{"id":"0540df5a-4de5-422a-b151-e84b928bd1e0","arxiv_id":"2608.11949","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Team roles induced from past trajectories and expressed as prompts and token-level markers improve multi-agent LLM performance on two multi-hop QA benchmarks but not on a full-Wikipedia retrieval stress test.","lead":"This paper introduces ExRole, a framework that learns role definitions for AI agent teams from past collaboration traces and turns them into instruction prompts and model-side routing signals. In tests on two multi-hop question answering benchmarks, the role-trained teams outperformed single-agent and role-free baselines, though a harder full-Wikipedia retrieval test did not show the same benefit.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's 2Wiki claim of a 7.7-point EM gain over the strongest control is not statistically supported: that control is the shuffled-role condition, and the margin is within sampling noise for 200 examples.","rationale":"The paper is a serious method paper with a clear controlled design, a provenance and split-isolation audit for the MuSiQue role library, a strict 2x2 ablation, and an honest HotpotQA stress test that explicitly bounds the generalization claim. The reader's stated weakest assumption, the manually weighted reward in Appendix B.1 Equation 21, is a real limitation, but it is not the most load-bearing threat to the central claim because all compared arms share the same reward and evaluation protocol; a reward perturbation would change absolute scores but would not by itself explain the ExRole-versus-no-role advantage. The more direct threat is statistical: the paper's headline improvement over the 'strongest non-ExRole control' on 2Wiki is 7.7 EM, and that control is the shuffled induced-role condition rather than a role-free baseline. For 200 examples the difference is within sampling noise, so the specific quantitative claim in the abstract is not supported without paired intervals or more evaluation data. This does not overturn the paper's overall direction, because the MuSiQue comparisons are large and the 2Wiki comparison against no-role MAS is also substantial; it means the exact headline margin over the strongest control should be reported with uncertainty or replaced by the more robust no-role comparison. The reader already flagged the absence of confidence intervals in the main table, so this reinforces the CONDITIONAL verdict rather than changing it. The lack of code or data artifacts further limits independent verification, but the disclosed diagnostics and limitations give enough internal evidence to treat the central mechanism as plausible pending the requested interval analysis.","tokens_in":22247,"tokens_out":15394,"duration_ms":154991,"concrete_test":"Using the saved per-example predictions on the 200-example 2Wiki split, compute paired-bootstrap 95% confidence intervals (20,000 resamples) and win/tie/loss counts for ExRole-Routed versus each Table 1 control, specifically the shuffled induced-role condition. If the EM-difference interval versus shuffled includes zero, the abstract should not claim a 7.7-point gain over the strongest control; the supported margin is instead the 12.0-point gain over the no-role MAS. If per-example predictions are unavailable, rerun the shuffled-role control with three seeds on the same 200 examples and report the paired ExRole-versus-shuffled bootstrap interval. The same analysis should be run on MuSiQue to confirm that the 11.5-point gain over no-role remains significant.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1 reports no confidence intervals for ExRole versus any control. On 2WikiMultiHopQA, the strongest non-ExRole control is not a role-free baseline but the shuffled induced-role condition (42.3 EM), and ExRole-Routed's advantage is 7.7 EM. With 200 binary examples, p1=0.50 and p2=0.423 give an approximate standard error of 4.1-5.0 percentage points for the difference (the lower value if the shuffled score is treated as a three-seed average), so a 95% paired or unpaired interval includes zero. The F1 margin of 9.7 points may survive, but the EM margin is a headline number in the abstract and is not established. The MuSiQue leg is more robust: 31.5 vs 20.0 EM is a difference of roughly 2.7 standard errors. The problem is specific to the 2Wiki 'strongest control' comparison, and it is compounded by the fact that the identity of the strongest control shifts across benchmarks (no-role on MuSiQue, shuffled on 2Wiki) and by the shuffled control's opposite behavior on the two benchmarks. Table 10 provides paired CIs only for Routed minus Shared, not for ExRole versus the controls that define the paper's central quantitative claims.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ExRole, a framework that learns \"executable roles\" from prior multi-agent LLM trajectories and uses them to condition team behavior and, in the Routed variant, to allocate shared LoRA capacity via role-conditioned sparse routing. Role induction is performed by training a prefix-local encoder with next-action, future-evidence, and final-return targets, clustering the embeddings, and deterministically resolving clusters into readable instructions and token-aligned markers. The policy is trained with turn-aligned GRPO, and a Shared variant is presented as a matched control for the Routed variant. The method is evaluated on MuSiQue and 2WikiMultiHopQA, comparing against single-agent search, no-role and manual-role MAS, random role prompts, and shuffled induced roles, with additional ablations and a HotpotQA stress test.","tokens_in":22540,"tokens_out":8554,"duration_ms":82568,"significance":"The core idea of making roles executable control variables that bridge behavioral abstraction, prompt conditioning, and parameter routing is a useful and timely contribution to multi-agent LLM systems. The paper ships a deterministic template resolver that avoids an extra role-labeling model, provides a split-isolation audit for the role library (Appendix B.8), and reports a matched Shared/Routed comparison with paired bootstrap CIs (Table 10). The MuSiQue result is reasonably robust (11.5 EM over the no-role control is roughly 2.5 standard errors), and the HotpotQA stress test in Appendix D.6 honestly reports a boundary where induced roles do not beat the strongest controls. However, the statistical support for the 2Wiki headline claim and the absence of reward-weight sensitivity analysis are significant caveats that presently limit confidence in the strength of the central claims.","major_comments":[{"comment":"The headline 2WikiMultiHopQA EM gain of 7.7 points over the strongest non-ExRole control is not statistically supported. With 200 binary examples, the difference between 50.0 EM (ExRole-Routed) and 42.3 EM (shuffled induced role) has an approximate standard error of 4–5 percentage points, so the 95% confidence interval includes zero; the F1 margin of 9.7 points is more likely to survive. Table 10 reports paired bootstrap CIs only for Routed minus Shared, not for ExRole versus the controls that define the central claim, and the identity of the strongest control shifts from no-role on MuSiQue to shuffled on 2Wiki, with the shuffled control behaving oppositely on the two benchmarks (8.5 EM on MuSiQue versus 42.3 on 2Wiki). Please provide paired or clustered bootstrap confidence intervals for ExRole versus each control on both benchmarks, or report the comparison as exploratory where the interval includes zero.","section":"Table 1 and Abstract"},{"comment":"The turn-level reward in Eq. (21) is manually weighted with several positive and penalty terms, and the paper acknowledges in Limitations C.1 that \"The turn-level reward components are manually weighted.\" Because the same reward defines the role-induction targets (Eq. 5), the turn-aligned credit (Eqs. 16–18), and the GRPO objective, the induced roles and the measured gains are all contingent on this specific reward design. No sensitivity analysis for any reward coefficient is provided. Please include at least one alternative weighting scheme (or a one-at-a-time perturbation of the major coefficients) to show that the qualitative ranking in Table 1 is robust to reward misspecification.","section":"Appendix B.1 (Eq. 21) and Limitations C.1"},{"comment":"The training status of the Table 1 controls is unclear. The paper states that all controlled systems share the same policy backbone, action interface, and evaluation protocol, and the manual-role control description mentions a \"training configuration,\" but it is not stated whether the single-agent, no-role, manual-role, random-role, and shuffled-role controls are optimized with the same turn-aligned GRPO objective as ExRole. This matters because the \"w/o role induction\" ablation in Table 6 (26.5 EM) differs substantially from the no-role MAS in Table 1 (20.0 EM) under the same shared-capacity setting, suggesting different training conditions. Please state explicitly, for each row of Table 1, whether the model was trained with the same GRPO objective, the same reward, and the same number of updates as ExRole, and if not, report the trained controls or restrict the claim accordingly.","section":"Table 1 and Table 6"}],"minor_comments":[{"comment":"The Succ. column exactly equals the EM column for all primary-benchmark rows, and Appendix C.2 reports D_Succ,EM = 0. Consider stating that this agreement is a property of the current predictions rather than a definitional identity, since the two metrics are defined differently.","section":"Table 1 and Appendix C.2"},{"comment":"The cross-seed agreement for K = 3 shows ARI ranging from 0.61 to 0.85 and NMI from 0.62 to 0.82, indicating moderate stability. The paper should mention this degree of variability in the main text rather than only in the appendix.","section":"Appendix D.2"},{"comment":"The abstract claims that \"the controlled results consistently favor trajectory-induced role conditioning over role-free, manual, random, and shuffled alternatives.\" This is accurate for the two primary benchmarks, but the HotpotQA stress test in Appendix D.6 shows the opposite. Consider adding a qualifier such as \"on the primary supporting-document benchmarks\" to avoid overgeneralization.","section":"Abstract"},{"comment":"The three-bar groups in Figure 1 have no per-bar color legend. Adding a legend or directly labeling the Manual, Random, and Shuffled bars would improve readability.","section":"Figure 1"},{"comment":"The symbol Ld (the total number of adapted module–rank slots) is used in Eq. (13) and later but is not defined at first use in the main text; a brief definition would help the reader.","section":"Section \"Role-Conditioned Sparse Routing\" (Eq. 13)"}],"recommendation":"major_revision","confidential_remarks":"The paper presents an interesting framework and the MuSiQue evidence is reasonably solid. The 2Wiki headline comparison, however, lacks the statistical support needed for an abstract-level claim, and the absence of clarity about whether the controls receive the same RL training is a fairness issue that must be resolved. The reward-weight sensitivity is also important given how much of the method depends on that reward. I would not reject the paper; these concerns are addressable with additional experiments and reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ExRole is a serious method paper: it induces roles from logged team trajectories, resolves them into prompt instructions and token markers, and optionally routes shared LoRA rank slots by role. The controlled setup is the best part—role-free, manual, random, and shuffled controls under a common interface, plus a provenance audit that checks the role-induction split doesn't leak. The Role-Agent-Turn disentanglement is also genuinely informative: roles predict action beyond agent identity and turn phase, and the cross-agent consistency measures are convincing. They even disclose a HotpotQA stress test where the method fails to beat controls, which is more honesty than most papers show.\n\nThe main quantitative claim needs work. The MuSiQue gains are solid: 31.5 vs 20.0 EM is well outside noise. But the 2Wiki headline—7.7 EM over the strongest control—is not established because that control is the shuffled-role condition and the margin is within sampling error for 200 examples. The abstract should not lead with that number. The F1 margin may hold, but the EM one doesn't, and the paper reports no CIs for these comparisons. The shuffled control behaving terribly on MuSiQue and best on 2Wiki also suggests something odd, and the manual-role baseline being below no-role makes that control weaker than it should be.\n\nThe routing variant is honestly presented as not reliably better than the shared path, but that undercuts the 'executable role' framing: the parameter-level path is the headline architectural contribution and it doesn't pay off. The turn-reward coefficients are manually weighted, and there's no sensitivity analysis; the HotpotQA result cautions against overclaiming generalization.\n\nOverall: the trajectory-induced prompt-level role conditioning result is plausible and useful; the parameter-level routing claim is not supported; the evaluation needs more rigor before the numbers are trustworthy. This deserves a serious referee—the framework and diagnostics are worth engaging with—but the authors should be pushed to add multiple seeds, confidence intervals, and to de-emphasize the 2Wiki EM claim.","headline":"A serious role-induction framework with honest reporting, but the headline 2Wiki gain is within sampling noise and the routed-capacity variant doesn't justify its billing.","tokens_in":23175,"tokens_out":2392,"would_cite":true,"duration_ms":24103,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Roles induced from past team trajectories outperform hand-written and role-free role prompts in multi-agent LLM search, with 11.5–16.1 point EM/F1 gains over the strongest controls.","keywords":["multi-agent language models","role induction","executable roles","multi-hop question answering","trajectory learning","LoRA routing","turn-aligned credit assignment","retrieval-augmented reasoning"],"falsifier":"Run ExRole-Shared from scratch on the same 200-example MuSiQue split with the reward coefficients in the turn-level formula perturbed (for example, doubling $\\lambda_{\\text{ans}}$ and $\\lambda_{\\text{ver}}$, or zeroing all penalty terms) and measure EM/F1 against the single-agent and no-role controls; if the induced-role advantage shrinks to within noise, the reward's manual weighting, not trajectory-derived roles, is carrying the result.","tokens_in":21942,"feed_emoji":"🤖","tokens_out":10872,"duration_ms":96721,"temperature":0.7,"pith_summary":"This paper claims that in multi-agent language-model teams, a useful role is not a hand-written prompt label but an executable control variable: it should summarize behavior that predicts future utility, guide subsequent interaction, and identify the trainable capacity responsible for that behavior. The framework induces role prototypes from prefix-local team traces, resolves them into readable instructions and token-aligned role markers, and uses the same role identity to condition prompts, route shared LoRA rank slots, and assign turn-level credit during policy optimization. On MuSiQue and 2WikiMultiHopQA, trajectory-induced roles beat single-agent search by 15.0/14.4 and 13.5/16.1 EM/F1 points, and the strongest role-free or hand-role controls by 11.5/11.6 and 7.7/9.7 points. A reader should care because role engineering is a bottleneck in building LLM agent teams; this work offers a way to learn roles from data and make them concrete inside the model.","feed_headline":"Team roles learned from logs beat hand-written prompts by 19 EM points","feed_subtitle":"Trajectory-derived roles plus turn-level credit lift multi-agent QA by 11–16 points over role-free and single-agent baselines.","key_machinery":"The load-bearing mechanism is the trajectory-to-role pipeline. Each logged agent turn is encoded into a 30-coordinate prefix-local feature vector covering action rates, timing, team context, evidence use, and answer behavior, and a small encoder is trained to predict the same agent's next action, future evidence hit, and final trajectory return; Euclidean K-means over these future-aware embeddings yields role prototypes. A deterministic resolver scores each prototype against a fixed functional vocabulary and binds it to a readable instruction and a token-aligned role marker, making the induced role an executable identity. That identity drives two complementary paths: it conditions the active agent's observation and prompt, and it computes a balanced sparse-delta gate over shared LoRA rank slots in the routed variant, while turn-aligned GRPO adds discounted local turn returns to the role-turn segment's tokens and trains the router to predict that credit. ExRole-Shared keeps the same roles and credit but uses a uniform LoRA path, isolating the effect of role-conditioned routing.","core_discovery":"The central discovery is that roles recovered from the statistics of past team trajectories, rather than from hand-authored descriptions, can function as executable control variables: the same induced identity conditions the active agent's instruction, stays aligned with generated role-turn tokens, selects a balanced subset of shared LoRA rank slots, and receives discounted turn-level credit in optimization. In the controlled comparison, ExRole-Shared reaches 31.5 EM and 43.2 F1 on MuSiQue, 15.0/14.4 points above single-agent search and 11.5/11.6 points above the strongest non-ExRole control; ExRole-Routed reaches 50.0 EM and 59.7 F1 on 2WikiMultiHopQA, 13.5/16.1 points above single-agent search and 7.7/9.7 points above the strongest control. The paper also reports that adding role identity to an action model that already observes agent identity and turn phase reduces held-out action log loss by roughly 19%, indicating that induced roles encode behavioral specialization beyond fixed agent indices or turn positions.","pith_inferences":["Editorial inference: a role library induced on one multi-hop QA benchmark should provide a useful warm start on a second benchmark with the same team interface, since the fixed agent-role assignment is an implementation choice rather than a requirement of the learned roles.","Editorial inference: the deterministic template resolver restricts roles to a fixed vocabulary (researcher, coordinator, verifier, analyst), but the induction step itself does not require that vocabulary; an open-vocabulary or learned resolver could discover roles the templates cannot name.","Editorial inference: the cleanest stress test of the whole framework is a reward-sensitivity sweep, because the turn-level reward coefficients are manually weighted and never varied; one would expect the induced-role advantage to survive plausible coefficient perturbations if the roles reflect real behavioral specialization rather than reward-shaped artifacts.","Editorial inference: if the mechanism generalizes, router gate deltas and role-turn markers give a token-level record of which capacity a role used, turning black-box team behavior into inspectable specialization that could be used for auditing and debugging agent teams."],"forward_implications":["If ExRole's results hold, role definition for LLM agent teams can shift from hand-written prompts to a learned, auditable artifact: a library of prototypes with deterministic instructions and markers, refreshable when the task distribution changes.","The same role identity can be used at three levels at once—prompt text, token alignment, and parameter routing—so role conditioning becomes a single control variable rather than three disconnected mechanisms.","Turn-aligned credit means a useful role turn reinforces both its own response tokens and, in the routed variant, the LoRA capacity that produced them, directly addressing the team credit-assignment problem in language-agent settings.","The reported HotpotQA full-Wikipedia stress test shows the benefit is not universal: when retrieval comes from an unbounded lexical index, induced roles do not beat role-free or prompt-role controls, so the method's gains are tied to the bounded supporting-document setting.","Because ExRole-Shared and ExRole-Routed differ only in the routing path, the matched comparison isolates the contribution of role-conditioned sparse LoRA routing; the paired differences on the 200-example splits are small and not statistically reliable across all metrics."],"supporting_citations":[{"why":"Establishes the role-decomposition line in multi-agent reinforcement learning that ExRole extends from compact states to language-agent trajectories.","marker":"Wang et al. 2021"},{"why":"Formulates counterfactual credit assignment for shared team rewards, the setting ExRole's turn-aligned credit adapts.","marker":"Foerster et al. 2018"},{"why":"Supplies the hand-written inception-prompting role baseline that ExRole's controlled comparisons must beat.","marker":"Li et al. 2023"},{"why":"Motivates interleaved retrieval and reasoning in the multi-hop question-answering evaluation tasks.","marker":"Trivedi et al. 2023"},{"why":"Provides a single-agent search-reinforcement-learning reference baseline and backbone comparison point.","marker":"Jin et al. 2025"},{"why":"Provides a stronger search-specialized reference that ExRole compares against but does not surpass.","marker":"Song et al. 2025"},{"why":"Shows why uniform outcome credit obscures useful intermediate role contributions, motivating role-decomposed credit.","marker":"Park, Cho, and Lee 2026"},{"why":"Provides role-typed credit assignment for agentic reinforcement learning, another antecedent for ExRole's turn-aligned credit.","marker":"Xu et al. 2026"}],"fun_headline_variants":["Trajectory-learned roles beat hand-written prompts by up to 19 EM","ExRole: roles from team logs boost multi-agent QA by 15 EM","Executable roles learned from history outperform manual prompts","Role prototypes from trajectories lift QA by 11–16 points","Learn roles from logs, not labels: ExRole gains 15 EM on QA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the hand-set scoring formula that grades each turn and serves as the shared target for role induction, credit assignment, and policy optimization; if those weights reward the wrong behavior, the induced roles and measured gains are artifacts of the scoring formula.","fun_headline_variants_meta":{"raw":{"variants":["Trajectory-learned roles beat hand-written prompts by up to 19 EM","ExRole: roles from team logs boost multi-agent QA by 15 EM","Executable roles learned from history outperform manual prompts","Role prototypes from trajectories lift QA by 11–16 points","Learn roles from logs, not labels: ExRole gains 15 EM on QA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000661,"raw_usage":{"total_tokens":3045,"prompt_tokens":995,"completion_tokens":2050,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":1956}},"tokens_in":611,"tokens_out":2050,"duration_ms":15433,"temperature":1.0,"reasoning_tokens":1956,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:21:55.520037+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ExRole-Shared from scratch on the same 200-example MuSiQue split with the reward coefficients in the turn-level formula perturbed (for example, doubling $\\lambda_{\\text{ans}}$ and $\\lambda_{\\text{ver}}$, or zeroing all penalty terms) and measure EM/F1 against the single-agent and no-role controls; if the induced-role advantage shrinks to within noise, the reward's manual weighting, not trajectory-derived roles, is carrying the result.","supporting_citations":[],"review_version":1}