{"id":"bb30d06d-8328-48c1-a9fe-166d3ee49ce6","arxiv_id":"2501.02888","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DRMAC shows that after agents exchange messages, the integrated message embeddings still contain redundant and decision-irrelevant dimensions, and it reduces both with a Barlow-Twins-style loss plus a meta-learned dimensional mask.","lead":"A new method, DRMAC, improves multi-agent communication by decorrelating the message representations that receiving agents build and by masking dimensions that do not help decisions. It reports consistent wins over state-of-the-art baselines in Hallway and StarCraft II tasks, although the evidence is mostly learning curves without error bars.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The decorrelation objective may discard task-critical information because its two views differ only by noise injected into encoder weights, not by a semantic augmentation; the paper provides no probe or information-theoretic check that decision-relevant content survives.","rationale":"Reader's CONDITIONAL verdict is reasonable; my read does not change it. The weakest assumption is the same one the reader flagged: decorrelation via Eq. (1) preserves task-relevant information. I refine it by noting that the specific augmentation—Gaussian noise on the copied IIE weights, not on the input messages—makes the Barlow Twins analogy especially fragile. The paper gives no measurement of information retention (e.g., probe accuracy or mutual information between z and reward-relevant quantities), so a performance gain, even if reproduced, could come from the regularizer's side effects rather than from the proposed mechanism. The absence of error bars and code, plus a single-scenario ablation, makes it impossible to distinguish these explanations. However, the paper does implement a clear, parameter-counted method (4 hyperparameters in Appendix A.2), and the plug-and-play experiments are a reasonable attempt at generality. If the proposed probe check comes out negative, the central claim would need to be weakened; if it comes out positive, the conditional verdict should be upgraded. For now, CONDITIONAL remains appropriate.","tokens_in":17790,"tokens_out":5848,"duration_ms":131608,"concrete_test":"Run DRMAC and the DRMAC w/o RR ablation on 1o_2r_vs_4r with the same 5 seeds and hyperparameters. At the end of training, (i) record mean and standard deviation of final test win rate for both, and (ii) freeze the online IIE and train a linear probe on z to predict the coordinates and health of the known enemy units from the global state. If beta>0 improves win rate while the probe's AUROC on z is significantly lower than the beta=0 variant, then Eq. (1) has removed decision-relevant information and the stated mechanism is unsupported; if probe accuracy is maintained or improved, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that receiver-side dimensional redundancy and confounders persist and that DRMAC's decorrelation plus masking removes them while preserving decision-relevant content. The load-bearing step is Eq. (1) in Section 4.1: the invariance term forces the cross-correlation of two views toward identity. Those views are produced by feeding identical messages through the IIE and through a copy of the IIE with Gaussian noise added to its weights. Thus the objective enforces invariance to weight perturbation, not to a content-preserving input augmentation. Barlow Twins' redundancy-reduction rationale assumes the two views share semantic content; here no such guarantee exists. An encoder can satisfy Eq. (1) by becoming insensitive to its own parameter noise, and such insensitivity can also suppress input-dependent variation that the policy needs. The paper reports only learning curves (Figs. 6-8), a one-scenario ablation (Sec. 5.3), and no error bars, no numerical table, no code, and no direct measurement of how much task-relevant information survives in z. Appendix B states results are averaged over 5 seeds, but variance is not shown. Without a probe or information-theoretic check, the observed performance gains cannot be attributed to removing dimensional redundancy rather than to an incidental regularizing effect. This is the weakest point of the argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that in cooperative MARL, sender-side optimization of communication content and timing is insufficient because the receiver-side integrated message embeddings still exhibit dimensional redundancy and dimensional confounders. To address this, the authors propose DRMAC, which adds a Barlow-Twins-style redundancy-reduction regularization term on two views of integrated messages and an Information Selective Network that learns a per-agent dimensional mask via meta-learning. The method is evaluated on Hallway and SMAC tasks against TarMAC, MAIC, SMS, MASIA, and QMIX, and is also applied as a plug-and-play module on top of several baselines. The central claim is that DRMAC consistently outperforms existing methods by removing receiver-side dimensional redundancy and confounders while preserving decision-relevant information.","tokens_in":18103,"tokens_out":6225,"duration_ms":56385,"significance":"If the central claim is validated, DRMAC offers a useful complementary perspective to sender-side communication-efficiency methods and provides a simple, plug-and-play module that can be integrated with existing MARL algorithms. The motivating experiments in Figures 2 and 3 are a strength: they give a concrete demonstration that residual redundancy and harmful dimensions persist after sender-side optimization. The plug-and-play experiments in Figure 8 are also valuable because they show compatibility with multiple learning algorithms and communication baselines. However, the current empirical support is not yet sufficient: the headline comparisons are learning curves without error bars or significance tests, the component ablation is performed on a single SMAC map, and there is no direct evidence that decorrelation preserves task-relevant information. The contribution is potentially interesting, but the load-bearing claims require additional support before the paper can be accepted.","major_comments":[{"comment":"The headline claim that DRMAC 'consistently outperforms the baselines in almost all environments' is supported only by learning curves. Appendix B states that results are averaged over 5 random seeds, but no variance, error bars, numerical table, or significance test is reported anywhere in Section 5.2 or Figure 6. As a result, the reader cannot assess whether the observed advantages are statistically reliable or within run-to-run noise. Please add a quantitative summary (e.g., final mean and standard deviation per task) and, where practical, statistical comparisons for at least the main environments.","section":"Section 5.2, Appendix B"},{"comment":"The redundancy-reduction objective follows Barlow Twins, whose rationale assumes the two views share semantic content. Here the two views are produced by feeding the identical message through the online IIE and through a copy with Gaussian noise added to its weights; the invariance term therefore enforces invariance to weight perturbation, not to a content-preserving input augmentation. An encoder can satisfy Eq. (1) by becoming insensitive to its own parameter noise, which does not guarantee that decision-relevant information is retained. The paper provides no probe or information-theoretic check that task-relevant content survives in the decorrelated embedding, so the observed performance gains cannot be attributed specifically to redundancy removal as opposed to an incidental regularizing effect. Please add a measuring experiment (e.g., predicting the global state or the optimal action from the masked representation) or an ablation with a semantic augmentation.","section":"Section 4.1, Eq. (1)"},{"comment":"The ablation study evaluates the two components on a single SMAC map, 1o_2r_vs_4r. The conclusion that redundancy reduction and ISN are mutually reinforcing is therefore based on one scenario; please extend the ablations to at least one additional task (for instance a Hallway variant or another SMAC map) to support the claimed generality.","section":"Section 5.3, Fig. 7"},{"comment":"The description of the ISN meta-learning update is ambiguous: the text says 'back-propagation of gradients is excluded' when computing the trial weights, yet immediately concludes that θ_ISN 'is refined through second-order gradient optimization.' If the trial weights are detached from the computation graph, the resulting meta-gradient is first-order; if they are not detached, the statement about excluding back-propagation is misleading. Please clarify exactly which computation graph is used and which order of derivatives is taken, since this determines both the optimization behavior and the reproducibility of the method.","section":"Section 4.3, Eqs. (5)-(6)"},{"comment":"The paper does not provide direct evidence that the learned ISN mask corresponds to decision-irrelevant dimensions. No analysis of the learned mask values, no comparison with masking the same number of random dimensions, and no visualization of which dimensions are suppressed is reported. Without such evidence, the improvement of DRMAC over DRMAC w/o ISN might reflect a generic input-dependent feature weighting rather than the removal of dimensional confounders, which is one of the paper's two central mechanisms.","section":"Section 4.2, Section 5.3"}],"minor_comments":[{"comment":"The Dec-POMDP tuple includes a message set M, but the formal formulation does not describe how messages are generated or how they enter the observation/reward structure; please make this explicit.","section":"Section 3"},{"comment":"The parameter set θ includes θ_IEE, while the encoder is abbreviated IIE in Section 4.1; please unify the notation.","section":"Section 4.3"},{"comment":"Section 5.1 states 'two well-known cooperative multi-agent environments, encompassing a total of eight tasks,' while Appendix B begins with 'The four test environments,' which is inconsistent.","section":"Section 5.1, Appendix B"},{"comment":"The normalization in Eq. (2) lacks an epsilon term for numerical stability; please add one or explain how zero-variance dimensions are handled.","section":"Section 4.1, Eq. (2)"},{"comment":"The paper says to refer to Appendix A for implementation details, but Appendix A only reports DRMAC's architecture and hyperparameters, not a complete training recipe or the baseline configurations; please add a reproducibility statement or release code.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The central idea is interesting and the plug-and-play experiments are useful, but the load-bearing empirical claims are not yet backed by quantitative evidence such as error bars, significance tests, or multi-scenario ablations, and the decorrelation mechanism lacks a check that decision-relevant information is preserved. These issues are fixable within the scope of the paper, so I recommend major revision rather than rejection. The self-citation to IMMAC appears only in related work and is not load-bearing, so I do not see a novelty concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Fairly concrete idea, thin evidence. The genuinely new part is the receiver-side perspective: after sender-side gating, the integrated message embeddings still carry redundant and decision-irrelevant dimensions, and zeroing some of those dimensions at test time can improve returns. The two fixes, a Barlow-Twins-style decorrelation loss and a meta-learned dimensional mask, are clean, modular, and explicitly positioned as complements to existing sender-side methods. The writing is clear and the related work is accurate; the single self-citation sits in related work and is not load-bearing.\n\nWhat the paper does well: the motivating experiment in Fig. 3 is a real observation, and it is independent of the proposed method. The method itself is simple enough to implement on top of any MARL algorithm, and the plug-and-play results across QMIX, VDN, MAPPO, MASIA, SMS, and TarMAC give the modularity claim some support. The ablation on 1o_2r_vs_4r shows both components needed for the reported gain.\n\nThe soft spots are mostly about verification, not conception. First, every headline comparison is a learning curve. Appendix B says results are averaged over 5 seeds but shows no variance, no numerical table, no significance tests. For the abstract's claim that DRMAC 'consistently outperforms' baselines, that is insufficient. Second, the ablation is run on a single SMAC map, so we do not know whether both components help elsewhere. Third, the stress-test concern about the decorrelation objective is legitimate: the two views are the same messages through the encoder and a weight-noised copy, so the invariance term enforces robustness to weight perturbation, not to a content-preserving augmentation. The joint RL loss probably prevents total collapse, but the paper never checks whether decision-relevant information survives in the decorrelated z; a simple probe or information-theoretic estimate would settle it. Fourth, 'dimensional analysis' is a misnomer; nothing dimensional-analytic is done. Minor, but it oversells the framing.\n\nNo code is released, so reproducibility is limited to the plots. That said, the central idea is testable and the field would benefit from a careful replication with error bars and ablation across maps. I would not cite this in my own work yet, but I would send it to a serious referee: the method is plausible, the motivation is honest, and the missing evidence is exactly the kind a good review can request.","headline":"Plausible receiver-side complement to sender-side MARL communication, with an honest motivation and clean modules, but the evidence needs error bars, multi-map ablations, and a check that decorrelation preserves task-relevant information.","tokens_in":18564,"tokens_out":2909,"would_cite":false,"duration_ms":28158,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Dimensional redundancy and confounders persist in receiver-side message embeddings after sender-side gating, and DRMAC removes them to improve multi-agent communication.","keywords":["Multi-Agent Reinforcement Learning","Multi-Agent Communication","Communication Efficiency","Dimensional Analysis","Redundancy Reduction","Representation Learning","Cooperative Multi-Agent Systems","Dimensional Confounders"],"falsifier":"On a diagnostic task where the optimal policy is known to depend on a specific, identifiable dimension of the integrated message embedding, train DRMAC while varying the noise scale used to create the twin view or removing the redundancy term; if performance collapses when that dimension is decorrelated or down-weighted, the assumption that redundancy reduction preserves decision-relevant information is contradicted.","tokens_in":17564,"feed_emoji":"🤝","tokens_out":9771,"duration_ms":81913,"temperature":0.7,"pith_summary":"This paper argues that communication efficiency in multi-agent reinforcement learning is not fully solved by optimizing what and when agents send. Even after gating and content optimization, the integrated message embeddings formed at the receiving end still carry redundant dimensions and decision-irrelevant confounders that degrade decisions. The authors introduce DRMAC, which decorrelates the dimensions of these embeddings with a redundancy-reduction regularization term and learns a per-agent dimensional mask, via meta-learning, to down-weight confounding dimensions. They report consistent performance gains over existing communication methods on Hallway, Hallway-Group, and StarCraft tasks, and show that DRMAC's modules can be plugged into other MARL and communication algorithms to improve them. If true, communication efficiency can be improved at the receiving end rather than solely at the sender.","feed_headline":"Receiver-side decorrelation boosts multi-agent communication","feed_subtitle":"Even after gating, integrated messages keep redundant and misleading dimensions; DRMAC masks them and beats baselines.","key_machinery":"The central mechanism is a redundancy-reduction regularizer applied to the cross-correlation matrix of two views of the same integrated message embedding, together with an Information Selective Network (ISN) that produces a dimensional mask. The regularizer, equation (1), pushes the diagonal of the cross-correlation matrix toward 1 and the off-diagonal entries toward 0, so each dimension of the embedding captures distinct information; the twin views come from the online integration encoder and a noise-perturbed copy of it. The ISN takes the agent's own observation as input and outputs a mask with values in [0,1], applied to the embedding by element-wise multiplication, and is trained through a meta-learning step using second-order gradients of the RL loss so it can adapt to which dimensions matter for the current decision.","core_discovery":"The central discovery is that, even after optimized and gated messages are sent, dimensional redundancy and confounders persist in the integrated message embeddings at the receiving end, and these negatively affect communication quality and decision-making. DRMAC mitigates both: a redundancy-reduction regularization term drives the cross-correlation matrix of twin message embeddings toward the identity matrix, decoupling the information carried by each dimension, and an Information Selective Network learns a dimensional mask that dynamically adjusts gradient weights so the policy focuses on decision-relevant dimensions while suppressing confounders. The paper reports that DRMAC consistently outperforms state-of-the-art communication baselines across a diverse set of cooperative multi-agent tasks, and that its key modules serve as a plug-and-play complement to existing communication strategies rather than a replacement.","pith_inferences":["The motivating evidence that randomly masking dimensions sometimes improves performance suggests that a substantial fraction of embedding dimensions are actively harmful rather than merely redundant; a natural follow-up is to measure how that fraction changes with task complexity and message length.","The redundancy-reduction objective is adapted from self-supervised learning, so a targeted ablation of the Gaussian noise injection and the projection layer could reveal whether the invariance term, rather than the decorrelation, is the main driver of the gains.","The ISN is trained only by RL performance while the encoder is trained with the redundancy loss; one extension would be to let the mask influence which dimensions are decorrelated, instead of assuming all off-diagonal correlations are equally harmful.","Because the evaluation is limited to cooperative tasks with shared rewards, it remains open whether dimensional confounders play the same role in competitive or mixed-motive settings, where a receiver may want to distrust certain message dimensions."],"forward_implications":["Sender-side message filtering of content and timing alone does not resolve communication efficiency; receiver-side dimensional analysis is a necessary complementary step.","Decorrelation of message-embedding dimensions plus masking of decision-irrelevant dimensions improves cooperative performance across tasks with different numbers of agents and difficulty levels.","DRMAC's modules are plug-and-play: adding them to different MARL training algorithms (VDN, QMIX, MAPPO) and to communication methods (MASIA, SMS, TarMAC) improves those methods' performance.","Ablation shows that the redundancy-reduction term and the ISN mask are mutually reinforcing: removing either one degrades performance."],"supporting_citations":[{"why":"Supplies the redundancy-reduction objective (cross-correlation matrix to identity) that DRMAC adapts to message embeddings.","marker":"[45]"},{"why":"MASIA; provides the motivating visualization of dimensional redundancy in integrated representations and serves as a baseline and plug-in target.","marker":"[11]"},{"why":"SMS; shows that gated messages still produce redundant and confounded integrated embeddings, and serves as a baseline and plug-in target.","marker":"[41]"},{"why":"TarMAC; attention-based communication baseline and plug-in target for DRMAC's modules.","marker":"[7]"},{"why":"MAIC; sender-side message-customization baseline that DRMAC outperforms.","marker":"[44]"},{"why":"QMIX; the value-decomposition MARL algorithm that DRMAC builds on and one of the training algorithms it is integrated with.","marker":"[24]"},{"why":"Supplies the SMAC-Communication task configuration used to stress communication in the experiments.","marker":"[39]"},{"why":"Supplies the meta-learning update used to train the Information Selective Network's dimensional mask.","marker":"[18]"}],"fun_headline_variants":["Decouple and mask message dimensions in MARL","Receiver-side decorrelation boosts MARL comms","Why sender-side gating isn't enough in MARL","DRMAC: Cut redundant and misleading dimensions","Decorrelation and masking fix MARL communication"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that forcing the cross-correlation of two noise-corrupted views of the same integrated message toward the identity removes redundant dimensions without also discarding information the policy needs; if the Gaussian noise augmentation or the decorrelation destroys task-critical features, DRMAC's performance gains would disappear.","fun_headline_variants_meta":{"raw":{"variants":["Decouple and mask message dimensions in MARL","Receiver-side decorrelation boosts MARL comms","Why sender-side gating isn't enough in MARL","DRMAC: Cut redundant and misleading dimensions","Decorrelation and masking fix MARL communication"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000147,"raw_usage":{"total_tokens":1164,"prompt_tokens":903,"completion_tokens":261,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":190}},"tokens_in":519,"tokens_out":261,"duration_ms":3375,"temperature":1.0,"reasoning_tokens":190,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:00:32.063371+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a diagnostic task where the optimal policy is known to depend on a specific, identifiable dimension of the integrated message embedding, train DRMAC while varying the noise scale used to create the twin view or removing the redundancy term; if performance collapses when that dimension is decorrelated or down-weighted, the assumption that redundancy reduction preserves decision-relevant information is contradicted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the redundancy-reduction objective (cross-correlation matrix to identity) that DRMAC adapts to message embeddings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MASIA; provides the motivating visualization of dimensional redundancy in integrated representations and serves as a baseline and plug-in target."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SMS; shows that gated messages still produce redundant and confounded integrated embeddings, and serves as a baseline and plug-in target."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"TarMAC; attention-based communication baseline and plug-in target for DRMAC's modules."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MAIC; sender-side message-customization baseline that DRMAC outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the SMAC-Communication task configuration used to stress communication in the experiments."}],"review_version":1}