{"id":"fe2c7a23-6639-41b0-8fa5-f39225291cad","arxiv_id":"2607.03473","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Value-guided unlearning of low Counterfactual Message Value channels from an unrestricted MARL policy yields 80–90% bandwidth cuts with bounded return loss.","lead":"MUTE cuts multi-agent communication by 80–90% after training a full-bandwidth expert, by unlearning low-value messages scored with a counterfactual return estimator. It matters for bandwidth-limited robot teams and games where prior info-theoretic pruning often discards return-critical signals.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged smoothness/higher-order gap; that gap is real but already explicit and does not overturn the central empirical claim.","rationale":"The reader's diagnosis of the smoothness + single-message CMV limitation is accurate and already the paper's own stated caveat. Because the empirical pipeline (three-stage curriculum, dual loss, multi-environment results, ablations on anchoring/MVE/λ/β) stands independently of a tight theoretical certificate, the concern does not move the verdict away from CONDITIONAL. The concrete leave-two-out check would either confirm that higher-order effects are negligible under the reported regimes or quantify how much the safety argument softens; either outcome leaves the paper's contribution intact as a practical unlearning method with an explicit (if conditional) bound. No other load-bearing flaw—e.g., unfair baselines, non-reproducible code claims, or contradictory ablations—emerges on close reading.","tokens_in":25702,"tokens_out":592,"duration_ms":5557,"concrete_test":"On SMAC MMM2 and Corridor, recompute the unlearning phase while replacing the single-message CMV target (Eq. 3) with a 2-message leave-two-out Monte-Carlo estimate for a random 20 % of batches; if the final win-rate gap versus the original MUTE widens by more than 5 points or the communication rate rises above 15 %, higher-order interactions are material and the bound is not merely formal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest_assumption correctly isolates the softest point: Theorem 4.2 rests on Assumption A.1 (L-smoothness of Q*_tot w.r.t. the continuous concatenated message vector) and a single-message CMV that does not enumerate subsets (Limitations §6). The Taylor-remainder argument in the proof of Lemma A.2 therefore only bounds pairwise-style interaction residuals under that smoothness; if the true value surface is non-smooth or if critical multi-message synergies are systematically under-scored by the attention MVE, both the numerical bound and the safety of the dual-objective unlearning step become non-guarantees. That said, the paper already states the limitation, supplies empirical checks that ||Q*_tot-Q_tot||_∞ and MVE loss stay small (Appendix I), and shows that behavioral anchoring prevents collapse when messages are pruned (Figure 4, RQ3). The central claim—that value-guided unlearning from a full-communication expert yields 80–90 % bandwidth reduction with near-baseline return—is therefore supported by the multi-benchmark evidence and ablations even if the theoretical certificate remains conditional. No stronger internal inconsistency or experimental confound is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper introduces MUTE, a three-stage framework that first trains an unrestricted full-communication MARL expert (value-based CTDE backbone), then fits an attention-based Message Value Estimator to approximate Counterfactual Message Value (CMV = difference in Q*_tot with vs. without a given message), and finally unlearns low-CMV messages via the dual objective L_sparse (L1 on messages below an adaptive threshold) + β L_anchor (behavioral anchoring of individual Q_i to the frozen expert). A performance-difference bound (Theorem 4.2) is derived under L-smoothness of Q*_tot and bounded estimation errors. Empirically, MUTE is evaluated on Hallway, Traffic Junction, SMAC, SMAC-Communication and SMACv2, reporting 80–99 % communication reduction while matching or nearly matching full-communication baselines, with dropout-AUC robustness, component ablations, hyper-parameter sweeps and empirical checks that estimation errors remain small.","tokens_in":26116,"tokens_out":1130,"duration_ms":24318,"significance":"If the empirical results hold under broader scrutiny, MUTE offers a practically useful alternative to information-theoretic or gating-based communication sparsification: by decoupling expert coordination learning from subsequent value-guided pruning, it avoids the unstable multi-objective trade-offs that plague joint training. The CMV estimator, dual unlearning loss, and explicit return-gap bound constitute clear technical contributions relative to SMS (Shapley), NDQ/IMAC (mutual information) and IC3Net/TarMAC. Strengths that raise the paper above typical MARL communication work include the multi-environment evaluation with 95 % CIs over five seeds, the dropout-AUC table, the budget-allocation and L1-vs-L2 ablations, the empirical verification of the bound assumptions (Appendix I), and the release of anonymous code. These elements make the central claim falsifiable and reproducible.","major_comments":[{"comment":"Theorem 4.2 / Lemma A.2 and Assumption A.1: the return-gap certificate rests on L-smoothness of the joint action-value with respect to the continuous concatenated message vector together with a first-order Taylor remainder that bounds multi-message interaction residuals. The Limitations section already acknowledges that single-message CMV does not enumerate subsets. Because the bound is presented as a guarantee of controlled degradation, an additional empirical diagnostic (e.g., measured return gap under simultaneous masking of the entire M_red set versus the sum of individual CMVs on held-out trajectories) would make the theoretical claim more robust and would directly address the higher-order interaction concern.","section":"Theorem 4.2, Lemma A.2, §6"},{"comment":"Algorithm 1 and §4.3: the Message Value Estimator is trained once on the unrestricted expert and then frozen; during the unlearning phase only θ and ψ are updated. Although Appendix I reports that “MVE loss” stays small, it is not stated whether this loss is recomputed against fresh counterfactual targets under the evolving (sparsified) message distribution or merely the original Phase-1 targets. Clarifying this point, and if necessary re-estimating CMV periodically or with a small online update, is load-bearing for the claim that the messages being driven to zero remain the truly low-value ones.","section":"Algorithm 1, §4.3, Appendix I"}],"minor_comments":[{"comment":"Abstract and §5.2 claim “80 % to 90 %” reduction while the Hallway results and text repeatedly cite up to 99 %. Align the abstract wording with the strongest reported figures.","section":"Abstract, §5.2"},{"comment":"Author email for Qinru Qiu is listed as qiqiu@syr.com (missing .edu). Correct for reproducibility.","section":"Title page"},{"comment":"Appendix K title contains the typo “Compuatational”. Fix throughout.","section":"Appendix K"},{"comment":"Figure 2 caption and main text should more explicitly remind the reader that the triangles mark the 2 M-step expert performance that MUTE subsequently sparsifies; this avoids any residual confusion about training-budget fairness.","section":"Figure 2, §5.1"},{"comment":"In Definition 4.1 the notation Q*_tot(τ, m) ≔ Q*_tot(τ, a) is convenient but slightly overloads the second argument; a short clarifying sentence would help readers.","section":"Definition 4.1"}],"recommendation":"minor_revision","confidential_remarks":"The unlearning framing is genuinely novel for MARL communication and the experimental package is unusually thorough for the area. The two major points are clarifications rather than fundamental flaws; once addressed the paper is ready for acceptance. No citation or novelty concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: train a full-communication expert first, estimate each message’s counterfactual Q-gap with an attention MVE, then unlearn the low-value channels under an L1 sparsity term plus behavioral anchoring to the frozen individual Qs. That pipeline is the actual novelty relative to NDQ/IMAC, TarMAC/MAIC, IC3Net, and SMS. It is not a foundational rewrite of MARL; it is a clean systems move that works.\n\nWhat they do well is the evidence package. Ten environments (Hallway, TJ, SMAC, SMAC-Comm, SMACv2), 95 % CIs over five seeds, dropout-AUC robustness (Table 2), ablations of anchoring, MVE, λ, β, L1 vs L2, and budget allocation, plus empirical checks that ||Q*−Q||∞ and MVE loss stay small. The dual loss really does prevent the collapse you see when you just penalize communication during joint training. The performance-difference bound (Thm 4.2) is standard but correctly derived under the stated L-smoothness and error assumptions, and the limitations section already flags that CMV is single-message and does not enumerate subsets.\n\nSoft spots are real but proportionate. The free parameters (λ, β, the 2 M/0.5 M/1.5 M schedule) matter; early unlearning hurts. Higher-order message synergies are only controlled indirectly by anchoring, so the theoretical certificate is conditional, not tight. SMACv2 results are a bit less decisive because stochastic starts reduce the value of communication itself. None of that overturns the central empirical claim of 80–90 % (sometimes ~99 %) bandwidth cut with near-baseline return.\n\nThis is for people who actually ship bandwidth-constrained multi-agent systems or who care about practical communication pruning. The math and citation pattern look solid; the anonymous code link is a plus if it stays public. I would send it to referees. Worth reading and, if the code holds up, worth citing when you next need sparse cooperative communication.","headline":"Solid systems paper: post-hoc value-guided unlearning from a dense expert beats joint MI/gating training for sparse MARL communication, with real multi-benchmark evidence and a conditional bound.","tokens_in":26738,"tokens_out":541,"would_cite":true,"duration_ms":5518,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"MUTE unlearns low-return messages so multi-agent teams keep almost full performance at a fraction of the bandwidth.","keywords":["multi-agent reinforcement learning","communication efficiency","machine unlearning","counterfactual message value","value-guided sparsity","Dec-POMDP","behavioral anchoring"],"falsifier":"On a map where several messages are jointly necessary but each alone looks low-value, check whether MUTE’s estimated Counterfactual Message Values still keep at least one of them and whether the observed return gap stays inside the bound of Theorem 4.2; a large, unexplained drop would falsify the claim.","tokens_in":26596,"feed_emoji":"📡","tokens_out":631,"duration_ms":5536,"temperature":0.7,"pith_summary":"Cooperative multi-agent reinforcement learning often needs agents to talk under partial observability, but real bandwidth limits make dense messaging impractical. Most prior methods compress or gate messages by information-theoretic proxies such as mutual information; those proxies can keep statistically salient messages that do not actually raise team return. MUTE instead first trains an unrestricted expert policy, then treats communication reduction as machine unlearning: it estimates each message’s Counterfactual Message Value (the drop in joint action-value if that message is masked) and systematically unlearns only the low-value channels. A dual loss enforces sparsity while anchoring agents to the original expert’s local value functions, and a theorem bounds how much return can fall. Across Hallway, Traffic Junction, SMAC and SMACv2 maps the method typically cuts communication 80–90 percent (sometimes near 99 percent) while matching or nearly matching full-communication baselines.","feed_headline":"Agents keep 80–90% less talk and nearly full team scores","feed_subtitle":"MUTE unlearns only low-return messages after full training, with a proven bound on the performance drop.","key_machinery":"Counterfactual Message Value (CMV) — the difference in joint action-value with versus without a given message — estimated in one attention pass by the Message Value Estimator and used to select the redundant set that is unlearned under the dual sparsity-plus-anchoring objective of Theorem 4.2.","core_discovery":"Communication reduction in cooperative MARL can be cast as return-preserving unlearning: after training an unrestricted joint policy, an attention-based Message Value Estimator scores each message by its Counterfactual Message Value, low-value messages are driven toward zero by an L1 sparsity loss, and a behavioral-anchoring loss keeps individual Q-functions close to the expert, yielding a controllable performance gap and large bandwidth savings.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["MUTE unlearns low-value messages for 80-90% less bandwidth","Return-preserving unlearning cuts MARL talk 80-90%","Attention scores messages; MUTE drops low-value ones with bound","Post-training unlearning yields sparse return-matching comms","Counterfactual value guides 80-90% communication reduction"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The joint value function is assumed smooth enough in the continuous message embeddings, and single-message counterfactual scores plus anchoring are assumed to catch the messages that matter even when several messages interact.","fun_headline_variants_meta":{"raw":{"variants":["MUTE unlearns low-value messages for 80-90% less bandwidth","Return-preserving unlearning cuts MARL talk 80-90%","Attention scores messages; MUTE drops low-value ones with bound","Post-training unlearning yields sparse return-matching comms","Counterfactual value guides 80-90% communication reduction"]},"model":"grok-4.5","effort":"low","cost_usd":0.00388,"raw_usage":{"total_tokens":1226,"prompt_tokens":773,"num_sources_used":0,"completion_tokens":77,"cost_in_usd_ticks":38800000,"prompt_tokens_details":{"text_tokens":773,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":376,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":773,"tokens_out":77,"duration_ms":3330,"temperature":1.0,"reasoning_tokens":376,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T02:14:28.294781+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a map where several messages are jointly necessary but each alone looks low-value, check whether MUTE’s estimated Counterfactual Message Values still keep at least one of them and whether the observed return gap stays inside the bound of Theorem 4.2; a large, unexplained drop would falsify the claim.","supporting_citations":[],"review_version":1}