{"id":"4156c0e1-315d-4a13-a8f1-a2dd37babc53","arxiv_id":"2611.17041","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LEAD-RL jointly trains main and subagents with contribution-aware credit assignment, improving accuracy and reducing critical-path cost in multi-agent deep research.","lead":"This paper introduces LEAD-RL, a training framework that jointly optimizes a main agent and its subagents in a multi-agent deep-research system. It uses a rubric to assess each subagent report's research utility and gates the team reward accordingly, improving answer quality and efficiency across five benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Contribution judge reliability is not independently validated, and the paper's central comparison (Table 3) may conflate rubric quality with the credit-assignment mechanism.","rationale":"The reader's weakest_assumption correctly identifies the reliance on an unvalidated rubric judge as a key concern. My analysis agrees and sharpens the point: the paper's central comparison does not disentangle the effect of the judge's accuracy from the credit-assignment mechanism itself. The proposed concrete test (random-label ablation) would directly address this. The paper's own acknowledgments of the hypothesis-like nature of the mechanism (b127, b181) support the need for such a test. I do not recommend changing the verdict from CONDITIONAL, because the concern is real but not fatal; the paper's results are still suggestive, and the proposed test is feasible. The agreement_with_reader is 'agree' because the reader's weakest_assumption and my load-bearing_attack converge on the same issue. The rationale explains why the concern is load-bearing and why the concrete test is appropriate.","tokens_in":17372,"tokens_out":727,"duration_ms":8815,"concrete_test":"Re-run the LEAD-RL training on BrowseComp with the contribution judge labels replaced by random labels (drawn from the same marginal distribution as the judge's labels, e.g., ~31% helpful, ~10% harmful, ~59% neutral) but still using the gating function of Eq. 6. If accuracy and critical-path cost remain similar to the reported LEAD-RL results, then the rubric's specific labels are not driving the improvement. If performance degrades to the level of joint RL with shared credit (or worse), then the judge's accuracy is essential. This test directly isolates the contribution of the judge's accuracy from the credit-assignment structure.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that contribution-aware credit assignment (LEAD-RL) improves accuracy and reduces critical-path cost relative to joint RL with shared team credit. The key mechanism is the rubric-based gate (Eq. 6) that filters the team advantage for each subagent report. However, the rubric judge (DeepSeek V4 Pro) is used without any independent validation of its accuracy, inter-rater reliability, or sensitivity to the specific prompts and evidence truncation described in Appendix A. The paper acknowledges that confidence below 0.8 is mapped to neutral to reduce noise, but this does not establish that the judge's labels are correct or consistent. If the judge frequently misclassifies helpful reports as harmful or vice versa, the gating function could distort credit assignment in unpredictable ways, potentially explaining the observed improvements (or lack thereof) without actually testing the credit-assignment hypothesis. The paper's controlled comparison (Table 3) shows that LEAD-RL outperforms joint RL with shared credit, but this comparison does not isolate the effect of the rubric: the two conditions differ in both the credit-assignment rule and the presence of the judge's labels. A proper ablation would compare LEAD-RL against a version that uses the same judge labels but applies them randomly or with a fixed gate (e.g., always neutral). Without such an ablation, the observed gains could be due to the judge's specific (possibly biased) labeling rather than the credit-assignment logic itself. The paper's own limitation statements (e.g., 'the mechanism remains a hypothesis rather than a directly measured effect') underscore this concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper introduces LEAD-RL, a multi-agent deep research framework that jointly trains a main agent and subagents using reinforcement learning with contribution-aware credit assignment. The key idea is to use a rubric-based judge to classify each subagent report as helpful, neutral, or harmful, and then gate the team advantage accordingly (Eq. 6). This prevents helpful reports from being penalized by team failures and harmful reports from being rewarded by team successes. The method is evaluated on five benchmarks (BrowseComp, BrowseComp-ZH, GAIA, HLE, DeepSearchQA) using two model families (Ling-3-Flash and Qwen3.6-35B-A3B). Controlled comparisons (Table 3) show that LEAD-RL improves accuracy by 2.49 percentage points and reduces critical-path cost by 24.5% over joint training with shared team credit on BrowseComp. Validation dynamics (Figure 2) show accuracy increases at approximately constant critical-path cost. The paper also reports comparisons with larger closed-source and open-source models.","tokens_in":17784,"tokens_out":1837,"duration_ms":16974,"significance":"If the results hold, the paper makes a meaningful contribution by addressing a specific credit-assignment problem in multi-agent delegation: the misalignment between team outcome and individual subagent report utility. The idea of using a rubric-based judge to gate the team advantage is novel and well-motivated. The experimental evaluation is thorough, covering multiple benchmarks and model families, with controlled ablations that isolate the effect of the credit-assignment mechanism. The paper also introduces useful efficiency metrics (critical-path cost, idealized speedup) that go beyond simple token counts. The main limitation is the reliance on an external judge (DeepSeek V4 Pro) for rubric labels without independent validation of its accuracy, which the authors acknowledge but do not fully address. Overall, the paper is a solid contribution to multi-agent RL for research agents.","major_comments":[{"comment":"The contribution judge (DeepSeek V4 Pro) is used without independent validation of its accuracy, inter-rater reliability, or sensitivity to the specific prompts and evidence truncation. The paper acknowledges that confidence below 0.8 is mapped to neutral (Appendix A), but this does not establish that the judge's labels are correct or consistent. The central comparison in Table 3 conflates the credit-assignment mechanism with the specific labeling behavior of this judge. A proper ablation would compare LEAD-RL against a version that uses the same judge labels but applies them randomly or with a fixed gate (e.g., always neutral). Without such an ablation, the observed gains could be due to the judge's specific (possibly biased) labeling rather than the credit-assignment rule itself. The authors should either provide evidence of judge accuracy (e.g., human evaluation on a sample of reports) or include an ablation that controls for the judge's presence.","section":"§3.3, Eq. (5)-(6), Appendix A","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The controlled comparison shows that LEAD-RL outperforms joint RL with shared team credit, but both RL variants underperform the SFT initialization in terms of accuracy (70.39% and 71.00% vs. 72.50%). This is a concerning pattern: the RL training appears to hurt performance unless the contribution-aware gating is used. The authors should discuss why main-only and shared-credit RL degrade accuracy relative to SFT, and whether this indicates a fundamental issue with the RL formulation (e.g., reward sparsity, instability) that LEAD-RL only partially mitigates. The fact that LEAD-RL recovers and slightly exceeds SFT accuracy is positive, but the degradation of the baselines raises questions about the robustness of the training pipeline.","section":"§4.4, Table 3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The validation dynamics (Figure 2) show that LEAD-RL improves accuracy at approximately constant critical-path cost, while baselines increase cost substantially. However, Appendix C (Figure 3) reveals that LEAD-RL's critical-path cost begins to rise more rapidly after approximately 50 updates without a sustained increase in batch correctness. The authors note this late-stage behavior and suggest that checkpoint selection should consider the joint accuracy-efficiency frontier. This is an important caveat: the reported gains (2.49 points, 24.5% cost reduction) are from a specific checkpoint, and the method may not be stable over longer training. The authors should clarify whether the reported results in Table 3 correspond to the best checkpoint on the validation set or the final checkpoint, and whether the gains are robust to checkpoint selection.","section":"§4.5, Figure 2, Appendix C","authors_acknowledge":"Because both questions and trajectories vary across updates, these measurements expose the behavior optimized during training but are not estimates of generalization performance.","authors_acknowledge_p":0.62}],"minor_comments":[{"comment":"The abstract states 'no token, step, or speedup reward is used' but the method does use a KL penalty coefficient (βL_KL in Eq. 8), which is a form of regularization that could affect token-level behavior. This is not a major issue but the phrasing could be clarified.","section":"§1, Abstract","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The normalization factor max(1, |I_j|) is used to prevent rollouts with more subagents from receiving disproportionately large aggregate updates. However, the denominator is the number of subagents with nonzero gated advantages, not the total number of subagents. This could incentivize the policy to produce more subagents with zero gated advantage (i.e., neutral or filtered out) to increase the per-subagent update. The authors should discuss this potential incentive.","section":"§3.3, Eq. (7)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The comparison with closed-source and large open-source models is informative but the paper correctly notes that external scores may use different scaffolds. The statement 'LEAD-RL (124B-A5.1B) outperforms DeepSeek V4 Flash (284B-A13B)' should be qualified: this is within the same swarm harness, but DeepSeek V4 Flash may perform differently under its own optimal harness. The paper already acknowledges this in §4.1 but the phrasing in §4.3 could be more precise.","section":"§4.1, Table 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The contribution rubric judge prompt is provided in full, which is good for reproducibility. However, the prompt includes the instruction 'Return JSON only' and specifies a schema. It would be helpful to report the average confidence scores and the proportion of labels mapped to neutral due to confidence < 0.8, to give readers a sense of how often the judge is uncertain.","section":"Appendix A","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The local and global verifiers are described as 'role-conditioned invocations of the shared policy' and 'use no reference answer at inference'. It would be helpful to clarify whether these verifiers are also trained during RL or are fixed. If they are trained, how does their training interact with the main agent and subagent updates?","section":"§3.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The table reports results for LEAD-RL on two model families, but the baseline Ling-3-Flash and Qwen3.6-35B-A3B results are from the same swarm harness. It would be helpful to also report the performance of these base models without any training (i.e., the SFT checkpoint) in the same table, to make the comparison more direct. Currently, the SFT initialization is only reported in Table 3 for BrowseComp.","section":"§4.2, Table 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The case study is illustrative but the paper correctly notes that it is 'not used in the aggregate evaluation' and the values are 'report contents rather than independently annotated ground truth'. This is appropriate, but the case study would be strengthened by including a failure case or an example where the rubric judge's classification is shown to be correct or incorrect.","section":"Appendix D","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"The paper's core idea is simple and well-motivated: when training a multi-agent research system, don't give every subagent the same team reward. Instead, use a rubric to classify each subagent report as helpful, neutral, or harmful, and gate the team advantage accordingly. This is a genuine contribution—prior work either trained only the main agent or broadcast the same signal to all agents. The experiments are thorough: five benchmarks, two model families, controlled comparisons that isolate the RL stage, and validation dynamics that show the accuracy-efficiency trade-off improving during training without an explicit cost penalty. The paper also ships a clear formulation of critical-path cost and idealized speedup, which are useful evaluation tools.\n\nThe main soft spot is the reliance on a proprietary rubric judge (DeepSeek V4 Pro) without independent validation of its accuracy or inter-rater reliability. The paper acknowledges this indirectly—it maps low-confidence judgments to neutral, and it states that the mechanism remains a hypothesis rather than a directly measured effect—but the central comparison (Table 3) could conflate rubric quality with the credit-assignment mechanism. A proper ablation comparing LEAD-RL against a version that uses the same judge labels but applies them randomly or with a fixed gate would strengthen the claim. That said, the paper's own limitation statements are honest, and the concern is not fatal: the judge is used only for credit routing, not as a reward signal, and the results are consistent across two model families and five benchmarks.\n\nMinor issues: no error bars or statistical significance tests for the main results, and code/data are not released. These are common in this area and don't undermine the core finding.\n\nWho this is for: researchers working on multi-agent RL for language agents, especially those interested in credit assignment for delegation. The paper deserves a serious referee—the idea is clean, the experiments are extensive, and the limitations are acknowledged. I would accept peer review and recommend the editor send it out.","headline":"LEAD-RL introduces a clean idea—rubric-gated credit assignment for subagent reports—and backs it with thorough experiments. The main soft spot is that the rubric judge's reliability is not independently validated, but the paper acknowledges this and the central claim holds up.","tokens_in":18335,"tokens_out":511,"duration_ms":7550},"pith_extraction":null,"created_at":"2026-11-13T07:11:08.381863+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":null,"supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}