acceptodds
Under review as a conference paper at ICLR 2027

SLRC: Structured Low-Rank Coupling for One-Round Communication in Cooperative Multi-Agent Reinforcement Learning

Abstract

Efficient coordination under partial observability remains a central challenge in multi-agent reinforcement learning. Communication supplements local observations, while multi-round exchanges enable agents to refine their responses using updated peer information. However, these exchanges introduce sequential dependencies across rounds. Inspired by this refinement process, we introduce SLRC (**S**tructured **L**ow-**R**ank **C**oupling), a one-round communication framework for cooperative MARL. SLRC adopts a structured affine response model that separates agent-specific transformations from rank- interaction dependencies. Each agent encodes its contribution into a compact packet of dimension . After one logical exchange, receivers aggregate available packets and solve an -dimensional linear system to locally infer the outcome of iterative refinement within the learned response model. Our analysis establishes exact decoding of the model-defined coupled response and its equivalence to the limit of iterative refinement under contraction. Experiments show that SLRC achieves the highest performance among the evaluated methods across 13 Cooperative Navigation, Predator Prey, and SMACv2 scenarios and remains effective under field-of-view (FoV) constraints. Sparse-reward Hallway tasks reveal a limitation, while rank analyses characterize the trade-off between interaction capacity and message payload. SLRC offers a new perspective on cooperative MARL communication by shifting from repeated exchanges to structured interaction inference within one round.

Primary area:
reinforcement learning
Keywords:
Multi-Agent Reinforcement Learning, Cooperative Multi-Agent System, Communication

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_kWynOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates efficient communication in cooperative multi-agent reinforcement learning (MARL) under partial observability. Existing multi-round communication methods allow agents to iteratively refine their representations using information from teammates, but introduce sequential dependencies and additional communication overhead.

To address this challenge, the authors propose Structured Low-Rank Coupling (SLRC), a one-round communication framework that models inter-agent response dependencies using a structured low-rank affine formulation. Each agent encodes its local intent and interaction factors into a message of dimension $ r^2 + r $. After a single communication round, each receiver aggregates the available messages and solves an $ r $-dimensional linear system to obtain its collaborative representation update.

The paper provides theoretical results establishing exact recovery of the model-defined coupled response under specified conditions, together with its relationship to iterative refinement under a contraction assumption. Experiments across Cooperative Navigation, Predator Prey, SMACv2, and Hallway evaluate coordination performance, communication efficiency, and component contributions. SLRC achieves the highest mean performance among the evaluated methods in 13 Cooperative Navigation, Predator Prey, and SMACv2 scenarios, while exhibiting limitations in sparse-reward Hallway tasks.

Critical strengths

1.The paper proposes a well-motivated structured low-rank coupling mechanism that enables collaborative inference within a single communication round, providing an interesting alternative to conventional multi-round communication methods.

2.The paper establishes exact recovery of the model-defined coupled response under explicit assumptions and provides a theoretical connection to iterative refinement under a contraction condition.

3.SLRC achieves the highest mean performance across all 13 CN, PP, and SMACv2 scenarios among the evaluated methods. The extensive experiments, including ablations, rank analysis, and efficiency comparisons, further support its effectiveness.

Critical weaknesses

1.Although the paper includes several ablation experiments in Section 7.4 and Appendix L, these primarily examine the effects of removing communication and auxiliary loss functions. The contribution of the proposed structured coupled decoder itself is not directly isolated. A controlled comparison with matched model capacity and communication overhead would help verify whether the performance improvements arise from the proposed decoding mechanism rather than other architectural differences.

2.As shown in Table 2, SLRC performs poorly on the Hallway tasks. In particular, its success rate on Hallway 3a is only 1.5%, compared with 100.0% for QMIX_COMM and 96.1% for SMS. The authors attribute this limitation to insufficient learning signals under sparse rewards. Further analysis would be valuable to identify the underlying causes of this degradation.

3.The main experiments remove teammate-related information from local observations to increase the importance of communication. While this is a reasonable setting for evaluating communication mechanisms, the paper does not systematically evaluate SLRC under standard observation settings where teammate information remains available. Additional comparisons would help clarify whether the advantages of SLRC generalize beyond the restricted observation settings.

Questions for authors

Q1: Could the authors provide a controlled ablation comparing the proposed coupled decoder with conventional message aggregation methods, while keeping model capacity and communication overhead comparable?

Q2: Could the authors further investigate the reasons for SLRC's significant performance degradation on the Hallway tasks, particularly Hallway 3a?

Q3: Could the authors provide additional results under standard observation settings, where teammate-related information is retained, to evaluate the generalizability of SLRC's performance improvements?

2 comments
AuthorsOpenReview ↗

Decoder ablations, Hallway failures, and generalization

Thank you for your constructive feedback. Revisions are marked in blue in the updated PDF. We respond point by point below.

Q1/W1. Does the coupled solve itself improve performance?

Table A. Decoder and packet ablations at r=32r=32

VariantParams., MFloatsCN 10 returnPP 10 vs. 4 returnZerg 5 vs. 5 win rate, %
Coupled solve1.51,056−0.474 ± 0.0233.28 ± 0.2745.69 ± 1.40
MLP1.61,056−0.615 ± 0.0042.03 ± 0.5511.62 ± 7.72
bib_i-only1.532−0.570 ± 0.0582.53 ± 0.2132.44 ± 1.74

We replace only Sc=bˉSc=\bar{b} with an MLP on (S,bˉ)(S,\bar{b}), keeping the sender architecture, packet format, payload, readout, auxiliary losses, and learner fixed. The MLP increases total parameters from 1.5M to 1.6M. SLRC achieves higher mean performance on all three tasks at identical payload, supporting coupled decoding under these controls. To test the matrix packet, bib_i-only sends bib_i and sets each zi=Irz_i=I_r at the receiver. Its lower performance across all three tasks shows the cost of omitting the matrix term.

The revised manuscript summarizes these results in lines 496–503 (page 10), with Figure 22 (page 38) and Table 7 (page 39).

Q2/W2. Why does SLRC struggle on Hallway?

Table B. Hallway diagnostics

ScenarioDecoderSuccess, %Early UB, %Timeout LB, %
3aCoupled solve1.53 ± 0.6898.47 ± 0.680.00 ± 0.00
3aMLP1.30 ± 1.4073.70 ± 78.1825.00 ± 79.56
5aCoupled solve0.00 ± 0.0094.05 ± 18.945.95 ± 18.94
5aMLP0.00 ± 0.0038.10 ± 76.5361.90 ± 76.53

We perform the same decoder replacement on Hallway 3a and 5a. Success remains below 2% on 3a and at 0% on 5a, while timeout lower bounds rise with substantial variation across seeds. Replacing the coupled solve alone leaves the success gap largely unchanged. Other baselines succeed under the same sparse reward.

The timeout lower bound counts unsuccessful episodes in evaluations whose mean episode length reaches the 19-step limit. Subtracting success and this lower bound from 100% gives the early-arrival upper bound.

Two shared components may contribute. Summed packets may obscure teammate readiness, and sparse feedback gives the sender few opportunities to learn to preserve it. Action values supervising the intent encoder may also provide weak guidance on coordinated waiting. These hypotheses motivate future work on timing-aware packets and auxiliary training signals.

We summarize the findings in Section 7.2, lines 422–426 (page 8), and report Table 3 (page 31), with analysis in Appendix I.2, lines 1650–1661 (page 31).

Q3/W3. Do SLRC's gains persist with teammate observations?

Table C. Win rates with teammate observations, %

ScenarioSLRCQMIX_COMMTGCNet
Protoss 5 vs. 550.10 ± 2.9444.94 ± 3.9238.87 ± 5.86
Terran 5 vs. 562.27 ± 1.6756.48 ± 3.4343.84 ± 3.09
Zerg 5 vs. 550.93 ± 4.0042.14 ± 1.7735.83 ± 1.89

We retain teammate features for SLRC, QMIX_COMM, and TGCNet on the Protoss, Terran, and Zerg 5 vs. 5 maps. SLRC achieves the highest mean win rate on all three maps, exceeding QMIX_COMM and TGCNet by an average of 6.58 and 14.92 percentage points, respectively. These results support the generalizability of SLRC's gains to the tested settings with teammate observations.

Appendix G.5 describes the observation setting in lines 1497–1499 (page 28). The revised manuscript summarizes the results in Section 7.2, lines 414–415 (page 8), with Figure 12 (page 31) and analysis in Appendix I.1, lines 1642–1646 (page 31).

Please do not hesitate to ask if you have further questions. If our responses address your concerns, please consider raising your rating or confidence score.

Thank you for the response and additional experiments. These clarifications have addressed most of my concerns, and I have decided to maintain my original score.

Reviewer_VxjWOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SLRC, a one-round communication framework for cooperative multi-agent reinforcement learning. The method models inter-agent response dependencies through a structured low-rank affine formulation. Each agent encodes its local contribution into a compact message, and receivers aggregate the messages and solve a reduced linear system to obtain collaborative representation updates without additional communication rounds.

The paper establishes theoretical properties concerning exact decoding and the iterative-limit interpretation of the proposed response model under specific conditions. Experiments across Cooperative Navigation, Predator Prey, SMACv2, and Hallway compare SLRC with several single-round, multi-round, and non-communication baselines. SLRC achieves the highest mean performance in 13 CN, PP, and SMACv2 scenarios, but performs poorly on the two sparse-reward Hallway tasks.

Critical strengths
  1. A well-motivated and coherent communication framework: The paper presents an interesting approach to reducing sequential communication dependencies through structured low-rank coupled inference. The connection between compact message construction and receiver-side linear solving is mathematically coherent, and the method provides an alternative to repeated message exchanges.

  2. Strong empirical performance across diverse cooperative tasks: The experimental evaluation includes several representative communication baselines across different environments and team sizes. SLRC achieves competitive or superior performance in the majority of evaluated scenarios. The additional rank, FoV, and component analyses further support the effectiveness of the overall framework.

Critical weaknesses
  1. Insufficient isolation of the core coupled-inference mechanism: Although the paper provides several ablations involving communication and auxiliary losses, it does not directly isolate the contribution of the proposed reduced linear solver. Comparisons with parameter- and communication-budget-matched single-round aggregation alternatives, such as MLP- or attention-based decoders, would help establish whether the observed improvements specifically arise from structured coupled inference rather than the message representation or overall model capacity.

  2. Incomplete evaluation of the communication-efficiency trade-off: SLRC reduces the number of logical communication rounds, but its default configuration (r = 32) transmits 1,056 values per sender, exceeding the 384 values transmitted across three rounds by QMIX COMM. While the rank–payload analysis demonstrates favorable trade-offs at smaller ranks, the overall balance between fewer rounds, total message payload, and receiver-side computation remains less clear for high-performance configurations. A more systematic comparison under comparable total communication budgets would better substantiate the claimed efficiency benefits.

Questions for authors

Q1. Could the authors compare SLRC with parameter- and communication-budget-matched single-round aggregation models that replace the reduced linear solve with an MLP or attention-based decoder? This would help clarify the contribution of the coupled-inference mechanism itself.

Q2. Could the authors further clarify the performance–communication trade-off under comparable total message budgets? In particular, when is the reduction in communication rounds sufficient to justify the larger payload at higher ranks? A latency analysis, if available, would help substantiate this trade-off.

Q3. Proposition 3 provides an iterative-limit interpretation under a contraction condition, which is not explicitly enforced during training. Could the authors report how frequently the learned response maps satisfy this condition? Additionally, since Proposition 2 assumes common aggregates across receivers, could the authors clarify the scope of its exact-decoding guarantee under FoV-constrained communication, where receiver-specific aggregates may differ?

Update: I have read the authors' response, and would like to keep my original scores.

2 comments
AuthorsOpenReview ↗

Decoder ablations, communication efficiency, and decoding guarantees

Thank you for your constructive feedback. Revisions are marked in blue in the updated PDF. We respond point by point below.

Q1/W1. Does the coupled solve itself improve performance?

Table A. Decoder and packet ablations at r=32r=32

VariantParams., MFloatsCN 10 returnPP 10 vs. 4 returnZerg 5 vs. 5 win rate, %
Coupled solve1.51,056−0.474 ± 0.0233.28 ± 0.2745.69 ± 1.40
MLP1.61,056−0.615 ± 0.0042.03 ± 0.5511.62 ± 7.72
bib_i-only1.532−0.570 ± 0.0582.53 ± 0.2132.44 ± 1.74

We replace only Sc=bˉSc=\bar{b} with an MLP on (S,bˉ)(S,\bar{b}), keeping the sender architecture, packet format, payload, readout, auxiliary losses, and learner fixed. The MLP increases total parameters from 1.5M to 1.6M. SLRC achieves higher mean performance on all three tasks at identical payload, supporting coupled decoding under these controls. To test the matrix packet, bib_i-only sends bib_i and sets each zi=Irz_i=I_r at the receiver. Its lower performance across all three tasks shows the cost of omitting the matrix term.

The revised manuscript summarizes these results in lines 496–503 (page 10), with Figure 22 (page 38) and Table 7 (page 39).

Q2/W2. When do fewer rounds justify a larger packet?

Table B. Performance and modeled latency

MethodFloats/round × roundsCN 10 returnPP 10 vs. 4 returnProtoss 10 vs. 10 win rate, %Modeled P50/P95, ms
Matched payload, 72 floats/sender/step
SLRC r=8r=872×1−0.53 ± 0.022.97 ± 0.2450.84 ± 8.353.10 / 4.27
QMIX_COMM24×3−0.61 ± 0.021.87 ± 0.0535.45 ± 5.948.93 / 10.84
TGCNet24×3−0.67 ± 0.021.16 ± 0.0530.77 ± 2.188.93 / 10.84
MASIA72×1−1.10 ± 0.180.11 ± 0.1021.39 ± 6.573.10 / 4.27
Original configurations
SLRC r=32r=321,056×1−0.47 ± 0.023.29 ± 0.2761.9 ± 4.16.96 / 8.13
QMIX_COMM128×3−0.50 ± 0.032.06 ± 0.0245.7 ± 3.710.15 / 12.06

At 72 floats per sender per step, one-round SLRC achieves the highest mean performance on all three tasks, with modeled median latency of 3.10 ms versus 8.93 ms for both three-round baselines. At the original payloads, SLRC achieves 61.9% Protoss win rate versus 45.7% for QMIX_COMM, with modeled medians of 6.96 and 10.15 ms.

Latency uses simulated transmission and assumed byte-based processing costs. Under this model, two fewer rounds save more time than transmitting and processing the larger packet adds. The benefit therefore depends on communication conditions and processing costs.

The revised manuscript reports the comparison in Section 7.3, lines 471–478 (page 9), Figure 17 (page 35), and Table 6 (page 35). Appendix K.2 specifies the latency model in lines 1878–1887 (page 35).

Q3. Does the contraction condition hold, and what does FoV change?

Proposition 3, lines 250–254 (page 5), gives the sufficient condition ∥W−1UV⊤∥2<1\lVert W^{-1}UV^{\top}\rVert_2<1 for convergence of SLRC's affine iteration. Training leaves this condition unenforced. In full-team test rollouts on CN 10, PP 10 vs. 4, and Protoss 10 vs. 10, 0.0% of decisions met it at the reported precision in each scenario. The iterative-limit interpretation remains conditional, and failure of this sufficient bound alone does not establish divergence.

SLRC directly solves for the model's self-consistent response. Proposition 2, lines 236–238 (page 5), guarantees exact decoding with common full-team aggregates and invertible SS. Solver failures were 0/61,440 on CN 10, 0/61,440 on PP 10 vs. 4, and 0/75,937 on Protoss 10 vs. 10, totaling 198,817 solves without error.

Under FoV, receivers solve their own neighborhood systems. Differing aggregates can produce adjustments that depart from the full-team fixed point. Section 6.2, lines 316–323 (page 6), defines these local updates, and Figure 15 (page 34) reports performance under FoV.

Please do not hesitate to ask if you have further questions. If our responses address your concerns, please consider raising your rating or confidence score.

AuthorsOpenReview ↗

Follow-up on Rebuttal Feedback

Dear Reviewer,

Hope you are doing well. We have submitted our rebuttal and would greatly appreciate it if you could share any additional feedback when you are available. Thank you for your valuable time.

Reviewer_qqm5OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a one-round communication for cooperative MARL: each agent encodes an affine response model into a packet of dimension r**2 + r. The receiver solves an r-dimensional linear system to recover their model-defined collaborative updates. Experiments evaluate coordination performance, communication rank, restricted connectivity, and component ablations.

Critical strengths
  1. The architectural idea is coherent. Communicating response coefficients enables receivers to compute mutually dependent adjustments after one exchange. The packet construction and decoder have a direct mathematical connection.

  2. The empirical coverage is broad. The evaluation includes several task families, restricted connectivity, rank sweeps, and ablations. Reporting the Hallway failures helps identify limitations.

Critical weaknesses
  1. The experiments do not isolate coupled decoding or establish the communication-efficiency trade-off. The strongest baseline, QMIX_COMM, is the authors' own extension of QMIX (Appendix B), which originally has no communication. No baseline is run at a matched payload (all use 128-d messages, Table 4; SMS and T2MAC select which to send rather than shrinking them). Efficient-communication methods NDQ (ICLR 2020), VBC (NeurIPS 2019), TMC (NeurIPS 2020), IMAC (ICML 2020) and Gated-ACML (AAAI 2020) are cited in Appendix C.1 but not evaluated; MAIC (AAAI 2022) is not cited. SMS and MAGIC underperform no-communication QMIX on all nine SMACv2 maps and T2MAC on six (Table 2). Default SLRC sends 1056 floats per agent per step versus QMIX_COMM's 384 over three rounds, so the headline configuration is not compact relative to the baselines; lower-payload SLRC (r = 4, 8) covers three scenarios (Table 4) with no baseline payload sweep, and no latency or byte-budget evaluation exists (cf. DACOM, AAAI 2023).

  2. The ablations do not isolate coupled decoding: removing communication or the auxiliary losses only shows they help. Since receivers share c = S⁻¹b̄ (with common aggregates), compare a nonlinear decoder on identical packets, or equal-payload mean, attention or sum-of-MLPs aggregation (LAUREL, Hu et al., ACML 2022) on the same learner; QMIX_COMM is not such a control (384 vs 1056 floats, no intent encoder or auxiliary losses). Fixing zᵢ = Iᵣ and transmitting only bᵢ would test whether the matrix packet is necessary; Remark 1 does not settle this, since a scaled-projector UVᵀ does not make vᵢᵀwᵢ⁻¹uᵢ constant.

  3. Compact communication (SchedNet, ICLR 2019; VBC; NDQ; TMC), one-round execution (MAIC, AAAI 2022; MASIA, NeurIPS 2022), low-rank value factorization (Deep Coordination Graphs, ICML 2020), inverse-based propagation (PPNP, ICLR 2019; IGNN, NeurIPS 2020; EIGNN, NeurIPS 2021) and information-form fusion (Grime & Durrant-Whyte, Control Engineering Practice 1994; Olfati-Saber's distributed Kalman filter, IEEE CDC 2005) all have precedents; what remains is the particular packet and Woodbury decoder, a modest transfer of a classical idea. The propositions (rank factorization, Woodbury decoding, affine-contraction convergence) are correct but routine.

  4. Exact decoding needs only invertible S, but the iterative-limit interpretation is proved under the sufficient condition ‖W⁻¹UVᵀ‖₂ < 1 (Prop. 3) and concerns SLRC's affine map, not the Transformer rounds. Neither is enforced (Sec. 6.1; Remark 1 is only idealized), and no contraction or conditioning statistics are reported. The 10⁻⁴ floor permits inverse entries up to 10⁴, L_enor constrains vᵢᵀuᵢ rather than vᵢᵀwᵢ⁻¹uᵢ, handling of singular S is unspecified, and under FoV constraints the aggregates are receiver-specific (Sec. 6.2), so Proposition 2's premise fails and the decoded adjustments need not form a joint fixed point (unaddressed in Fig. 14).

  5. Why should coordination dependencies be well represented by a locally generated, low-rank affine operator? That low-rank affine coupling admits efficient decoding follows from the construction; that it captures the important coordination dependencies is the substantive hypothesis, and is never tested directly. The rank sweep changes capacity and payload, not the problem's interaction rank, yet Appendix K reads it as sufficient capacity. The sparse-reward explanation for Hallway is incomplete: QMIX_COMM and SMS reach 100% and 96.1% on hallway_3a under the same reward (Table 2), so sparse reward alone does not explain why SLRC fails where they succeed, and the authors should identify which component is responsible.

  6. Learner differences may confound the gains. Algorithm 1 gives QMIX_COMM an s_rc-conditioned mixer like SLRC's (Sec. 6.3), which the other baselines, run from their public code (Appendix H), lack, and a TD(λ) target with λ unspecified, whereas Appendix H describes QMIX_COMM's learner as standard EPyMARL QMIX and SLRC uses one-step TD (Sec. 6.4); both need clarification. No-communication SLRC reaches about 1.9 on PP_10_vs_4 (Fig. 16b; Appendix L) versus QMIX's 0.04, exceeding TGCNet's 1.40 (Table 2). Learner-matched controls are essential.

Questions for authors

See weakness section. Also, how was r = 32 chosen, and how should r be selected for a new task under a payload budget? Table 4's best point estimates differ by task (CN_10 at r = 48; PP_10_vs_4 and protoss_10_vs_10 at r = 32), within seed noise, and r fixes the r² + r packet. Would a budget-adaptive rank, shared by all agents at each step since the receiver sums the r × r blocks (Def. 2), be feasible?

6 comments
AuthorsOpenReview ↗

Part 1: Additional baselines, decoder ablations, and novelty

Thank you for your feedback. Revisions are marked in blue in the updated PDF. Inspired by multi-round gains, SLRC uses one response model to define a single exchange and a local solve for mutual adjustments, yielding a simple, effective protocol.

Reply to W1: Are the baselines competitive and budgets matched?

Performance on selected tasks

MethodCN 10 returnPP 10 vs. 4 returnProtoss 10 vs. 10 win rate, %
SLRC r=32r=32−0.47±0.043.28±0.4461.93±6.61
SLRC r=4r=4−0.57±0.022.49±0.3751.82±6.73
QMIX_COMM−0.50±0.042.06±0.0445.51±5.82
TGCNet−0.62±0.031.40±0.1138.67±2.64
MAIC−1.24±0.57−0.03±0.0618.46±6.87
TMC−2.54±3.09−0.64±0.3618.83±22.74

Performance and modeled latency by payload

MethodFloats × roundsCN 10PP 10 vs. 4Protoss 10 vs. 10, %P50/P95, ms
Matched: 72 floats/sender/step
SLRC r=8r=872×1−0.53±0.022.97±0.2450.84±8.353.10/4.27
QMIX_COMM24×3−0.61±0.021.87±0.0535.45±5.948.93/10.84
TGCNet24×3−0.67±0.021.16±0.0530.77±2.188.93/10.84
MASIA72×1−1.10±0.180.11±0.1021.39±6.573.10/4.27
Original configurations
SLRC r=32r=321056×1−0.47±0.043.28±0.4461.93±6.616.96/8.13
QMIX_COMM128×3−0.50±0.042.06±0.0445.51±5.8210.15/12.06

Poor SMS, MAGIC, and T2MAC results under our teammate-observation restriction motivated competitive QMIX_COMM with controllable communication depth. SLRC leads all nine SMACv2 maps in mean win rate, ahead of QMIX_COMM and published TGCNet. We cite MAIC and evaluate it and TMC on all 15 scenarios as one-round and reduced-frequency references. At 20 floats per sender per step, SLRC r=4r=4 beats all four references above on PP and Protoss, while trailing QMIX_COMM on CN. The W1/W2 ablation tests “a nonlinear decoder on identical packets.”

We cite MAIC on page 3, lines 121–122. We describe the added baselines on page 7, lines 367–369, and page 29, lines 1550–1555. Table 2 (page 30) reports the results, and we discuss the r=4r=4 setting on page 8, lines 401–403, and in Appendix I.1 at the end of page 30 and in lines 1639–1640 (page 31).

Following your budget suggestion, SLRC leads all three tasks at 72 floats per sender per step, with modeled median latency of 3.10 ms versus 8.93 ms for both three-round baselines. Reducing their widths from 128 to 24 lowers mean performance on all three tasks. At original payloads, modeled medians are 6.96 ms for SLRC and 10.15 ms for QMIX_COMM: fewer rounds offset the larger packet. Processing costs are assumed, not measured inference or solver times.

See lines 471–478 (page 9), Figure 17 and Table 6 (page 35). Appendix K.2 gives the model in lines 1878–1887 (page 35), and analysis in lines 1888–1889 (page 35) and 1911–1915 (page 36).

Reply to W1/W2: Does coupled decoding help at fixed payload?

DecoderFloats/sender/stepCN 10 returnPP 10 vs. 4 returnZerg 5 vs. 5 win rate, %
Coupled solve1056−0.474±0.0233.28±0.2745.69±1.40
MLP decoder1056−0.615±0.0042.03±0.5511.62±7.72
bib_i-only32−0.570±0.0582.53±0.2132.44±1.74

We replace only Sc=bˉSc=\bar{b} with an MLP on (S,bˉ)(S,\bar{b}), fixing sender architecture, packets, payload, readout, auxiliary losses, and learner. Both are trained end to end. The bib_i-only control sends bib_i and sets each zi=Irz_i=I_r at the receiver to test the matrix packet. Remark 1 establishes invertibility under its assumptions, but does not make weighted zi=vi⊤wi−1uiz_i=v_i^{\top}w_i^{-1}u_i constant.

The coupled solve leads all three tasks. On Zerg, it improves win rate by 34.07 percentage points over the MLP at the same payload. Omitting ziz_i lowers win rate by 13.25 points. These results support the coupled solve and the inclusion of the matrix term on the tested tasks. See lines 496–503 (page 10), Figure 22 (page 38), and Table 7 (page 39).

Reply to W3: What is new beyond the mathematical tools?

Repeated exchanges in CommNet (Sukhbaatar et al., 2016), TarMAC (Das et al., 2019), MAGIC (Niu et al., 2021), and TGCNet (Zhang et al., 2025), and AC2C's conditional second round (Wang et al., 2023), inspired our question: could one exchange recover the limit of repeated updates in a learned response model?

We appreciate the parallels with DCG's low-rank payoffs (Boehmer et al., 2020), EIGNN's implicit propagation (Liu et al., 2021), and information-form fusion (Grime & Durrant-Whyte, 1994). SLRC's design contribution lies in what agents exchange. Agents send locally computed response coefficients once, then solve for mutual adjustments. A low-rank affine model and the classical Woodbury identity realize this idea directly. Propositions 2 and 3 establish when decoding is exact and when the same model's repeated updates converge to that solution. This offers a complementary perspective to attention-based message routing and aggregation.

AuthorsOpenReview ↗

Part 2: Theoretical scope, Hallway failures, and implementation clarifications

Reply to W4: When do the decoding guarantees apply?

Proposition 2 gives exact decoding with common full-team aggregates and invertible SS. Proposition 3 concerns SLRC's affine iteration. Transformer comparisons are empirical. Soft penalties enforce neither invertibility nor contraction. Remark 1 is idealized. The 10−410^{-4} floor permits inverse entries near 10410^4, and LenorL_{\mathrm{enor}} constrains vi⊤uiv_i^\top u_i, not its weighted counterpart.

The sufficient bound ∥W−1UV⊤∥2<1\lVert W^{-1}UV^\top\rVert_2<1 held in 0.0% of decisions in each task at reported precision. Failing this sufficient bound does not establish divergence. SLRC directly solves the self-consistent response under Proposition 2's conditions. Full-team diagnostics are:

  • CN 10: P95 κ2(S)=16.8±9.7\kappa_2(S)=16.8\pm9.7. 0/61,440 failed solves.
  • PP 10 vs. 4: P95 κ2(S)=14.0±10.9\kappa_2(S)=14.0\pm10.9. 0/61,440 failed solves.
  • Protoss 10 vs. 10: P95 κ2(S)=28.0±10.8\kappa_2(S)=28.0\pm10.8. 0/75,937 failed solves.

P95 is computed per seed, then reported as mean ± 95% CI. All 198,817 solves completed without error. Detected singularities raise an error. This concerns direct solves, not iterative convergence. Under FoV, different neighborhood aggregates need not yield a joint fixed point. See Propositions 2–3, lines 236–254 (page 5), and Section 6.2 (page 6).

Reply to W5a: Why model coordination with low-rank responses?

Shared tactical factors, such as target priority, can couple agents' adjustments along a few directions. Local histories inform each agent's contribution, while rr bounds the coupling rank. Raising rr from 4 to 32 improves win rate by 12.77/10.11 points on Protoss, 11.04/9.12 on Terran, and 5.50/6.31 on Zerg, for 5 vs. 5/10 vs. 10 respectively. Higher rank helps all six settings, but the gain is larger for the larger team only on Zerg. See Appendix K.1, lines 1833–1835 (page 34) and 1870–1874 (page 35), and the original sweep in Figure 16 and Table 5 (page 34).

Reply to W5b: What explains the Hallway failures?

TaskDecoderSuccess, %Early UB, %Timeout LB, %
3aCoupled solve1.53±0.6898.47±0.680.00±0.00
3aMLP1.30±1.4073.70±78.1825.00±79.56
5aCoupled solve0.00±0.0094.05±18.945.95±18.94
5aMLP0.00±0.0038.10±76.5361.90±76.53

Replacing the solve with an MLP leaves success below 2% on 3a and at 0% on 5a. Timeout lower bounds rise and vary substantially across seeds. Other baselines succeed under the same sparse reward, so reward sparsity alone cannot explain SLRC’s poor performance. Since the decoder swap leaves the success gap largely unchanged, we identify two possible bottlenecks shared by both variants.

  • Packet timing cues: summation may obscure teammate readiness, while rare successful arrivals provide little feedback to learn to preserve it.
  • Intent supervision: learned action values may weakly guide coordinated waiting when synchronized arrivals are rare.

These remain hypotheses for future work on readiness-preserving packets and auxiliary signals for coordinated waiting. See lines 422–426 (page 8), Table 3 (page 31), and Appendix I.2, lines 1650–1661 (page 31).

Reply to W6: Are learner differences controlled?

QMIX_COMM shares SLRC's srcs_{\mathrm{rc}}-conditioned mixer for a competitive multi-round reference. Published baselines retain their own learners, including TGCNet's GraphCoarsening mixer. The one-step TD expression in the original manuscript was a typo. Both SLRC and QMIX_COMM used TD(λ\lambda) with λ=0.6\lambda=0.6 in the reported experiments. We corrected the target in Section 6.4, lines 344–345 (page 7), and clarified QMIX_COMM's mixer in Appendix H, lines 1527–1529 (page 29). The SLRC and TGCNet mixer descriptions appear in lines 326–335 (page 7) and 1534–1537 (page 29).

On PP, SLRC's local architecture already contributes to performance. Figure 18(b) (page 36) shows further communication gains with mixer and TD target fixed. W1/W2 isolates decoding at fixed payload and learner.

Reply to Q1: How should rank be selected under a budget?

r=32r=32 gives best or near-best means, without establishing taskwise optimality within seed noise. The added r=4r=4 configuration sends 20 floats. See lines 401–403 (page 8) and Table 2 (page 30).

For BB floats per sender per step, start with the largest rr satisfying r2+r≤Br^2+r\le B. With B=128B=128, r=10r=10 sends 110 floats. Smaller feasible ranks can be chosen by validation. See lines 467–468 (page 9).

Adaptive rank requires a shared budget and channel order before sending to retain one exchange. It is future work in Section 8, from line 539 (page 10).

Please do not hesitate to ask if you have further questions. If our responses address your concerns, please consider raising your rating.

I thank the author for the detailed response and expanded experiments. However, I still have following concerns:

  1. Comparison with direct baseline on compact message design in MARL is still missing. The added MAIC and TMC comparisons are relevant but do no belong this line of works.
  2. The baselines with communication even perform worse than no-communication. Appendix L says removing communication "reduces performance by … 42.4% … PP 10 vs 4". That gives 3.29 × 0.576 = 1.89, which is better than all baselines with communication (except QMIX-COMM). Similarly, 5 of 7 baselines perform worse than no-communication QMIX in MPE in table 2. This aligns with my existing concern on baseline selection/tuning.
  3. In the newly added decoder control study, MLP performs worse than no-communication SLRC on Zerg 5v5 (11.6% vs 23.9%). But (1) existing works in MARL have shown MLP/sum of MLP work well. (2)the study cannot lead to conclusion of superiority over competitive learned decoders.
  4. The proposed encoding and decoding scheme is similar to earlier works. For example, Katzfuss and Hammerling[1] transmit Ri=Bi⊤Vi−1BiR_i=B_i^\top V_i^{-1}B_i and γi=Bi⊤Vi−1yi\gamma_i=B_i^\top V_i^{-1}y_i, then compute ν=(I+∑iRi)−1∑iγi\nu=(I+\sum_iR_i)^{-1}\sum_i\gamma_i under an identity, zero-mean prior. They match SLRC’s packet construction in Section 6.1 (Definition 2) and the reduced solve preceding Eq. (8) in Section 6.2 when ui=vi=Biu_i=v_i=B_i, wi=Viw_i=V_i, and fi=yif_i=y_i. Although SLRC extends this construction through learned asymmetric factors and local representation updates, the novelty remains limited.

Reference: [1] Katzfuss, M., & Hammerling, D. (2017). Parallel inference for massive distributed spatial data using low-rank models. Statistics and Computing, 27(2), 363-375.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Rebuttal and Revisions

Dear Area Chair, Program Chairs, and Reviewers,

Thank you for the thoughtful reviews. Reviewer VxjW described the connection between SLRC’s message construction and receiver-side solve as “mathematically coherent.” Reviewer kWyn highlighted SLRC’s one-round design and its highest mean performance across all 13 Cooperative Navigation, Predator Prey, and SMACv2 scenarios. Reviewer qqm5 suggested comparing the coupled solve with “a nonlinear decoder on identical packets” to isolate its contribution.

  • Decoder contribution. Following Reviewer qqm5’s suggestion, we tested an MLP on identical packets with the same learner. The coupled solve led on all three tasks. On Zerg 5 vs. 5, win rate reached 45.69% with the coupled solve, 11.62% with the MLP, and 32.44% when agents sent only the vector term (Table 7).
  • Communication efficiency. We added MAIC, TMC, and rank-4 SLRC across all 15 scenarios (Table 2). At a matched budget of 72 floats per sender per step, one-round SLRC led all three tested tasks. Modeled median latency was 3.10 ms versus 8.93 ms for both three-round baselines (Table 6).
  • Generalization and limitations. With teammate observations retained, SLRC led on all three tested SMACv2 maps, averaging 6.58 points above QMIX_COMM and 14.92 above TGCNet (Figure 12). On Hallway, summing messages may obscure which teammates are ready, and sparse rewards may make it difficult to learn when to wait for others.
  • Decoding guarantees. All 198,817 direct solves completed without error. Across the three tasks, 0.0% of decisions met the sufficient contraction bound at the reported precision. The iterative-limit interpretation therefore remains conditional.

The revised PDF includes the new experiments, tables, learning curves, and analyses, with revisions marked in blue.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.