedback or alternative auditors.
-
Unclear auditor implementation. Please specify the frozen auditor’s model, version, size, inference settings, and relationship to the policy and trajectory-generation teacher, including any weight sharing or additional training/calibration. These details are essential for reproducing the optimization signal.
-
Unreported training-time auditing costs. Auditing workload grows with rollout count and the number of action/state-update segments. Please report audit calls, token usage, batching/caching strategies, and additional training time and compute. Compare total costs against outcome-only GRPO and additive reward, preferably under matched compute budgets. Table 6’s execution statistics do not establish training-time auditing costs; this concern is distinct from deployment-time inference overhead.
Questions for authors
See in Weaknesses
4 comments
Follow-up on our rebuttal
Dear Reviewer XYwK,
Thank you for your constructive feedback on our paper. As the discussion period is drawing to a close, we would like to briefly follow up to ensure our response has adequately addressed your concerns. In particular, in our response and revised manuscript, we have:
- Clarified the auditor setup, available evidence, and abstention rules.
- Compared three auditors against human labels on 1,926 segments and tested how noisy audit feedback affects training.
- Reported auditing costs and compared RevGRPO with additive reward under the same compute budget.
Please let us know if any points remain unclear or require further clarification. We would be glad to address them before the deadline. If our revisions have resolved your concerns, we would greatly appreciate it if you could reconsider your assessment of our work.
Thank you very much for your time!
Follow-up on our rebuttal
Dear Reviewer XYwK,
I hope this message finds you well.
As the discussion period is drawing to a close, with fewer than two days remaining, we wanted to kindly follow up on our rebuttal and ask whether there are any additional questions or concerns we could address. We would be grateful for any further feedback you may have, as your comments are invaluable in helping us improve our work. If our response has resolved your concerns, we would be very grateful if you could kindly consider updating your score.
Thank you very much for your time and effort in reviewing our paper.
Best regards,
The authors
Final Follow-up on Our Rebuttal
Dear Reviewer XYwK,
Thank you again for your valuable feedback.
With the discussion deadline approaching, we would like to kindly follow up once more. In response to your concerns, we have provided detailed auditor specifications, human validation results, noise sensitivity experiments, and training-time cost comparisons.
We hope these additions have adequately addressed your questions. We would be very grateful if you could share any remaining concerns or, if our responses have resolved them, kindly reconsider your assessment before the discussion closes.
Thank you very much for your time and consideration.
Best regards,
The authors
Response to Reviewer XYwK
Thank you for the careful reading and constructive comments. Below, we address your concerns about the auditor. All revisions are highlighted in red in the updated manuscript.
Auditor implementation and selection
We use the official Qwen3-VL-30B-A3B weights as our auditor, with a decoding temperature of 0 and a maximum output length of 16,384 tokens. The policy model is Qwen3-VL-8B-Instruct, and the trajectory teacher is GPT-5.6-Sol. The auditor shares no weights with the policy or teacher, receives no additional training or calibration, and remains frozen throughout RL.
The auditor judges whether each action or state update is supported by the evidence available at that decision point. Action audits use the original image, question, current state, and interaction history before the action, excluding the observation returned by that action. State update audits may also use the newly returned observation. Reference answers, later observations, final answers, and outcome rewards are excluded. The auditor returns supported, unsupported, or abstain. On abstention, the original outcome credit is retained and the local reward is zero.
We randomly sampled 200 search trajectories and manually labeled their 1,926 action and state update segments using the information available at each decision point. We evaluated three candidate auditors on these same segments. Human annotators labeled each segment as supported, unsupported, or requiring abstention. Human agreement is the fraction of all segments whose audit labels match the human labels. The admitted-verdict error rate is the fraction of non-abstaining verdicts that disagree with human labels, including cases where the human label calls for abstention. Coverage is the fraction of all segments receiving a non-abstaining verdict.
We selected Qwen3-VL-30B-A3B to balance judgment quality and auditing cost.
Sensitivity to noisy audit feedback
To test the effect of incorrect feedback on training, we compare standard GRPO, RevGRPO without added noise, and RevGRPO with an additional 10% or 20% of admitted verdicts randomly flipped. All runs use the same SFT checkpoint, training data, and number of training steps. We keep the admission gates unchanged and recompute both the conflict mask and local reward using the flipped verdicts. Final answer accuracy on MMSearch is reported below.
RevGRPO improves over GRPO by 2.3 percentage points without added noise, but both noisy variants fall below GRPO. Thus, audit gains depend on verdict quality: errors can corrupt local rewards and outcome-credit suppression.
Thank you for the suggestion. We provide details on training-time auditing and compare estimated total costs over the same number of training steps.
Auditing cost
We use a frozen Qwen3-VL-30B-A3B on NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, with a batch size of 16 and 16 concurrent requests. Audit verdicts are cached for reuse. Training on 5,000 questions for two epochs, with five rollouts per question, produces about 50,000 trajectories and 482,000 audited segments. With an average of 1,000 input and 150 output tokens per audit, total usage is approximately 482M input and 72.3M output tokens. At 16,000 segments per hour on a single GPU, the estimated auditing cost is 30.1 GPU-hours.
Total cost comparison
All methods share the SFT checkpoint, training questions, update steps, and rollouts per question. Costs include rollout generation, rewards, auditing, and policy updates. Non-auditing RL costs ~400 GPU-hours per method; Additive Reward and RevGRPO share the same auditing workload:
At the same 430.1 GPU-hours, RevGRPO outperforms additive reward on all five benchmarks in Table 3 and surpasses outcome-only GRPO’s step-80 reward in fewer steps, as shown in Figure 6(c).
Thank you again for your constructive feedback. We welcome your further feedback and would be happy to address any remaining questions.