acceptodds
Under review as a conference paper at ICLR 2027

Cog-Research: Mitigating Grounding Drift in Long-Horizon Multimodal Deep Research

Abstract

Multimodal deep research agents can acquire reliable evidence yet fail to preserve it through long-horizon reasoning and report synthesis. We characterize this failure as Grounding Drift: previously established visual observations, claim–source bindings, or task constraints become omitted, corrupted, or misbound at later stages. We introduce Cog-Research, a post-training framework that converts intermediate research evidence into supervision for grounded report synthesis through source-linked, task-specific rubrics constructed from research traces, workspace artifacts, and human reports. We build a process-grounded dataset of 500 Plan-and-Act research tasks and perform trajectory SFT followed by rubric-guided GRPO with rewards for evidence-informed criterion coverage and report quality. Built on Qwen3-VL-8B-Instruct, Cog-Research improves the base agent by 4.85/3.54/4.91 points on MMDeepResearch-Bench, DeepResearch-Bench II, and MiroEval-Multimodal, respectively, and reduces the task-level incidence of primary Grounding Drift errors from 30.7% to 17.1% on MMDeepResearch-Bench.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multimodal Deep Research, Grounding Drift, Long-Horizon Reasoning, Reinforcement Learning, Deep Research Agents

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

This paper introduces Grounding Drift, a failure mode in long-horizon multimodal research where correctly established evidence or source/entity bindings become corrupted, omitted, or reassigned later in the reasoning process. To address this, the authors propose Cog-Research, which combines trajectory SFT with rubric-guided GRPO over Plan-and-Act traces, workspace artifacts, and task-specific evidence-linked criteria. Experiments on MMDeepResearch-Bench, DeepResearch-Bench II, and MiroEval show consistent improvements over the base model, including a reduction in Grounding Drift from 30.7% to 17.1%. Overall, the work is technically solid, well validated, and provides a useful framework for improving evidence preservation in multimodal deep-research agents.

Critical strengths

1、The paper identifies an important failure mode at the trajectory level. Requiring a correct upstream grounding followed by a downstream omission, corruption, or misbinding provides a concrete operational distinction from first-occurrence perception errors and unsupported hallucination.

2、The paper provides an annotation protocol, representative raw cases, an inter-annotator agreement study (Cohen's kappa = 0.766), and broader error-distribution analyses rather than relying only on final-answer metrics.

3、The experimental package is strong. The work reports three benchmarks, Base/SFT/SFT-RL comparisons, training-strategy ablations, reward-component ablations, run-to-run standard deviations, repetition diagnostics, atomic-capability retention metrics, compute costs, and reproducibility details.

Critical weaknesses

1、A report that preserves the same entities, values, and URLs but swaps their pairings can still receive substantial lexical/source-coverage reward. A controlled entity/value/source-swap stress test would make the mechanism more convincing, although this limitation is explicitly acknowledged and does not invalidate the current empirical results.

2、Some benchmark gains are intertwined with generic output stabilization. The Presentation Quality / hard-gate bundle has a large effect, and severe repetition drops substantially after RL. A more fine-grained separation of grounding-specific gains from output-quality stabilization would sharpen the causal story.

3、The novelty is strongest as a trajectory-level diagnostic lens and training formulation rather than as an entirely new primitive failure type. The appendix shows overlap with context loss, citation/source-reference errors, and entity confusion. I think the current framing is still useful, but this boundary should remain explicit.

4、The central 43-to-24 Grounding Drift reduction is presented as task counts from a diagnostic pass. A paired confidence interval, bootstrap analysis, or similar uncertainty estimate would further strengthen the main claim.

Questions for authors

1、Can the authors provide a controlled entity/value/source-swap evaluation of the deterministic reward, showing how much the reward changes when a correct report is modified only by swapping bindings while preserving the same tokens and URLs?

2、The Process-only variant has 23 primary Grounding Drift errors versus 24 for the full SFT-RL model. Do the authors interpret the main benefit of the joint objective as stronger cross-benchmark transfer rather than strictly lower drift incidence?

3、Can the authors provide a paired uncertainty analysis for the reduction from 43/140 to 24/140 Grounding Drift cases?

4、Could Presentation Quality, repetition penalties, and hard-error gates be ablated separately rather than as a bundle, to better isolate the contribution of grounding-specific supervision?

3 comments
AuthorsOpenReview ↗

Response to Reviewer aUwC

Thank you for your valuable feedback to help us improve our paper. We have revised our paper based on your feedback. We detail our response below and please kindly let us know if our response addresses your concerns.

W1. A report that preserves the same entities, values, and URLs but swaps their pairings can still receive substantial lexical/source-coverage reward. A controlled entity/value/source-swap stress test would make the mechanism more convincing, although this limitation is explicitly acknowledged and does not invalidate the current empirical results.

Q1. Can the authors provide a controlled entity/value/source-swap evaluation of the deterministic reward, showing how much the reward changes when a correct report is modified only by swapping bindings while preserving the same tokens and URLs?

Controlled relation-level stress test. The reviewer's quoted sentence described the token-coverage evaluator of the submitted manuscript. We did not test it against relation-level perturbations, and we presented the limitation as applying to the whole reward. This was too broad, and we correct it here. The source-indexed verifier resolves report citation markers through the reference list, indexes (source, entity, attribute, value) facts, and checks parseable claims against the cited source (Appendix A.5). Table R1 reports its decisions on eight tasks, including Q28, Q101, and Q133, with 100 controlled cases per class.

The source-indexed verifier flagged all 100 entity–value swaps, 100 citation-pointer swaps, and 100 same-token contradictions, but only 8 of 100 unsupported claims carrying a valid URL. Detection refers to verifier decisions, not to the magnitude of the final reward. The verifier therefore reliably detects parseable binding corruption, while verification of individual free-form claim support remains a stated limitation.

Table R1. Source-indexed verifier sensitivity to controlled relation changes.

Counterfactual classCasesDetectedΔreward
Entity–value swap100100+0.0002
Citation-pointer swap1001000.0000
Same-token contradiction100100−0.0013
Unsupported claim with valid URL1008+0.0038

It should be acknowledged that the detection does not cover everything. It applies to parseable assertions whose cited source has a sufficiently complete fact index. Unsupported free-form claims that carry a valid URL mostly remain undecided. The verifier is therefore a targeted check of binding relations, not a general semantic entailment judgment. Section 4.4.1 and Limitations (Appendix A.7) state these boundaries.


W2. Some benchmark gains are intertwined with generic output stabilization. The Presentation Quality / hard-gate bundle has a large effect, and severe repetition drops substantially after RL. A more fine-grained separation of grounding-specific gains from output-quality stabilization would sharpen the causal story.

Q4. Could Presentation Quality, repetition penalties, and hard-error gates be ablated separately rather than as a bundle, to better isolate the contribution of grounding-specific supervision?

We ablated Presentation Quality and Hard Gates separately rather than as one bundle (Table R2a, Table R2b).

Table R2a. Ablations on MMDeepResearch-Bench.

OverallGENEVIVEFMM
Full28.0746.1424.7329.9118.12
w/o Presentation Quality26.1042.9222.6924.3517.54
w/o Hard Gates24.0342.5720.5022.9718.91

Table R2b. Ablations on MiroEval-Multimodal.

OverallCoverageInsightInstruct. Follow.Clarity
Full53.2763.2953.8860.6655.29
w/o Presentation Quality50.7260.2652.6760.0850.79
w/o Hard Gates49.1156.6146.9162.5550.45

Removing Presentation Quality lowers GEN and Clarity (−3.22 and −4.50 points), while removing Hard Gates causes the larger degradation (Overall −4.04 on MMDeepResearch-Bench and −4.16 on MiroEval-Multimodal). Both removals also lower the grounding-specific metrics (EVI by 2.04 and 4.23 points; VEF by 5.56 and 6.94 points), so the controls are not purely stabilization terms. Their effect on grounding is nonetheless coupled to output stabilization (Appendix A.15).

W3. The novelty is strongest as a trajectory-level diagnostic lens and training formulation rather than as an entirely new primitive failure type. The appendix shows overlap with context loss, citation/source-reference errors, and entity confusion. I think the current framing is still useful, but this boundary should remain explicit.

We agree that this boundary should remain explicit, and the manuscript states it in Section 3 and Appendix A.1. Grounding Drift is proposed as a trajectory-level failure pattern rather than an entirely new primitive error type. It requires three conditions: (1) correct grounding was established upstream; (2) the information remains relevant to the task; and (3) a downstream state omits, corrupts, or misbinds it without supporting evidence. This temporal requirement distinguishes Grounding Drift from initial perception errors and isolated hallucinations.


W4. The central 43-to-24 Grounding Drift reduction is presented as task counts from a diagnostic pass. A paired confidence interval, bootstrap analysis, or similar uncertainty estimate would further strengthen the main claim.

Q3. Can the authors provide a paired uncertainty analysis for the reduction from 43/140 to 24/140 Grounding Drift cases?

For each pairing, the exact McNemar test and a 95% paired bootstrap interval over tasks are well defined, and Table R3 reports their envelope across pairings.

Table R3. Sensitivity bounds for the Grounding Drift reduction.

StatisticResult
Base Drift incidence43/140 (30.71%)
Full Drift incidence24/140 (17.14%)
Absolute reduction13.57 percentage points
Two-sided exact McNemar p-value across the 25 pairings0.0000038–0.0271
Envelope of 95% bootstrap intervals across pairings (Full − Base)[−25.00, −2.14] pp

The McNemar p-value does not exceed 0.0271 under any pairing, so the reduction does not depend on one particular pairing.


Q2. The Process-only variant has 23 primary Grounding Drift errors versus 24 for the full SFT-RL model. Do the authors interpret the main benefit of the joint objective as stronger cross-benchmark transfer rather than strictly lower drift incidence?

Yes. We interpret the main additional benefit of joint RL as stronger cross-benchmark generalization, rather than strictly lower Grounding Drift incidence (Section 5.4).

Table R4. Process-only vs. joint RL (paper Table 3 and Table 2).

MetricProcess-onlyFull
MMDeepResearch-Bench Overall27.9828.07
Primary Grounding Drift23/14024/140
MiroEval-Multimodal Overall49.5353.27

Process-only has the lowest observed primary Drift count, but the joint objective gives the stronger cross-benchmark result. Combining process and outcome rewards improves transfer across different deep-research evaluation settings.

AuthorsOpenReview ↗

We would like to hear back from reviewer aUwC

Dear reviewer aUwC,

We would like to follow up to see if the response addresses your concerns. We would really appreciate the opportunity to discuss this further if our response has not already addressed your concerns. Thank you again!

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

This submission studies grounding drift in long-horizon multimodal deep-research agents, where an agent may gradually lose alignment between the evidence it has collected and the claims made in its final report. The paper distinguishes this problem from visual perception errors and generic hallucination, and proposes a post-training framework that converts intermediate research evidence into supervision for grounded report generation. In particular, it constructs source-linked, task-specific rubrics from research traces, workspace artifacts, and human-written reports, and uses deterministic Python-based evaluators to provide process- and outcome-level rewards. The method is evaluated on three benchmarks: MMDeepResearch-Bench, DeepResearch-Bench II, and MiroEval.

Critical strengths

S1. The paper targets a meaningful and relatively well-defined failure mode. The distinction between grounding drift and more general visual perception errors or hallucinations is useful. Focusing specifically on whether claims remain connected to accumulated evidence over long research trajectories makes the problem concrete and relevant to practical deep-research agents.

S2. The use of deterministic evaluators is an appealing design choice. Using Python-based evaluators for process and outcome rewards reduces dependence on online LLM/VLM judges, potentially making the supervision more reproducible, less expensive, and less sensitive to judge-model instability.

Critical weaknesses

W1. The grounding supervision may be too weak to establish genuine claim–evidence alignment. The proposed evaluators appear to rely substantially on signals such as token overlap and URL presence. These signals may be easy to satisfy without actually grounding a claim in the corresponding evidence. For example, an agent could include relevant URLs or reuse rubric-related keywords while still making claims that are unsupported, misattributed, or only loosely connected to the cited sources. This raises the concern that the reward may measure superficial evidence mention rather than semantic claim–source entailment.

W2. There is a potentially important mismatch between the training and evaluation settings. The method is trained using a fixed offline evidence corpus, whereas the target setting involves agents operating with live search and dynamically retrieved information. It is therefore unclear how well the learned grounding behaviour transfers when the evidence distribution, source quality, or search trajectory changes at test time. The performance degradation associated with search information in Table 13 appears consistent with this concern and suggests that robustness to live-search evidence remains insufficiently demonstrated.

Questions for authors

Q1 How do the proposed evaluators distinguish true claim–evidence support from superficial correlations such as URL inclusion or lexical overlap? Have you tested adversarial cases where the correct URL and relevant keywords are present but the claim is unsupported or contradicted by the source?

Q2 Have you compared the deterministic grounding reward with stronger semantic alternatives, such as claim-level entailment, citation correctness, or source attribution metrics? If so, how strongly do they correlate?

Q3 How much of the performance gain comes specifically from improving claim–source grounding, rather than simply encouraging more citations or greater lexical similarity to the training rubrics?

Q4 How does the method generalize from the fixed offline evidence used during training to live-search settings with unseen sources and dynamically changing evidence?

Q5 Could the authors provide a more detailed analysis of the degradation reported in Table 13 for search information? In particular, does this drop result from retrieval noise, source-distribution shift, failures in evidence selection, or failures in associating retrieved evidence with generated claims?

4 comments
AuthorsOpenReview ↗

Response to Reviewer RLFz

Thank you for your constructive comments and suggestions. We have revised our paper according to your comments. We respond to your questions below and would appreciate it if you could let us know if our response addresses your concerns.

W1. The grounding supervision may be too weak to establish genuine claim–evidence alignment. The proposed evaluators appear to rely substantially on signals such as token overlap and URL presence. These signals may be easy to satisfy without actually grounding a claim in the corresponding evidence. For example, an agent could include relevant URLs or reuse rubric-related keywords while still making claims that are unsupported, misattributed, or only loosely connected to the cited sources. This raises the concern that the reward may measure superficial evidence mention rather than semantic claim–source entailment.

Q1. How do the proposed evaluators distinguish true claim–evidence support from superficial correlations such as URL inclusion or lexical overlap? Have you tested adversarial cases where the correct URL and relevant keywords are present but the claim is unsupported or contradicted by the source?

We tested such adversarial cases. Token-coverage and URL-consistency signals alone cannot distinguish support from superficial overlap, as the controlled step-branch comparison in Appendix A.18 shows. The report evaluator includes a source-indexed verifier for parseable claims (Appendix A.5). Table R1 reports its decisions on controlled perturbations. Entity–value swaps, citation-pointer swaps, and same-token contradictions are detected in all 100 cases per class. The adversarial case, a valid URL with an unsupported claim, is detected in only 8 of 100 cases. The deterministic reward is therefore not a complete semantic entailment verifier, which motivates the independent semantic audit below.

Table R1. Source-indexed verifier sensitivity to controlled relation changes.

Counterfactual classCasesDetectedΔreward
Entity–value swap100100+0.0002
Citation-pointer swap1001000.0000
Same-token contradiction100100−0.0013
Unsupported claim with valid URL1008+0.0038

It should be acknowledged that the detection does not cover everything. It applies to parseable assertions whose cited source has a sufficiently complete fact index. Unsupported free-form claims that carry a valid URL mostly remain undecided. The verifier is therefore a targeted check of binding relations, not a general semantic entailment judgment. Section 4.4.1 and Limitations (Appendix A.7) state these boundaries.


Q2. Have you compared the deterministic grounding reward with stronger semantic alternatives, such as claim-level entailment, citation correctness, or source attribution metrics? If so, how strongly do they correlate?

We conducted an independent semantic grounding audit covering 137 candidate claims from 48 reports on 12 matched tasks, with four sets of blinded annotations (Appendix A.16). Human annotators assessed whether the actual cited source supported each factual claim, and we compared these judgments against the deterministic reward.

Table R2. Correlation with independently audited semantic support.

Reward MetricSpearman ρ95% CI
Overall deterministic report reward0.281[0.095, 0.487]
Evidence-specific reward0.528[0.312, 0.712]

The evidence-specific reward shows a stronger association with independently audited source support than the overall reward. This suggests that the evidence component captures information related to semantic grounding despite its structural scoring rules.

Q3. How much of the performance gain comes specifically from improving claim–source grounding, rather than simply encouraging more citations or greater lexical similarity to the training rubrics?

To disentangle "citation coverage" from "correct source support," we conducted a blind evaluation; the results are presented in Table R3.

Table R3. Independent claim–source audit.

MetricSFTFullFull − SFT95% paired CI (pp)
Citation coverage among factual claims35.5%49.3%+13.8 pp[0.3, 24.9]
Correct source support among cited factual claims23.2%25.6%+2.5 pp[−11.0, 19.6]
Correct source support among all factual claims8.2%12.6%+4.4 pp[−1.7, 11.8]
Report-level support rate, equal report weights8.4%16.2%+7.7 pp[0.6, 14.8]

These results provide independent, albeit preliminary, evidence of improved semantic grounding beyond the training reward.

Source-provenance ablation. We trained matched Full and w/o provenance variants from the same SFT checkpoint, retaining the other reward components.

Table R4. Matched source-provenance ablation on MMDeepResearch-Bench.

Metricw/o ProvenanceFullGain
Overall26.0228.07+2.05
VEF27.3929.91+2.52
EVI23.5624.73+1.17

Removing provenance supervision lowers Overall, VEF, and EVI under identical SFT initialization and RL budget, supporting its contribution.


W2. The method is trained using a fixed offline evidence corpus, whereas the target setting involves agents operating with live search and dynamically retrieved information. It is therefore unclear how well the learned grounding behaviour transfers when the evidence distribution, source quality, or search trajectory changes at test time. The performance degradation associated with search information in Table 13 appears consistent with this concern and suggests that robustness to live-search evidence remains insufficiently demonstrated.

Q4. How does the method generalize from the fixed offline evidence used during training to live-search settings with unseen sources and dynamically changing evidence?

The fixed evidence corpus is used only during RL training, whereas all benchmark evaluations use live web tools (Appendix A.13). The reported improvements therefore already reflect transfer from offline training to live-search execution. To further examine sensitivity to changing evidence, we conducted controlled retrieval-perturbation experiments using cached search responses for three multimodal tasks, three checkpoints, and four evidence conditions.

Table R5. Correctly source-supported focal answers under evidence perturbations.

ModelOriginalSource SubstitutionReorderingAdded NoiseTotal
Base2/31/31/31/35/12
SFT2/31/31/32/36/12
Full2/32/32/32/38/12

Full keeps 8 of 12 focal answers source-supported across the original and perturbed conditions, versus 5 for Base and 6 for SFT, suggesting better robustness to changing evidence.

Q5. Could the authors provide a more detailed analysis of the degradation reported in Table 13 for search information? In particular, does this drop result from retrieval noise, source-distribution shift, failures in evidence selection, or failures in associating retrieved evidence with generated claims?

We analyzed the archived workspaces and trajectories to identify the observable causes of the Search Info decline. Search Info is a structural evidence-availability metric (Appendix A.11), which is not a direct measurement of claim–source correctness.

Table R6. Decomposition of Search Info failures using archived intermediate artifacts.

Observable failureBaseFull
No eligible URL-bearing intermediate artifact2835
Eligible artifact, but insufficient URL/host diversity20
Failed tasks3035
Passed tasks110105

These results localize the observed decline primarily to the availability and diversity of information search, rather than to evidence misassociation. Thus suggests a trade-off between making full use of evidence already obtained and continuing to explore new sources.

I have carefully reviewed the authors' rebuttal. Some core weaknesses remain unaddressed. Thus, I maintain original scores.

AuthorsOpenReview ↗

Response to Reviewer RLFz

Dear Reviewer RLFz,

Thank you for reading our rebuttal and for identifying the points that remain open. We agree that two concerns are central, and we state below what the evidence does and does not establish.

1. Semantic claim–source support (W1, Q1). The deterministic reward is not a semantic entailment verifier, and we do not claim it is. Its verified scope is the parseable relations described in Appendix A.5. Controlled entity–value swaps, citation-pointer swaps, and same-token contradictions are detected in 100/100 cases each, whereas unsupported free-form claims carrying a valid URL are detected in 8/100 and 79/100 (Table R1).

To check the sensitivity of the reward itself, we re-evaluated the same controlled cases with the fused report reward Rfused∈[0,1]R_{\mathrm{fused}}\in[0,1], which folds the new evidence criteria into the evidence component and then applies penalties, clipping, and caps. For case ii, ΔRfused,i=Rfused(yiperturbed)−Rfused(yivalid)\Delta R_{\mathrm{fused},i}=R_{\mathrm{fused}}(y_i^{\mathrm{perturbed}})-R_{\mathrm{fused}}(y_i^{\mathrm{valid}}). A case is reward-detected when ΔRfused,i<−10−9\Delta R_{\mathrm{fused},i}<-10^{-9}; cases within tolerance or with an increase are undetected. The threshold is applied per case, the mean is the average of the 100 paired differences computed at full precision, and the table shows four decimals.

Table R1. Sensitivity of the source-indexed verifier and of the fused reward to controlled relation changes (100 cases per class; Appendix A.17).

Counterfactual classCasesDetected (verifier)Detected (ΔR_fused)Mean ΔR_fused
Entity–value swap100100100−0.2700
Citation-pointer swap100100100−0.2700
Same-token contradiction100100100−0.1350
Unsupported claim with valid URL100879−0.0076

The unsupported-claim row is the case that matters for Q1. Under the verifier, most such claims remain undecided. Under the reward, 79 of 100 cases decrease beyond tolerance, but the mean change is only −0.0076. This is a sensitivity result: the reward moves, but only slightly, and this movement does not establish that a semantic error holds. Whether a semantic error holds is still decided by the verification status or manual judgment. This is a limitation of the current reward, and it is why we give the independent audit below higher weight.

2. Live-search generalization (W2, Q4). The controlled perturbation study (three tasks, three checkpoints, four evidence conditions, 36 cases) is small. We present it as a sanity check, not as evidence of robustness to arbitrary live-search distribution shift. Appendix A.19 already states this limitation: the study uses three tasks and cached evidence, and it complements rather than replaces the live-search results.

We hope these clarifications address the concerns you consider most important. If there are specific weaknesses you still consider unaddressed, we would be grateful if you could point to them, so that we can respond before the discussion closes.

Reviewer_nNL1OpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper identifies a specific failure mode in long-horizon multimodal deep research agents, termed Grounding Drift, where evidence to source and entity bindings that are correctly established upstream through perception and retrieval get dropped, corrupted, or misbound during downstream reasoning and report synthesis. A diagnostic on a base agent shows this correct-before and incorrect-after pattern accounts for a substantial share of primary errors and is separable from first-occurrence perception errors and hallucination. The paper defines Grounding Drift, releases a Plan-and-Act dataset with source-linked rubrics, and proposes Cog-Research, a post-training recipe combining trajectory SFT with rubric-guided GRPO, showing gains on three benchmarks and reduced Drift incidence.

Critical strengths
  1. The paper forms a complete diagnosis-to-method-to-evaluation loop, locating Grounding Drift by error attribution, writing it into the RL reward via rubrics.

  2. The work is well motivated.

Critical weaknesses
  1. Methodological novelty is limited. SFT, Plan-and-Act, GRPO, rubric-as-reward, and the process-plus-outcome composition are all established components from prior work, and the paper's contribution is marginal.

  2. The Presentation regression on DeepResearch-Bench II and the Instruction Following drop on MiroEval-Multimodal jointly suggest that the rubric weighting and process reward may be squeezing report readability and instruction adherence, yet the authors offer only a verbal explanation without an ablation on dimension weights or any analysis of possible reward hacking.

  3. The Appendix concedes that the reward evaluator cannot detect reports that swap values between entities while preserving entity names, numbers, and URLs—precisely the Q28-style failure used to motivate Grounding Drift—so the Drift reduction cannot be causally attributed to the source-linked rubric reward and more likely reflects SFT teacher imitation and the step reward's alignment to teacher trajectories, which substantially weakens the claimed effectiveness of the proposed method.

Questions for authors
  1. Since the appendix concedes the reward evaluator cannot detect value swaps that preserve entity names, numbers, and URLs. Could the authors ablate the source-metadata terms of the report reward to show the gain truly comes from source-linked supervision rather than teacher imitation.

  2. Since the step credit assignment relies on greedy one-to-one matching to teacher steps, which penalizes non-teacher-like but equally valid trajectories, could the authors compare it against globally optimal alignment and isolate how much of the step reward's gain reflects a genuine process signal rather than extended behavior cloning.

  3. The reward formulation stacks many hand-tuned components. but the paper neither justifies these choices nor demonstrates that each term contributes meaningfully during training, raising the concern that the design is over-parameterized and that only a small subset of terms is actually responsible for the reported gains.

6 comments
AuthorsOpenReview ↗

Response to Reviewer nNL1

Thank you for reviewing our paper and for your valuable feedback. Below, we address your concerns point by point and we’ve revised our paper according to your suggestions. We would appreciate it if you could let us know whether your concerns are addressed by our response.

W1. Methodological novelty is limited. SFT, Plan-and-Act, GRPO, rubric-as-reward, and the process-plus-outcome composition are all established components from prior work, and the paper's contribution is marginal.

We would like to clarify that our central contribution is not the invention of SFT, GRPO, or Plan-and-Act individually, but the identification and mitigation of Grounding Drift in long-horizon multimodal deep research:

  1. Problem definition and diagnosis. We characterize Grounding Drift through a verifiable correct-upstream → incorrect-downstream transition, separating it from hallucination and perception errors.
  2. Data and trajectories for Plan-and-Act multimodal research. Prior trajectory data has mostly been built on the ReAct framework, with little work on Plan-and-Act data, particularly in multimodal settings. Our automated production pipeline is novel and will be very useful for future multimodal reasoning research.
  3. Rubric-guided GRPO in multimodal environments. Adapting GRPO to multimodal plan-and-act agents raises a reward-decomposition problem. First, we design a process reward that scores the Planner's plan alignment and the Actor's matched execution steps, combined with the report reward through a joint objective. Second, we design source-linked rubrics that mitigate Grounding Drift.

Table R1. The comparison with closely related work.

Related workFrameworkMethod
WebWatcher [1]ReActsynthetic multimodal trajectories (GRPO)
Tongyi DeepResearch [2]ReActcustomized GRPO
Vision-DeepResearch [3]ReActmultimodal trajectory synthesis (GRPO)
MiroThinker [4]ReActDPO & GRPO
CaRR [5]ReActcitation-aware rewards (GRPO)
Co-Sight [6]Plan-and-ActTraining-Free
Cog-Research (ours)Plan-and-ActRubric-guided GRPO

W2. The Presentation regression on DeepResearch-Bench II and the Instruction Following drop on MiroEval-Multimodal jointly suggest that the rubric weighting and process reward may be squeezing report readability and instruction adherence, yet the authors offer only a verbal explanation without an ablation on dimension weights or any analysis of possible reward hacking.

The reward-hacking probe (Table R2) shows that the rule-based reward can be increased by keyword additions. Inspection of the generated reports shows that the trained policy did not exploit these shortcuts. Component-removal ablations (Table R7) quantify each dimension's contribution.

Table R2. Controlled reward-hacking probe (36 fixed reports).

ProbeReportsReward increasedTiedDecreasedMean Δ report reward
Keyword-only addition (isolated rubric-term list)363600+0.18814
Four repeated empty headings362781+0.00179

The probe confirms that the rule-based reward can be increased by keyword addition, which we treat as a known limitation. These probes measure evaluator sensitivity, not exploitation observed during policy rollouts.

[1] Geng, Xinyu, et al. "Webwatcher: Breaking new frontiers of vision-language deep research agent." International Conference on Learning Representations. 2026.

[2] Team, Tongyi DeepResearch, et al. "Tongyi deepresearch technical report." arXiv preprint arXiv:2510.24701.

[3] Huang, Wenxuan, et al. "Vision-deepresearch: Incentivizing deepresearch capability in multimodal large language models." Forty-third International Conference on Machine Learning. 2026.

[4] Team, MiroMind, et al. "Mirothinker: Pushing the performance boundaries of open-source research agents via model, context, and interactive scaling." arXiv preprint arXiv:2511.11793.

[5] Zhang, Jiajie, et al. "Chaining the evidence: Robust reinforcement learning for deep search agents with citation-aware rubric rewards." Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics. 2026.

[6] Zhang, Hongwei, et al. "Co-sight: Enhancing llm-based agents via conflict-aware meta-verification and trustworthy reasoning with structured facts." arXiv preprint arXiv:2510.21557.


W3. The Appendix concedes that the reward evaluator cannot detect reports that swap values between entities while preserving entity names, numbers, and URLs—precisely the Q28-style failure used to motivate Grounding Drift—so the Drift reduction \cannot be causally attributed to the source-linked rubric reward and more likely reflects SFT teacher imitation and the step reward's alignment to teacher trajectories, which substantially weakens the claimed effectiveness of the proposed method.

Q1. Since the appendix concedes the reward evaluator cannot detect value swaps that preserve entity names, numbers, and URLs. Could the authors ablate the source-metadata terms of the report reward to show the gain truly comes from source-linked supervision rather than teacher imitation.

The reviewer's concern is correct for the token-coverage terms of the evaluator, which we confirm in Table R6b. Our additional experiments address four points: whether the drift reduction is independent of the training reward, whether it holds under an independent semantic audit, whether source-provenance supervision contributes, and whether the report evaluator detects relational corruptions.

(1) Drift reduction is measured independently of the training reward. As reported in Table 2, the primary-error diagnostic yields from 43/140 to 24/140.

(2) Independent claim–source evaluation. We audited 137 candidate claims from 48 reports on 12 matched tasks using blinded source-support assessments.

Table R3. Independent claim–source audit.

MetricSFTFullFull − SFT95% paired CI (pp)
Citation coverage among factual claims35.5%49.3%+13.8 pp[0.3, 24.9]
Correct source support among cited factual claims23.2%25.6%+2.5 pp[−11.0, 19.6]
Correct source support among all factual claims8.2%12.6%+4.4 pp[−1.7, 11.8]
Report-level support rate, equal report weights8.4%16.2%+7.7 pp[0.6, 14.8]

Full cited a larger share of factual claims than SFT (+13.8 percentage points, 95% paired CI [0.3, 24.9]). Its estimated rate of correctly source-supported claims was also higher (+4.4 points; 95% paired CI [−1.7, 11.8]). These results provide independent, albeit preliminary, evidence of improved semantic grounding beyond the training reward.

(3) Source-provenance ablation. We trained matched Full and w/o provenance variants from the same SFT checkpoint for 100 GRPO steps, retaining the other reward components.

Table R4. Matched source-provenance ablation on MMDeepResearch-Bench.

Metricw/o ProvenanceFullGain
MMDR Overall26.0228.07+2.05
VEF27.3929.91+2.52
EVI23.5624.73+1.17

Removing provenance supervision lowers Overall, VEF, and EVI under identical SFT initialization and RL budget, supporting its contribution.

(4) Controlled relation-level stress test. The reviewer's quoted sentence described the token-coverage evaluator of the submitted manuscript. We did not test it against relation-level perturbations, and we presented the limitation as applying to the whole reward. This was too broad, and we correct it here. The source-indexed verifier resolves report citation markers through the reference list, indexes (source, entity, attribute, value) facts, and checks parseable claims against the cited source (Appendix A.5). Table R5 reports its decisions on eight tasks, including Q28, Q101, and Q133, with 100 controlled cases per class.

Table R5. Source-indexed verifier sensitivity to controlled relation changes.

Counterfactual classCasesDetected (verifier)Detected (ΔR_fused)Mean ΔR_fused
Entity–value swap100100100−0.2700
Citation-pointer swap100100100−0.2700
Same-token contradiction100100100−0.1350
Unsupported claim with valid URL100879−0.0076

It should be acknowledged that the detection does not cover everything. It applies to parseable assertions whose cited source has a sufficiently complete fact index. Unsupported free-form claims that carry a valid URL mostly remain undecided. The verifier is therefore a targeted check of binding relations, not a general semantic entailment judgment. Section 4.4.1 and Limitations (Appendix A.7) state these boundaries.

Q2. Since the step credit assignment relies on greedy one-to-one matching to teacher steps, which penalizes non-teacher-like but equally valid trajectories, could the authors compare it against globally optimal alignment and isolate how much of the step reward's gain reflects a genuine process signal rather than extended behavior cloning.

We separate two questions: whether the matching procedure is suboptimal, and whether the step reward carries a validity signal beyond similarity to teacher-style decomposition. The first is answered by global-assignment control (Table R6a). The second is answered by a controlled comparison (Table R6b).

(1) Greedy vs. globally optimal assignment. We replaced the greedy matcher with Hungarian maximum-weight one-to-one assignment on identical scoring matrices, for plan and execution-step alignment.

Table R6a. Alignment comparison on 12 groups of four rollouts.

MetricOptimal Plan AlignmentOptimal Execution Alignment
Mean joint-reward increase+0.014+0.001
Joint-reward ranking reversals1/720/72
Changed top-ranked rollout groups0/120/12

Although globally optimal alignment improves matching scores, the greedy approximation has limited impact on joint-reward rankings in this evaluation. We attribute this to two factors: ① the Deep Research tasks involve similar sequences of execution steps; and ② the atomicity of the plans means that no steps are combined (e.g., searching A and B simultaneously).

(2) Teacher-likeness versus validity. To test whether the step reward distinguishes valid from teacher-like trajectories, we scored three fixed final reports (G251, G368, G479), each in four variants:

  • Teacher-like, valid: follows the teacher's step with correct bindings.
  • Teacher-like, invalid: follows the teacher's step decomposition but misbinds facts.
  • Non-teacher-like, valid: merges steps and with correct bindings..
  • Non-teacher-like, invalid: merges steps and misbinds facts.

Table R6b. Step reward under controlled teacher-likeness × validity variants (greedy matching unless stated).

TaskTeacher-like, validTeacher-like, invalidNon-teacher-like, validNon-teacher-like, invalidGap, greedy
G2510.8230.8230.3480.348+0.475
G3680.9040.9040.3110.311+0.593
G4790.9290.9290.3040.304+0.625

Gap = teacher-like invalid − non-teacher-like valid.

Across classes, teacher-like trajectories score 0.82–0.93, and valid non-teacher-like plans score 0.30–0.35. The gap is 0.47–0.62 under greedy matching and remains of the same size under optimal plan or execution matching. We acknowledge that token-coverage criteria cannot see bindings, , and we state this explicitly.

Q3. The reward formulation stacks many hand-tuned components. but the paper neither justifies these choices nor demonstrates that each term contributes meaningfully during training, raising the concern that the design is over-parameterized and that only a small subset of terms is actually responsible for the reported gains.

Table R7 reports ablations of the dimension-associated reward components on MMDeepResearch-Bench. Each reward term is motivated a priori by a diagnosed failure mode in Section 3 and Table 2. The weights assign the largest share (0.60 in total) to the two grounding dimensions, which are the primary target of this work, and the remaining dimensions follow in decreasing order of diagnostic relevance.

Table R7. Ablations on MMDeepResearch-Bench.

OverallGENEVIVEFMM
Full28.0746.1424.7329.9118.12
w/o Visual Grounding26.7545.4025.1121.5516.09
w/o Evidence Grounding25.9445.9420.8826.5017.29
w/o Task Following26.9144.3324.0628.1617.30
w/o Quantitative Analysis27.1645.6024.4227.3917.45
w/o Presentation Quality26.1042.9222.6924.3517.54
w/o Hard Gates24.0342.5720.5022.9718.91

The contributions are therefore distributed across the design rather than concentrated in a small subset of terms. Whether the weights are optimal is left to future work (R1-W2).

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of Paper Revision

We sincerely appreciate all reviewers for their insightful and constructive feedback. According to these comments, we have improved the paper (new pdf uploaded) and highlighted the main changes with blue text. Below, we summarize all changes:

Reviewer nNL1

  1. Source-linked rubrics and verification. Section 4.2 describes rubric source metadata and the source-indexed verifier (Page 5, Line 266–269). Section 4.4.1 states the scope of claim–source verification and its fallback rules (Page 6, Line 312–317).
  2. Presentation regression. Section 5.2 now reports the Presentation decrease during SFT and its partial recovery after RL (Page 8, Line 406–411). The output-quality analysis and reward-hacking probes are in Appendix A.15 (Page 30–31, Line 1595–1624).
  3. Teacher-likeness versus validity. Section 5.4 states that the step branch tracks teacher-style decomposition rather than validity (Page 9, Line 468–473). Appendix A.18 reports the controlled teacher-likeness × validity comparison and the greedy-matching analysis (Page 32–33, Line 1695–1730).
  4. Reward component ablations. Presentation Quality and Hard Gates are separated in Table 4 (Page 9, Line 443–450) and discussed in Section 5.5 (Page 9, Line 482–485).
  5. Weights and design priors. Appendix A.5 states that the reward weights, λ, and hard-error caps are design priors not searched on the evaluation benchmarks (Page 19, Line 998–1000).
  6. Source-indexed verifier. Appendix A.5 gives the verifier and its effect on the evidence reward (Page 20, Line 1026–1050). Appendix A.7 states its limitations (Page 25, Line 1315–1320).

Reviewer RLFz

  1. Adversarial and relation-level tests. Section 4.4.1 describes the verifier's scope (Page 6, Line 312–317). Appendix A.17 reports controlled counterfactual tests, including unsupported claims with valid URLs (Page 31–32, Line 1667–1693).
  2. Independent semantic audit. Section 5.3 introduces the blinded claim–source audit (Page 8, Line 431–436). Appendix A.16 reports citation coverage, correct-support rates, and their correlation with the deterministic reward (Page 31, Line 1627–1647).
  3. Source-provenance ablation. A matched 100-step ablation of provenance supervision is reported in Appendix A.16 (Page 31, Line 1649–1665).
  4. Live-search transfer. Appendix A.13 notes the controlled evidence-perturbation experiments (Page 28, Line 1472–1473). Appendix A.19 reports them, together with the Search Info decomposition (Page 33, Line 1732–1767).
  5. Search Info. Appendix A.11 clarifies that Search Info is a structural traceability proxy rather than a claim–source support metric (Page 27, Line 1433–1436). Appendix A.19 gives its failure decomposition (Page 33, Line 1753–1767).

Reviewer aUwC

  1. Controlled entity/value/source-swap stress test. Appendix A.17 reports the counterfactual results for entity–value swaps, citation-pointer swaps, and same-token contradictions, together with the unsupported-claim case (Page 31–32, Line 1667–1693).
  2. Separated ablations. Presentation Quality and Hard Gates are separated in Table 4 (Page 9, Line 443–450), in Section 5.5 (Page 9, Line 482–485), and in Appendix A.5 and A.15 (Page 23, Line 1201–1205; Page 30, Line 1598–1603).
  3. Process-only versus joint RL. Section 5.4 states that the main benefit of the joint objective is cross-benchmark transfer rather than a lower Drift count (Page 9, Line 468–473).
  4. Uncertainty analysis. Appendix A.20 reports the sensitivity bounds for the Grounding Drift reduction across all 25 admissible pairings. (Page 33, Line 1769–1779).
  5. Output-quality discussion. Section 5.2 reports the Presentation and Instruction Following trends across training stages (Page 8, Line 406–411).

Supplementary Material

Added the anonymized code files corresponding to the paper.

Dear Reviewers,

We are approaching the deadline of the discussion period (Nov. 18). Please go through authors' rebuttal at your earliest convenience, and engage in a discussion or leave any follow-up comments.

Thank you for your time and participation.

AC

AuthorsOpenReview ↗

Rebuttal Summary for AC

Dear Program Chairs, Senior Area Chairs, Area Chairs, and Reviewers,

We sincerely thank you for your time and constructive feedback. We have responded point by point and revised the manuscript; the changes are summarized in the Summary of Paper Revision and highlighted in blue.

We have addressed the concerns raised by Reviewer aUwC (Rating 3); we followed up after our responses and have not received further comments. Reviewers nNL1 and RLFz maintained their scores. Reviewer nNL1's remaining concern is novelty beyond prior work. We have clarified our three contributions: a temporal diagnosis of Grounding Drift as a correct-before/incorrect-after failure, a Plan-and-Act multimodal research dataset with source-linked rubrics, and rubric-guided GRPO with process and outcome rewards. Reviewer RLFz's remaining concerns are the scope of the reward's semantic verification and generalization to live search. We have stated the scope of these claims precisely: the source-indexed verifier targets parseable relations, the step branch is teacher-referenced, and the claim-level support gain is preliminary (Section 5.4; Appendices A.16 and A.17).

We understand this is a busy period for ACs. We hope the outcomes of the rebuttal process will be fully considered in your final decision, and we deeply thank you for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.