{"id":"62bced3e-22ec-4fd6-a8dd-c998244ab2e2","arxiv_id":"2502.09846","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CDE-GIB couples consensus-oriented multi-agent communication with graph information bottleneck compression and adaptive event triggering, reporting efficiency gains on a small formation-control benchmark.","lead":"This paper proposes CDE-GIB, a multi-agent reinforcement learning method that combines a graph information bottleneck regularizer with a variable-threshold event-triggering rule to compress inter-agent messages. The method is tested on a decentralized formation task, where it reports lower communication volume and faster reward improvement than three baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (19) is not the KL divergence claimed by Lemma 2: the log-ratio sign is inverted and the q_h=N(0,I) constants are wrong, so the printed GIB regularizer may not minimize the stated information-bottleneck bound.","rationale":"The reader's weakest_assumption correctly identifies Eq. (17) as the soft spot; I agree. The manuscript has no code, no error bars, and omits the closest prior baselines (MAGI and ETCNet), so the strongest empirical claim is under-supported on those grounds as well. However, the more fundamental issue is that the printed regularizer is not the KL divergence claimed: the sign of the log-ratio term is flipped and the constants for the q_h=N(0,I) KL are wrong. This bears directly on the central contribution rather than on experimental polish, because Table IV's ablation is the only direct evidence that the GIB module works as advertised. I nonetheless do not reject: the error may be typographical, the ablation direction is consistent with the intended mechanism, and a corrected derivation plus seed-averaged results would resolve the concern. Therefore I keep the reader's CONDITIONAL verdict, meaning no change to the verdict is needed.","tokens_in":10783,"tokens_out":5619,"duration_ms":53270,"concrete_test":"Independently re-derive Eq. (19) from Lemma 2, then re-run the formation-control experiment with (i) the printed Eq. (19), (ii) the corrected KL regularizer, and (iii) no regularizer, using at least 5 seeds per condition and reporting mean±std for the Table IV consensus error and Fig. 3 reward curves. If (i) and (ii) differ materially, the paper's derivation is not a typo and the claimed GIB advantage must be re-attributed; if they are identical, the sign/constant errors are notation-only. Also check that the means and variances in Eq. (19) are computed over the batch as Assumption 2 requires.","verdict_should_be":"UNCHANGED","load_bearing_attack":"CDE-GIB's central contribution is the GIB regularizer whose loss is Eq. (19), derived from Lemma 2 and Assumption 2. For diagonal Gaussians f=N(μ_m,Σ_m), g=N(μ_l,Σ_l), the standard KL is Σ_k[log(σ_l/σ_m)+(σ_m^2+(μ_m−μ_l)^2)/(2σ_l^2)−1/2]. Eq. (17) inverts the log ratio to log(σ_m/σ_l). Eq. (19) then uses this inverted ratio in the first KL term; for the second KL with q_h=N(0,I), it has +log σ_h and an overall −1 outside the sums, whereas the correct expression is Σ_k[−log σ_h+(σ_h^2+μ_h^2)/2−1/2]. Hence the printed objective is not the KL divergence of Lemma 2, and Theorem 1's 'upper bound' is not minimized by Eq. (19) unless the implementation silently uses a different formula. The only direct evidence for the GIB contribution is Table IV, which compares ConsMAC with/without GIB; if the implemented regularizer differs from the derived bound, this ablation does not validate the claimed mechanism. There is also a notation mismatch in Theorem 1: Eq. (12) writes DKL(p(o_g)||p(h)), while the proof and Assumption 2 use p(FψE(h)).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CDE-GIB, a multi-agent reinforcement learning framework that integrates a graph information bottleneck (GIB) regularizer with a variable-threshold event-triggering mechanism (VT-ETM) for consensus-driven communication and control. The authors derive an upper bound on the GIB objective (Theorem 1), approximate it using Gaussian assumptions (Assumption 2, Corollary 1), and train with a combined loss (Eq. (15)). Experiments on a formation-control task in the multi-agent particle environment compare CDE-GIB against ConsMAC, TarMAC, and MASIA, reporting lower communication volume and consensus error with VT-ETM and GIB ablations, and higher reward in learning curves.","tokens_in":11107,"tokens_out":6793,"duration_ms":56856,"significance":"If the derivation is corrected, the contribution is meaningful: it provides a principled approach for jointly compressing the communication graph and message features for consensus inference, and an event-triggering mechanism that uses both historical and current information with a time-decaying threshold. The paper includes component-wise ablations (Tables III–IV) and comparisons to multiple baselines (Figs. 3–4), which are useful and show consistent improvements. However, the theoretical derivation currently contains sign errors, an unproven positivity condition, and an unvalidated Gaussian assumption, so the significance cannot be fully assessed until these points are addressed.","major_comments":[{"comment":"The KL divergence formula printed in Eq. (17) is not correct for diagonal Gaussians: for f = N(μm, Σm) and g = N(μl, Σl), the log-ratio term is log(σ_l/σ_m), while Eq. (17) has log(σ_m/σ_l). This inverted sign carries into Eq. (19), where the second term against q_h = N(0, I) also shows +log σ_h and an overall −1 instead of the correct Σ_k[−log σ_h + (σ_h^2 + μ_h^2)/2 − 1/2]. Therefore the printed objective (19) does not equal the KL divergence of Lemma 2, and minimizing it does not minimize the bound in Theorem 1 as claimed. Please correct Eqs. (17) and (19), and clarify whether the implementation used the corrected formula; if so, update the displayed equations to match the experiments.","section":"III-B, Eq. (17) and Eq. (19)"},{"comment":"The inequality (b) in Eq. (13) is not justified: it relies on the condition 1 − E_{p(o_g)p(h)}[p(FψE(h))/p(o_g)] > 0, which is asserted without proof and is not generally true for continuous densities (where E[1/p(o_g)] may diverge). Without a valid bound for this expectation, the step from (a) to (b) does not follow, and the upper bound in Eq. (12) is not established. Please either prove this condition under the model assumptions or replace the bound with one that does not require it.","section":"III-B, Theorem 1 proof"},{"comment":"The closed-form GIB loss (19) relies on Assumption 2 that o_g, FψE(h), and h are Gaussian over training batches, and on the variational prior q_h = N(0, I). No evidence is provided that these Gaussian approximations hold for the learned representations in the formation-control experiments. Since the KL terms are computed under this assumption, the regularizer is only justified if the assumption is empirically supported or the method is shown to be insensitive to it. Please add validation (e.g., normality diagnostics on batch statistics) or discuss the consequences of misspecification.","section":"III-B, Assumption 2 and Corollary 1"},{"comment":"The ablation comparing consensus error with and without GIB in Table IV is intended to validate the GIB module, but it reports results under the printed loss (19). Given the sign error in Eq. (19), this ablation does not currently isolate the claimed GIB mechanism. After correcting the regularizer, please rerun the ablation and report the updated table and learning curves.","section":"IV-B, Table IV"}],"minor_comments":[{"comment":"The zero-vector label o(t)_k is not defined; please specify its dimensionality so that the MSE loss in Eq. (9) is well-formed.","section":"III-A, Eq. (9)"},{"comment":"The theorem statement in Eq. (12) writes DKL(p(o_g) || p(h)), whereas the proof and Eq. (19) use p(FψE(h)); please align this notation throughout.","section":"III-B, Theorem 1"},{"comment":"The value Vϕ(s(t)) is computed after the episode loop; it appears this should be Vϕ_old(s(t)) to match the MAPPO objective in Eq. (1).","section":"Algorithm 1, line 12"},{"comment":"The statement \"Consistent with Ref. [8]\" for Assumption 2 is not elaborated; a brief justification of why the same Gaussian approximation is appropriate here would improve readability.","section":"III-B, Assumption 2"}],"recommendation":"major_revision","confidential_remarks":"The theoretical core of the paper contains repairable but nontrivial errors. I recommend requesting a corrected derivation of Eqs. (17)–(19), a rigorous justification or replacement of the positivity condition in Theorem 1, and a rerun of the GIB ablation. If the implementation already uses the correct KL formula, the authors should state this explicitly, because the current text misrepresents the implemented objective and undermines the stated mechanism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is combining a graph information bottleneck regularizer with a variable-threshold event trigger inside a consensus-oriented MARL framework. That combination is not in MAGI or ETCNet, which handle the two ideas separately. The ablation tables do show the GIB and ETM modules each move the reported metrics in the intended direction, and the idea of replacing stale memorized messages with zeros under an event trigger is a sensible, testable design choice. The writing is clear about what the framework adds over ConsMAC.\n\nThat said, the soft spots are real and in the load-bearing parts. The stress-test note is correct: Eq. (17) has the KL log-ratio sign inverted, and this propagates into Eq. (19). The q_h = N(0,I) constants are also off. So the printed objective is not the KL divergence of Lemma 2, and unless the implementation silently uses a different formula, the GIB loss does not minimize the stated bound. That is not a cosmetic typo; it is the paper's central mechanism. The proof of Theorem 1 also leans on an unproven positivity condition after inequality (b), and there is a notation mismatch between Eq. (12) and the proof. These need to be fixed or the claims need to be softened.\n\nThe empirical side has a different but addressable weakness: no error bars or seeds, no comparison against MAGI or ETCNet (the closest prior methods), and no code. The claims about beating state-of-the-art are therefore not yet supported, even though the direction of the ablations is consistent with the authors' story.\n\nOn the positive side, the paper is honest about what it builds on, the framework is clearly specified, and the proposed integration is a legitimate research direction. This is not a paper with a fabricated result or a circular argument; it is an under-validated algorithmic paper with a fixable derivation error.\n\nWho is this for? Researchers working on communication-efficient MARL who want a concrete recipe for combining GIB-style compression with event-triggered transmission. They will get a useful baseline idea even if the current derivation needs correction.\n\nI would send this to peer review, not desk reject it, but only with the expectation that the authors must correct the KL derivation, add error bars and seeds, and benchmark against MAGI and ETCNet. As printed, I would not cite Eq. (19) as a valid bound.","headline":"A plausible engineering integration of GIB and event-triggered communication for MARL, but the printed KL derivation has a sign error and the experiments lack the statistical grounding to support the central claim.","tokens_in":11638,"tokens_out":611,"would_cite":false,"duration_ms":7933,"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":"This paper claims that a consensus-driven, event-based graph information bottleneck method lets multi-agent reinforcement learning reach consensus with fewer, more informative messages.","keywords":["multi-agent reinforcement learning","event-triggered communication","graph information bottleneck","consensus inference","communication and control co-design","formation control","partial observability"],"falsifier":"Run the same formation-control benchmark with observations drawn from a multimodal Gaussian mixture while keeping every other component fixed; if the closed-form GIB loss in Eq. 19 degrades relative to a version that estimates the KL terms with a Monte Carlo sample-based estimator, the Gaussianity assumption in Eq. 18 is the load-bearing simplification.","tokens_in":10561,"feed_emoji":"📡","tokens_out":7730,"duration_ms":64309,"temperature":0.7,"pith_summary":"This paper aims to show that in multi-agent reinforcement learning with partial observability, agents can reach a shared consensus with far less communication than standard protocols require. It proposes CDE-GIB, which adds two mechanisms to the consensus-based ConsMAC architecture: a graph information bottleneck (GIB) regularizer that jointly compresses the communication graph and the message features, and a variable-threshold event-triggering module that decides, from both history and current observations, whether a message is worth sending. The authors argue that these modules together extract concise, informative messages and suppress redundant exchanges, and they report that the combined method lowers communication volume, reduces consensus error, and raises formation-control reward relative to ConsMAC, TarMAC, and MASIA.","feed_headline":"GIB and event triggers cut multi-agent chatter, keep consensus","feed_subtitle":"A GIB regularizer plus an event-triggering gate lowers communication volume and improves task reward.","key_machinery":"The load-bearing objects are the GIB upper bound and the VT-ETM gate. The GIB bound, derived in Theorem 1, is $L_{GIB} \\leq E[D_{KL}(p(o_g)\\|p(F_{\\psi_E}(h)))] + \\eta E[D_{KL}(p(h|E_{mi})\\|q_h(h))]$, which under a Gaussian assumption on $o_g$, $F_{\\psi_E}(h)$, and $h$ becomes the closed-form loss in Eq. 19 using Lemma 2's KL formula. The VT-ETM gate computes an importance score $\\kappa(F_{\\psi_T}(o_i^{(t)}, o_i^{(\\tilde{t}_i)}), \\hat{e}^{(t-1)})$ and triggers communication only when this score exceeds the exponential threshold $G_{threshold}=c\\zeta^t$; otherwise the agent's outgoing message is replaced by a zero vector. Together they convert raw consensus updates into a sparse, compressed message stream.","core_discovery":"Integrating a graph information bottleneck regularizer and a variable-threshold event-triggering mechanism into the consensus-oriented multi-agent communication framework ConsMAC yields more concise message representations and lower communication volume while preserving or improving task performance. The GIB term jointly compresses the communication graph and information flow through a single upper bound, avoiding the inner-loop complexity of separate compression; the event-triggering term evaluates information importance from both historical and current observations against a time-decaying threshold and suppresses low-importance messages. The paper reports that CDE-GIB achieves lower consensus error, lower communication volume, and higher formation-control reward than ConsMAC, TarMAC, and MASIA.","pith_inferences":["The time-decreasing threshold $c\\zeta^t$ is a general mechanism for counteracting stale-update accumulation; the same principle could be ported to other event-triggered MARL protocols that currently use fixed thresholds.","Because the GIB regularizer only needs batch-level Gaussian statistics, it could in principle be attached to any consensus-oriented communicator that produces a latent vector, not only ConsMAC, suggesting a testable plug-in extension.","The roughly two-fold reduction in consensus error reported in Table IV hints that the GIB term is acting as a representation-learning objective that aligns the latent with the global observation distribution, a hypothesis a reader could isolate by ablating the CE loss while keeping GIB.","Replacing un-triggered messages with zero vectors, rather than with stale memorized messages, implies a design principle: in bandwidth-limited multi-agent systems, silence should signal 'no new information' rather than repeat old data."],"forward_implications":["Integration of VT-ETM reduces the volume of transmitted messages, with the reduction persisting across different maximum observation ranges (Table III).","The GIB regularizer roughly halves the consensus-estimation loss relative to the same architecture without it (Table IV).","CDE-GIB attains higher formation-control reward than ConsMAC, TarMAC, and MASIA in the multi-agent particle environment (Fig. 3).","The VT-ETM module can be added to other consensus algorithms and improves their reward as well, indicating plugin-like adaptability (Fig. 4).","Joint GIB optimization avoids the high computational cost of inner-loop separate compression of graph and features, making the method more scalable in principle."],"supporting_citations":[{"why":"TarMAC, an attention-based multi-agent communication baseline whose performance is used as a comparison and which VT-ETM is shown to improve.","marker":"[3]"},{"why":"MASIA, a supervised information-aggregation baseline used as a state-of-the-art comparison for the formation-control task.","marker":"[4]"},{"why":"ConsMAC, the consensus-oriented communication framework that CDE-GIB extends and the backbone into which the GIB and VT-ETM modules are inserted.","marker":"[5]"},{"why":"Graph information bottleneck theory, which provides the local-dependence assumption and the bound used in Theorem 1.","marker":"[9]"},{"why":"MAGI, the prior GIB-based MACRL method whose separate graph and feature compression is contrasted and improved upon.","marker":"[10]"},{"why":"ETCNet, the event-triggered communication method whose use of historical data motivates the variable-threshold design of VT-ETM.","marker":"[13]"},{"why":"MAPPO, the multi-agent PPO algorithm used as the RL backbone and the policy optimization objective in Eq. (2).","marker":"[16]"},{"why":"Source of Lemma 2's closed-form KL divergence between diagonal Gaussian distributions, used to compute the GIB bound in Eq. (19).","marker":"[23]"}],"fun_headline_variants":["CDE-GIB: Event triggers and graph bottleneck shrink comms","Less communication, same consensus: GIB plus event gating","Multi-agent control with leaner messages via event-triggered GIB","Graph information bottleneck and event triggers cut chatter","Efficient consensus: event-triggered communication with GIB"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The closed-form GIB loss (Eq. 19) assumes that, for each training batch, the global observation, the estimated global-state embedding, and the consensus latent are Gaussian with diagonal covariance, and that the variational posterior can be fixed as a standard normal.","fun_headline_variants_meta":{"raw":{"variants":["CDE-GIB: Event triggers and graph bottleneck shrink comms","Less communication, same consensus: GIB plus event gating","Multi-agent control with leaner messages via event-triggered GIB","Graph information bottleneck and event triggers cut chatter","Efficient consensus: event-triggered communication with GIB"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000251,"raw_usage":{"total_tokens":1505,"prompt_tokens":841,"completion_tokens":664,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":577}},"tokens_in":457,"tokens_out":664,"duration_ms":7114,"temperature":1.0,"reasoning_tokens":577,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T20:17:58.505785+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same formation-control benchmark with observations drawn from a multimodal Gaussian mixture while keeping every other component fixed; if the closed-form GIB loss in Eq. 19 degrades relative to a version that estimates the KL terms with a Monte Carlo sample-based estimator, the Gaussianity assumption in Eq. 18 is the load-bearing simplification.","supporting_citations":[{"cited_title":"Robust multi-agent communication with graph informa- tion bottleneck optimization,","cited_arxiv_id":null,"evidence_quote":"MAGI, the prior GIB-based MACRL method whose separate graph and feature compression is contrasted and improved upon."},{"cited_title":"Tarmac: Targeted multi-agent communication,","cited_arxiv_id":null,"evidence_quote":"TarMAC, an attention-based multi-agent communication baseline whose performance is used as a comparison and which VT-ETM is shown to improve."},{"cited_title":"Efficient multi-agent communication via self-supervised information aggregation,","cited_arxiv_id":null,"evidence_quote":"MASIA, a supervised information-aggregation baseline used as a state-of-the-art comparison for the formation-control task."},{"cited_title":"Decentralized adaptive formation via consensus- oriented multi-agent communication,","cited_arxiv_id":null,"evidence_quote":"ConsMAC, the consensus-oriented communication framework that CDE-GIB extends and the backbone into which the GIB and VT-ETM modules are inserted."},{"cited_title":"Graph information bottleneck,","cited_arxiv_id":null,"evidence_quote":"Graph information bottleneck theory, which provides the local-dependence assumption and the bound used in Theorem 1."},{"cited_title":"Event-triggered communication network with limited- bandwidth constraint for multi-agent reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"ETCNet, the event-triggered communication method whose use of historical data motivates the variable-threshold design of VT-ETM."},{"cited_title":"The surprising effectiveness of ppo in cooperative multi- agent games,","cited_arxiv_id":null,"evidence_quote":"MAPPO, the multi-agent PPO algorithm used as the RL backbone and the policy optimization objective in Eq. (2)."},{"cited_title":"Recovering from out-of-sample states via inverse dynamics in offline reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Source of Lemma 2's closed-form KL divergence between diagonal Gaussian distributions, used to compute the GIB bound in Eq. (19)."}],"review_version":1}