acceptodds
Under review as a conference paper at ICLR 2027

Semantic Information-bottleneck Guided Multi-Agent Communication Learning

10 views

Abstract

Multi-agent reinforcement learning relies critically on inter-agent communication for effective coordination in partially observable environments. Recent work has pursued either information-theoretic compression for efficiency or language grounding for interpretability, yet neither addresses the fundamental relationship between what information to communicate and how precisely to communicate it. To address this gap, we introduce **S**emantic **I**nformation-bottleneck **G**uided **M**ulti-**A**gent communication (SIGMA), a framework combining these two perspectives through the variance parameter of a variational communication encoder. SIGMA uses to quantify uncertainty from partial observability and to weight messages in Bayesian fusion, while a KL regularizer guides the latent space during pretraining; at execution, information flow is governed by confidence-weighted fusion. Concretely, SIGMA leverages a one-time LLM annotation of a small set of real observations to pretrain a semantically structured latent space; agents then transmit distribution parameters from this space conditioned on their local views, and the encoder is jointly optimized with the policy network, allowing to weight messages by confidence. Empirically, on standard MARL benchmarks, SIGMA achieves strong performance, remains robust under reduced communication budgets, and transfers effectively to unseen tasks.

Primary area:
reinforcement learning
Keywords:
Multi-Agent Reinforcement Learning, Information-bottleneck, Communication learning

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

The paper proposes SIGMA, a multi-agent communication framework that uses a variational encoder to produce Gaussian message parameters from agents' local observations. An offline LLM-annotation stage supplies semantic targets for variational pretraining, and an inverse-variance fusion mechanism weights incoming messages based on their predicted uncertainty. The fused means are subsequently used by a QMIX-style MARL policy. Experiments on SMAC/SMACv2, LBF, and GRF evaluate task performance, communication budgets, uncertainty calibration, and transfer to unseen scenarios.

Critical strengths
  1. The paper considers how agents should weight received information under partial observability. The empirical study spans multiple MARL benchmarks and includes communication-budget experiments, component ablations, transfer experiments, and uncertainty analyses.

  2. Variance-weighted aggregation is computationally straightforward, and explicitly communicating a mean and uncertainty parameter offers a useful design for investigating confidence-sensitive message fusion.

Critical weaknesses

1. Limited novelty and incomplete positioning against closely related probabilistic and information-bottleneck MARL methods

The claimed contributions rely heavily on established techniques: Gaussian variational encoding, KL regularization, inverse-variance weighting, offline semantic pretraining, and value-based MARL. The main combination is potentially useful, but the manuscript does not convincingly explain which part constitutes a new methodological insight beyond assembling these components.

In particular, the Related Work should discuss several closely relevant methods more precisely:

  • Bayesian Ego-graph Inference for Networked Multi-Agent Reinforcement Learning (NeurIPS 2025; BayesG) treats communication structure as a latent stochastic variable and optimizes variational edge masks under local observations. It infers which links should be active, whereas SIGMA assigns uncertainty-based weights to messages on available links. The mechanisms are distinct, but both use Bayesian/variational uncertainty to guide decentralized interaction; this distinction is important to the novelty argument.
  • Bandwidth-constrained Variational Message Encoding for Cooperative Multi-agent Reinforcement Learning (AAMAS 2026; BVME) is particularly close to SIGMA's variational content encoding and KL-based compression claims. BVME already uses Gaussian message posteriors and information-bottleneck-inspired regularization under explicit dimensional constraints. SIGMA should explain what is new relative to this setup, beyond LLM semantic targets and inverse-variance fusion.
  • Heterogeneous Information-Bottleneck Coordination Graphs for Multi-Agent Reinforcement Learning (HIBCG, arXiv:2605.17393, 2026) integrates information-bottleneck principles with heterogeneous coordination-graph structure and per-agent/message capacity control. This provides another relevant reference point for SIGMA's information-bottleneck and communication-efficiency motivation.

These are relevant comparisons for establishing the research gap, not automatically mandatory direct baselines: BayesG and HIBCG primarily address graph structure/capacity, whereas SIGMA emphasizes semantic priors and confidence-weighted aggregation. However, omitting their discussion leaves the contribution insufficiently differentiated from recent graph-based and variational MARL research.

2. The key uncertainty and Bayesian-fusion assumptions are not adequately justified

SIGMA interprets the variational encoder's learned variance as an estimate of partial-observability uncertainty and uses its inverse to weight messages. Yet a Gaussian encoder trained through reconstruction and KL regularization does not automatically provide calibrated epistemic uncertainty: its variance may reflect optimization, compression, or model misspecification rather than how reliable an agent's message is for a particular receiver.

Furthermore, the Gaussian product/precision-weighting argument assumes that different agents' messages are conditionally independent estimates of a shared latent quantity. In MARL, observations are correlated, and each encoder may represent different local information. It is unclear whether these assumptions hold sufficiently for the fusion rule to be described as optimal Bayesian aggregation rather than a useful heuristic.

There also appears to be a mathematical inconsistency in Eq. (5): if σk\sigma_k denotes standard deviation as defined in Eq. (1), the fused standard deviation should be σfuse=(∑kσk−2)−1/2\sigma_{\mathrm{fuse}}=(\sum_k\sigma_k^{-2})^{-1/2} (component-wise), or the fused variance should be σfuse2=(∑kσk−2)−1\sigma_{\mathrm{fuse}}^2=(\sum_k\sigma_k^{-2})^{-1}. The displayed expression instead appears to set σfuse=∑kσk−2\sigma_{\mathrm{fuse}}=\sum_k\sigma_k^{-2}. The authors should verify the equation, figure, and implementation and clarify whether this is a notation error or affects the reported uncertainty calibration. The reported action network receives the fused mean only, so any consequences for policy performance should be distinguished from those for calibration.

Questions for authors
  1. What concrete methodological capability does SIGMA add beyond prior Bayesian/variational graph and message-learning approaches, especially BVME's Gaussian KL-regularized messages and HIBCG's bottleneck-based communication-capacity control?
  2. What justifies interpreting independently encoded local messages as conditionally independent estimates of the same latent variable? Can the authors verify Eq. (5) and clarify whether the fusion rule is a formal Bayesian estimator or a precision-weighted heuristic?
  3. How is the learned σ\sigma distinguished from representation compression/noise, and what evidence shows it is calibrated as sender reliability under shift?
  4. Since online updates omit the LLM semantic loss and TD gradients do not update the VAE, what ensures that the claimed semantic organization persists and remains relevant to subsequent policy learning?
5 comments
AuthorsOpenReview ↗

Responses by Authors

We sincerely thank you for the insightful and constructive comments. We address each concern below.

Part 1/3: Responses to Weakness 1 and Question 1

Weakness 1 and Question 1: Novelty and Positioning vs. Probabilistic and IB-based MARL

Response: We thank you for pointing this out. We agree the components are established and SIGMA is insufficiently differentiated from BayesG, BVME, and HIBCG. Our different angle on the variational variance is to use the same σ that shapes the latent code during training as the execution-time weight for message trust, so content and trust come from one encoder rather than two mechanisms. We will add a Related Work comparison and a mechanism control isolating σ-as-fusion-weight from variational encoding.

The three methods differ in what the uncertainty acts on:

MethodMain questionCore mechanismDifference from SIGMA
BayesGWho should communicate with whom?Variational edge masks, stochastic topologyUses uncertainty to decide which links are active, not to weight messages on active links
BVMEHow to encode messages under bandwidth constraints?Gaussian variational encoding, KL regularization, sampled messageUses σ for stochastic encoding and compression; not as a receiver-side confidence weight
HIBCGHow to jointly control topology and capacity?Graph information bottleneck, group priors, capacity allocationFocuses on edge selection and capacity, not semantic uncertainty fusion
SIGMAWhose message should the receiver trust?LLM semantic pretraining, transmit (μ,σ)(\mu,\sigma), inverse-variance fusionσ directly determines each message's influence on the fused representation

The closest comparison is BVME. BVME samples zi=μi+σi⊙ϵz_i=\mu_i+\sigma_i\odot\epsilon during training and feeds the sampled representation downstream, coupling stochastic encoding with task optimization; at evaluation it uses the deterministic mean. SIGMA transmits mi=(μi,σi)m_i=(\mu_i,\sigma_i) and uses σi−2\sigma_i^{-2} as the receiver-side fusion weight. The real difference is that BVME uses σ inside stochastic encoding and regularization, whereas SIGMA uses σ directly as a fusion confidence signal. The original manuscript described σ mainly as an uncertainty proxy; we will state explicitly that its receiver-side fusion role is the contribution relative to BVME.

Mechanism control. BVME has no official implementation, and its GNN backbone, end-to-end training, and bandwidth-constrained message dimension differ from SIGMA's QMIX backbone. A head-to-head comparison would mix encoding, training, and fusion factors and cannot isolate receiver-side confidence weighting. We instead compare five variants within the SIGMA backbone on 2o_20b_vs_2r. (i) SIGMA (full): Gaussian VAE, frozen/slow updates plus LLM semantics, inverse-variance fusion. (ii) SIGMA w/o LLM: same encoder and fusion without LLM supervision. (iii) SIGMA + E2E VAE: end-to-end training with TD gradients entering the VAE and inverse-variance fusion. (iv) SIGMA + E2E VAE + simple fusion: same end-to-end encoder but simple averaging, so σ is not used in fusion. (v) BVME-style: SIGMA's Gaussian VAE but samples z=μ+σ⊙ϵz=\mu+\sigma\odot\epsilon during training, feeds the sampled message downstream under end-to-end training, and uses the deterministic mean at evaluation, matching BVME's evaluation protocol; σ is not used as a receiver-side weight.

VariantWin Rate (%)
SIGMA (full)84.5
SIGMA w/o LLM74.2
SIGMA + E2E VAE79.6
SIGMA + E2E VAE + simple fusion73.8
BVME-style72.4

The design isolates one factor at a time: (iii) vs (iv) isolates σ as a fusion weight, (i) vs (iii) isolates frozen encoder plus LLM semantics, and (v) tests sampled variational encoding without receiver-side weighting. The key comparison is SIGMA + E2E VAE (79.6) versus SIGMA + E2E VAE + simple fusion (73.8). With encoding and training fixed, using σ as a fusion weight yields a substantial gain, isolating receiver-side confidence weighting from variational encoding itself. SIGMA + E2E VAE (79.6) versus SIGMA (full) (84.5) shows that freezing the encoder and adding LLM semantics contributes further, so the gain is not due to end-to-end training alone. BVME-style (72.4) is comparable to the variant without receiver-side confidence weighting, consistent with sampled variational encoding without σ-based fusion not producing the same benefit. All variants use the same QMIX backbone, training budget, and actual bit budget. The BVME-style variant is not a full BVME reimplementation or a head-to-head baseline; it simulates BVME's key mechanism to isolate σ's fusion effect.

We will not claim a new Gaussian encoder or information-bottleneck objective, or better compression without direct comparison. We will make this positioning explicit in Related Work and the contribution statement.

AuthorsOpenReview ↗

Responses by Authors

Part 2/3: Responses to Weakness 2 and Questions 2-3

Weakness 2 and Questions 2–3: Uncertainty Interpretation, Fusion Assumptions, and Calibration

Response: We thank you for these concerns. We agree that a VAE variance is not calibrated epistemic uncertainty, shared encoding does not guarantee conditional independence, and Eq.5 needs clearer notation. We describe σ as a relative confidence proxy and fusion as a precision-weighted product-of-experts aggregation with a conditional Bayesian interpretation, verify Eq.5, and add σ permutation and constant-σ controls.

Two claims about σ. The concern splits into whether σ reflects sender knowledge of global state (A) and whether it indicates receiver-specific trust (B). SIGMA supports A only as a relative proxy: σ ranks observations by how much they constrain the hidden state, not as a calibrated estimate of Var⁡(s∣τi)\operatorname{Var}(s|\tau_i). Appendix C.3 reports r=0.89r=0.89 with global-state reconstruction error, lower error using (μ,σ,τ)(\mu,\sigma,\tau) than μ\mu only (0.23 vs. 0.41), stronger response to occlusion than additive noise, and ECE 0.11→0.04 after temperature scaling; these are empirical associations, not theoretical identities. σ is not an explicit compression-rate control or noise estimate; the occlusion-vs-noise contrast supports this. For B, a globally uncertain sender may still be most informative; SIGMA weights by per-dimension precision without modeling receiver-specific relevance, so σ is not receiver-specific reliability.

Fusion as Gaussian product of experts. A shared encoder guarantees a shared function, not that each latent dimension estimates the same quantity. Under a shared prior with conditionally independent observations, the joint posterior needs a prior correction, p(z∣o1,…,oN)∝∏ip(z∣oi)/p(z)N−1p(z|o_1,\ldots,o_N)\propto\prod_i p(z|o_i)/p(z)^{N-1}, rather than unconditional multiplication that double-counts the prior. When the prior is weak relative to message likelihoods, this correction is small, and Eq. (4) is a useful approximation; we state this condition rather than claiming an exact posterior. Inverse-variance weighting minimizes variance among linear unbiased combinations under unbiased messages, uncorrelated errors, and matching variances; we treat it as a statistically motivated approximation.

Eq.5 verification. We thank you for flagging Eq.5, and we realize the notation can be read in more than one way. The intended meaning is the fused precision, σfuse−2=∑kσk−2\sigma_{\mathrm{fuse}}^{-2}=\sum_k\sigma_k^{-2}, equivalently σfuse=(∑kσk−2)−1/2\sigma_{\mathrm{fuse}}=(\sum_k\sigma_k^{-2})^{-1/2}, which is consistent with standard inverse-variance fusion and with the implementation. To remove the ambiguity, we will rewrite Eq.5 to show precision, variance, and standard deviation on separate lines. The action network receives only μfuse\mu_{\mathrm{fuse}}, so calibration affects reported uncertainty, not policy.

Temperature scaling calibrates reported variance but does not enter fusion weights; multiplying all variances by a positive scalar leaves normalized weights unchanged. Improved ECE therefore does not itself show that sender ranking or fusion decisions improve. Table 17 reports transfer-target calibration (ECE 0.13→0.06, NLL 1.18→0.92), supporting meaningful reported uncertainty under shift, not a perfectly calibrated sender-reliability estimator under arbitrary shift.

Inference-time controls. We permute σ across senders and fix all variances to a constant, without retraining.

Inference conditionWin rate (%)Δ vs. unmodified
SIGMA (original σ)84.5—
σ permuted across senders77.8−6.7
σ fixed to constant76.3−8.2

Permuting σ across senders drops the win rate to 77.8%, showing that content-confidence matching matters. Fixing all variances to a constant drops it further to 76.3%, showing that a non-uniform confidence distribution also contributes. We also train a probe on local history only:

Probe (no σ)Pearson rrSpearman ρ\rhoAUC
Hidden global-state prediction error0.740.720.71

Because the probe never sees μi\mu_i or σi\sigma_i, this supports the proxy claim independently.

Task-relevant reliability check (2o_20b_vs_2r). We compare fusion weights with task-relevant error for enemy position. (i) Visible-Roach Overseer messages have lower error and higher weight than non-visible Baneling messages. (ii) For the same Overseer, occlusion raises error and lowers weight, so weight tracks current visibility rather than sender identity. (iii) Highly overlapping senders receive comparable weights, so inverse-variance fusion does not explicitly de-duplicate redundant messages; we treat this as a limitation. These checks support a task-relevant fusion value for σ without claiming calibrated posterior variance.

We will revise Eq.5 and Section 3.4, and state explicitly that σ is a relative confidence proxy whose benefit comes from content-confidence matching rather than calibrated absolute uncertainty.

AuthorsOpenReview ↗

Responses by Authors

Part 3/3: Response to Question 4

Question 4: What Ensures Semantic Organization Persists Online?

Response: We thank you for pointing this out. We agree that removing the online LLM semantic loss and blocking TD gradients from the VAE does not guarantee preservation of the pretrained semantic geometry. Dual-speed reduces but does not eliminate drift; we describe it as mitigation, not a guarantee. We report drift measurements and an inference-time intervention testing reliance on semantic organization.

Online, the VAE trains only with reconstruction and KL (Lsem=Lobs+βLKL\mathcal{L} _{\mathrm{sem}} = \mathcal{L} _{\mathrm{obs}} + \beta \mathcal{L} _{\mathrm{KL}}), updates every K=50K=50 episodes at a learning rate an order of magnitude smaller than the policy, and blocks TD gradients. This reduces parameter change but cannot guarantee unchanged semantic geometry; reconstruction may also pull the latent toward reconstruction-friendly directions that conflict with pretrained semantics. Appendix C.4/C.5 indirectly support this: freezing outperforms fine-tuning without slow updates (74.8 vs. 68.2 and 81.0 vs. 75.4), but these are not direct measurements.

Persistence: direct drift measurement. Using saved LLM annotations, we freeze policy snapshots at several stages and report semantic cosine between decoded and original embeddings, CKA between pretrained and online latents, and nearest-neighbor preservation for semantically similar observations. CKA and NN preservation are primary; cosine is secondary because decoder updates can confound it. We set a 0.90 "semantic-retention" threshold from the validation lower bound after pretraining. Four update schemes are compared under the same budget.

Update schemeSemantic cosine↑CKA↑NN preservation↑Win Rate (%)
Fully frozen VAE1.001.001.0076.6
Dual-speed (SIGMA)0.930.910.8984.5
Frequent VAE updates0.740.680.6579.2
Periodic semantic anchoring0.960.950.9484.1

Dual-speed stays above 0.90 while frequent updates fall well below it. Preservation and performance are non-monotonic: frozen preserves best but performs lowest, while dual-speed preserves above the threshold and performs best. Preservation alone is insufficient; the representation must also adapt. The drop under frequent updates suggests semantic organization is relevant, not incidental. Periodic anchoring recovers most similarity but does not improve performance over dual-speed; the difference is within seed noise.

Relevance: inference-time semantic intervention. Drift measurement shows whether structure is retained, not whether the policy relies on it. We identify semantic dimensions with a linear probe predicting LLM embeddings; top-kk by probe weight are semantic. We permute only semantic dimensions of μfuse\mu_{\mathrm{fuse}} across samples, keeping σfuse\sigma_{\mathrm{fuse}}, non-semantic dimensions, and policy unchanged; this breaks content-semantic correspondence while preserving marginals.

Inference conditionWin Rate (%)Δ vs. unmodified
Unmodified SIGMA84.5—
Semantic dimensions of μfuse\mu_{\mathrm{fuse}} permuted72.1−12.4
Non-semantic dimensions of μfuse\mu_{\mathrm{fuse}} permuted83.6−0.9

The large semantic drop and small non-semantic drop indicate that the policy relies on semantic organization of the fused representation rather than treating it as a generic feature vector. The non-semantic drop is within seed noise. This addresses "remains relevant to subsequent policy learning" more directly than downstream win rate alone.

Mechanism decomposition. To separate update frequency from online loss, we compare three VAE update rules at the same frequency K=50K=50.

Online lossSemantic cosine ↑CKA ↑Downstream WR (%)
KL only0.880.8482.4
Reconstruction only0.830.7882.9
Reconstruction + KL (SIGMA)0.930.9184.5

Reconstruction-only drifts most, consistent with the reconstruction objective pulling the latent away from semantic structure. KL-only also drifts, though less, because it pushes the posterior toward the prior; to keep this variant trainable, we freeze the decoder and use the same low KL weight as in SIGMA, so it does not collapse. Keeping both terms is more stable: the online gradient stays closer to the pretraining balance than either alone, which is why the combined loss drifts least. At dual-speed frequency, the loss combination is therefore the main factor preserving semantic structure; this does not show that frequency is irrelevant.

We therefore present semantic preservation as an empirically measured property of the dual-speed scheme, not as a theoretical guarantee.

We are grateful for these constructive suggestions. We hope these responses and the planned revisions address your concerns.

Reviewer_4a8BOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper focuses on addressing two multi-agent communication problems together: given only partial observations, how confident are the agents, and what should be the corresponding communication precision of their transmitted messages? The authors also incorporate LLMs to improve interpretability and capture entity-level semantics that remain meaningful and transferable across tasks. This communication method could be integrated with other value-based CTDE MARL methods. The evaluations are conducted around six research questions to support the proposed methods.

Critical strengths
  1. This paper is very well written, clear, and easy to follow. The introduction provides a high-level explanation of why the proposed idea is novel: given only partial observations, how precisely should agents communicate their local views? The authors then clearly explain how their design addresses this problem, with each component serving a specific purpose. Section 3 is also very well organized, introducing the SIGMA framework step by step. After reading it, I had a clear understanding of each component and the motivation behind its design.

  2. The overall design is good, the proposed SIGMA use a variance parameter σ\sigma to serve as a tractable proxy for agents' partial-view uncertainty and Bayesian fusion also use σ\sigma as a confidence weight to modulate the influence of the messages agents transmitting. This design link the agents' confidence and their comm precision. This design directly solved their identified challenges. The semantic encoder is also a good addition, since it helps improve transferability across similar MARL tasks, the use of LLM pre-training is a reasonable here. The framework is clearly designed and well explained, including the role of each parameter and the motivation behind each part.

  3. The training process and objectives are well explained in the main text. I think the clear writing makes the method easy to understand and provides enough information for reproducibility for researchers in this field.

  4. The experiments are designed around six research questions, which are all closely related to the advantages the authors proposed to give at the first place, and those questions are well-designed. For each question, the experiments provide good evidence to support their conclusions and appropriate, solid details about baselines, implementations are provided. Also the author gave a very good analysis about their experimental results. I am familiar with these marl-comm baselines, the reported scores and their performance make sense there.

  5. The ablation studies are well designed. The authors evaluate the contributions of different components, including LLM pretraining, probabilistic communication, bayesian fusion, and uncertainty weighting. I think these experiments are important because SIGMA combines several parts, and the ablation results help explain why each component is useful and how they contribute to the improved performance.

  6. The authors also consider the practical use of their framework. The LLM is only used for offline pretraining, so it does not introduce additional LLM inference costs during execution. The authors also evaluate communication budgets, scalability with increasing numbers of agents, and transferability across tasks. These are important considerations for marl-comm methods, and I think they make the proposed framework more useful.

Critical weaknesses
  1. The authors use the variance of a variational encoder as a proxy for uncertainty and then use it to weight messages during Bayesian fusion. I am wondering that if the σ\sigma really represent and agent's uncertainty under partial obs about the global view. The variance learned by a VAE is not necessarily equivalent to an agent's uncertainty about the environment. It can also reflect reconstruction difficulty or properties of the learned latent representation. The paper reports a correlation between variance and reconstruction error, but this does not fully establish that σ\sigma captures task-relevant uncertainty.

  2. The authors stated that SIGMA links agents' uncertainty to communication precision. But agents transmit the parameters μ\mu and σ\sigma, and uncertainty primarily changes how the received messages are weighted during fusion. This is different from dynamically changing the number of transmitted bits or the precision of message encoding. The experiments showed improved performance under fixed communication budgets, but do not necessarily demonstrate adaptive communication precision.

  3. In Table 4, SIGMA transfers a pretrained encoder and communication protocol to new tasks, while the baselines are trained from scratch. Although all methods receive the same 100 target-task episodes, SIGMA benefits from source-task pretraining. This makes it difficult to determine how much of the improvement comes specifically from semantic representations rather than pretraining in general. I think a stronger comparison would allow the baselines to transfer pretrained representations under comparable conditions.

  4. Agents in cooperative MARL often observe overlapping parts of the environment, so their messages can be correlated. Inverse-variance fusion may then overestimate confidence. (I noticed the authors actually acknowledge this limitation in Appendix C and provide calibration experiments. )

Questions for authors

Q1: Can the authors validate σ\sigma against a more direct measure of partial-observability uncertainty of each agent about the global view, rather than relying primarily on reconstruction error?

Q2: How does SIGMA adapt communication precision? One of the main motivations of this paper is to determine how precisely agents should communicate their local observations. But SIGMA appears to use σ\sigma mainly to adjust the weights of received messages during Bayesian fusion, while the transmitted messages still contain both μ\mu and σ\sigma. Could the authors clarify how this design changes the actual communication precision, rather than only changing the influence of received messages?

Q3: Could the authors provide a fairer comparison for cross-task transferability? In Table 4, SIGMA uses a pretrained semantic encoder and communication protocol, while the baselines are trained from scratch on the target tasks. So I think SIGMA could benefit from additional source-task pretraining. Could the authors compare against baselines that also use source-task pretraining?

4 comments
AuthorsOpenReview ↗

Responses by Authors

We sincerely thank you for the insightful comments and positive recognition of our design. We address each concern below.

Part 1/3: Responses to Weakness 1 and Question 1

Weakness 1 / Question 1: Does σ Capture Partial-Observability Uncertainty?

Response: We sincerely thank you for this important conceptual distinction. We agree that a VAE posterior variance is not necessarily equivalent to the agent's epistemic uncertainty about the global state. Our intended claim is a relative proxy, not a calibrated uncertainty estimate, and we realize a few places overstated this; we state it more precisely below and add a probe that never sees σ to match both our intent and the evidence.

We thank you for helping us state precisely what σ is and is not. We do not claim σ is a calibrated estimate of Var⁡(s∣τi)\operatorname{Var}(s|\tau_i); it is a tractable relative proxy that ranks observations by how much they constrain the global state and serves as a fusion confidence weight. A proxy need not be ontologically equivalent to the target, since fusion weights depend only on relative precision. Appendix C.3 supports this view: average σ correlates with global-state reconstruction error (Pearson r=0.89r=0.89); a decoder using (μ,σ,τ)(\mu,\sigma,\tau) achieves lower error than one using only μ\mu (0.23 vs. 0.41); occlusion raises σ more than additive noise; and temperature scaling improves ECE from 0.11 to 0.04. These are empirical associations, not a theoretical identity. Because the original r=0.89r=0.89 decoder consumed σ, we remove this confound below:

Direct test without σ on 2o_20b_vs_2r. Using global states available during centralized training, we train a probe fω(τi)f_\omega(\tau_i) that receives only local history, never sees μi\mu_i or σi\sigma_i, and predicts the hidden part of the global state (Roach positions that Banelings cannot observe). We compute ei=∥shid−s^hid∥2e_i=\|s_{\mathrm{hid}}-\hat{s}_{\mathrm{hid}}\|^2 and compare it with σˉi=1d∑jσi,j\bar{\sigma}_i = \frac{1}{d}\sum_j \sigma_{i,j}. Since the probe has no access to σ, any remaining association cannot be explained by the evaluator using σ. We report this single hypothesis test below.

Validation targetPearson rrSpearman ρ\rhoAUC
Hidden global-state prediction error0.740.720.71

AUC 0.71 shows that σ separates high-error from low-error samples above chance even when the predictor never sees σ. The association is weaker than the original 0.89, as expected once the evaluator's access to σ is removed, but it remains substantial and is not attributable to reconstruction difficulty alone, since the probe predicts exactly the information absent from the local view. This is precisely the evidence the proxy claim requires: σ need not equal Var⁡(S∣O)\operatorname{Var}(S|O), but it must rank observations usefully for fusion, and it does. Repeating the analysis on bane_vs_hM and 5z_vs_1ul and within Overseer and Baneling subsets gives correlations in 0.74–0.81 and AUC in 0.71–0.77, so the signal is not specific to one map or role. Under controlled occlusion σ remains predictive, whereas under additive observation noise the correlation drops to 0.45 and AUC to 0.55, consistent with σ tracking missing information rather than generic noise.

Theoretical rationale for fusion. Inverse-variance fusion has a conditional minimum-variance justification: when messages are unbiased estimates of a common latent variable, their errors are uncorrelated, and their variances reflect true error variances, the inverse-variance weighted estimate minimizes variance among linear unbiased combinations. This provides a conditional rationale for SIGMA's fusion rule without requiring that an ordinary VAE recovers the global-state posterior covariance. We will state this conditionality explicitly.

Task-relevant case study (2o_20b_vs_2r). We compute a position-specific prediction error for enemy location and check whether σ's position-related component tracks that error, rather than comparing role-averaged σ. We observe that Overseer observations with visible Roaches show lower position-specific uncertainty than Baneling observations without them, and that σ rises under occlusion and falls when the Roach is re-observed.

Wording sharpening. The proxy interpretation is the paper's intended meaning, and the reviewer's concern helps us see where the text overstates it. We will revise the Figure 3 caption from "confirming that σ reflects uncertainty" to "consistent with σ tracking partial observability, with the no-σ probe providing the direct test"; replace "adaptive communication" with "uncertainty-weighted fusion" and state that the message length is fixed at 2d2d; and clarify in Appendix C.3 that temperature scaling calibrates only the reported variance without changing fusion weights or the policy. If the AUC and Spearman ρ\rho weaken substantially after controlling for role and map, we will report σ as a fusion heuristic rather than a proxy.

AuthorsOpenReview ↗

Responses by Authors

Part 2/3: Responses to Weakness 2, Question 2, and Weakness 4

Weakness 2 and Question 2: Communication Precision vs. Uncertainty-Weighted Fusion

Response: We thank you for this observation. SIGMA's "precision" is the statistical precision of the latent Gaussian, Λi,j=1/σi,j2\Lambda_{i,j}=1/\sigma_{i,j}^2, setting each message's fusion weight wi,j=Λi,j/∑kΛk,jw_{i,j}=\Lambda_{i,j}/\sum_k\Lambda_{k,j}. It is not transmission precision, and SIGMA does not adapt bits, quantization, or message dimensions; we state this distinction.

SIGMA transmits (μi,σi)(\mu_i,\sigma_i) at fixed size, and σ modulates received-message weights in Eq. (5): lower-uncertainty messages receive larger weights, lower-confidence messages are down-weighted. Communication precision thus refers to the statistical precision with which a message participates in fusion, not its encoding precision or bit count. Section 3.2 already states that adaptive communication refers to uncertainty-weighted effective information flow rather than reduced nominal message size, and that SIGMA still transmits 2d2d floats per agent per timestep.

The fixed budget enables selective trust: since fusion weights depend only on relative precision, a confident message dominates the fused representation while an uncertain one is suppressed, even though both occupy the same nominal bandwidth. This is why Table 2 shows SIGMA degrading more slowly than deterministic baselines under restricted fixed budgets: the gain comes from weighting messages by confidence, not from reducing their size. We will present Table 2 accordingly, not as evidence of dynamic transport optimization.

We will replace "adaptive communication precision" with "uncertainty-adaptive message fusion," remove wording suggesting adaptive bitrate allocation or dynamic compression, and state in Section 3.4 that message length is fixed at 2d2d floats per agent per step and that temperature scaling affects only the reported variance without changing fusion weights or the policy. The core contribution is fixed-budget, uncertainty-adaptive fusion, not dynamic bandwidth control.

Weakness 4: Correlated Messages and Inverse-Variance Fusion

Response: We thank you for this analysis. Overlapping observations make the independence assumption approximate, so correlated messages can underestimate fused variance and cause overconfidence. Temperature scaling calibrates only the reported uncertainty, not the fused mean or policy, so we do not claim it resolves correlation-induced fusion bias. We treat independent Gaussian fusion as a lightweight approximation and state this limitation.

Appendix C.3 notes that correlated messages can underestimate fused variance; our experiments examine part of this consequence rather than eliminate it. At low observation overlap, temperature-scaled SIGMA achieves ECE = 0.04 and NLL = 0.87, versus ECE = 0.05 and NLL = 0.91 for covariance intersection (CI); task win rates are 84.5% and 84.1%. This shows comparable calibration at lower tuning cost, but does not show that temperature scaling corrects fusion weights or the fused mean under correlation.

We further add a controlled observation-overlap study that varies the fraction of shared visible entities across agents and reports ECE, NLL, and win rate for inverse-variance fusion, simple averaging, and CI under the same trained encoder.

OverlapFusion ruleWin Rate (%)ECENLL
LowInverse-variance (SIGMA)84.50.040.87
LowSimple average80.20.091.08
LowCovariance intersection84.10.050.91
HighInverse-variance (SIGMA)79.80.131.21
HighSimple average77.40.101.12
HighCovariance intersection80.90.060.94

At low overlap, inverse-variance fusion is competitive in win rate, better calibrated than simple averaging, and close to CI. At high overlap, it becomes overconfident (ECE 0.13, NLL 1.21), while CI stays well calibrated (ECE 0.06, NLL 0.94). Simple averaging is less sensitive to correlation (ECE 0.10) but consistently weaker in win rate. The 1.1-point win-rate gap between inverse-variance fusion and CI at high overlap is within seed noise, so we do not claim CI is significantly better in task performance; the clear difference is in calibration. Inverse-variance fusion is thus appropriate when observations are approximately independent, while high-overlap settings benefit from a correlation-aware rule such as CI.

We will describe inverse-variance fusion in Section 3.4 as a low-cost approximate rule conditional on approximate independence and unbiasedness, state that temperature scaling calibrates only the reported variance without repairing correlation-induced fusion bias, and present CI as a robustness-oriented alternative requiring per-pair correlation estimates and additional tuning. Explicitly modeling inter-agent covariance remains important future work, which we will add to the limitations rather than claiming that the current calibration step addresses it.

AuthorsOpenReview ↗

Responses by Authors

Part 3/3: Responses to Weakness 3 and Question 3

Weakness 3 and Question 3: Fair Comparison for Cross-Task Transfer

Response: We thank you for this fair point. Table 4 does not isolate semantic pretraining, so we add a SIGMA-internal comparison that varies only LLM semantic supervision under fixed architecture, source/target budgets, and adaptation protocol, a source-pretrained CommFormer variant with a target adapter, and an explanation of why equivalent pretraining is not well-defined for the other baselines; we also revise the transfer claim to match Table 4.

Table 4 freezes the source VAE and protocol, adds a lightweight adapter, and trains new heads on 100 target episodes; it shows representation reusability but does not isolate the semantic effect.

Key comparison internal to SIGMA. The key question is whether SIGMA's transfer gain comes from semantic structure or from pretraining in general. We answer this by holding architecture, source budget, target budget, and adaptation protocol fixed, and varying only whether LLM semantic supervision is used during source pretraining.

Variantbane_vs_hM5z_vs_1ul
SIGMA (LLM semantic)74.881.0
SIGMA (observation-only)66.571.2
SIGMA (no pretrain)58.364.7

VAE pretraining contributes a substantial gain over no pretraining (58.3 → 66.5 and 64.7 → 71.2), and LLM semantic supervision contributes a further gain beyond reconstruction pretraining (66.5 → 74.8 and 71.2 → 81.0). Because architecture and budgets are fixed and only semantic supervision varies, this comparison directly isolates the semantic contribution.

Source-pretrained baseline check with CommFormer. CommFormer is the most architecture-compatible baseline for this test: it has an official implementation, and its end-to-end training can be adapted to a pretrain-then-finetune protocol. We train CommFormer on the source task, add a linear adapter to align target observations to the source observation dimension, and fine-tune on the target task under the same 100-episode budget. We choose full fine-tuning rather than partial freezing because CommFormer does not expose a separable task-independent representation layer.

Methodbane_vs_hM5z_vs_1ul
CommFormer (scratch)26.137.5
CommFormer (source pretrain + adapter + finetune)38.448.2
SIGMA (no pretrain)58.364.7
SIGMA (observation-only)66.571.2
SIGMA (LLM semantic)74.881.0

Source-task pretraining helps CommFormer (+12.3 on bane_vs_hM and +10.7 on 5z_vs_1ul), confirming the reviewer's concern that the original Table 4 did not give baselines a comparable opportunity. However, even with source pretraining and a target adapter, CommFormer remains far below SIGMA without any pretraining (38.4 vs. 58.3 and 48.2 vs. 64.7). SIGMA's advantage is therefore not solely due to source-task pretraining; the communication mechanism itself contributes a gap that pretraining a baseline does not close. Within SIGMA, semantic supervision adds a further gain beyond pretraining, isolating the specific contribution the reviewer asked about.

Other baselines lack an equivalent pretraining path. CommFormer is the only baseline for which a pretrain-then-finetune protocol is well-defined, because it learns a continuous communication architecture end-to-end and has an official implementation. The other baselines do not expose a separable, task-independent representation that can be frozen and adapted. LMAC generates task-specific protocol code: adapting it to a new task requires regenerating the protocol, not transferring a learned representation, and adding an adapter does not change this. QMIX, TarMAC, NDQ, and MASIA similarly learn their communication or value modules jointly with the task, so source pretraining would amount to whole-model initialization plus target fine-tuning or an adapter, mixing the original method's capability with our modification. Presenting such variants as a fair source-pretrained comparison would therefore confound attribution rather than clarify it. CommFormer remains the cleanest baseline-side check, and we report it above.

Scope of the transfer claim. Appendix Table 21 shows that, under the same target-domain interaction budget, transferred SIGMA and from-scratch SIGMA achieve comparable performance. The experiment therefore supports the reusability of the pretrained representation and the retention of performance after transfer, not a reduction in target-domain interactions. We will revise Table 4, Appendix C.5, and the transfer discussion accordingly, and state that the semantic-transfer advantage is established by the LLM-versus-observation-only comparison within SIGMA, while the CommFormer variant shows that source pretraining alone does not close the gap to SIGMA's communication mechanism.

We are grateful for these constructive suggestions. We hope these responses and the planned revisions address your concerns.

Dear Reviewer 4a8B,

Thank you for your time and effort in reviewing our paper. Your comments are valuable in improving the quality of this paper. As the author–reviewer discussion period ends today, we would be grateful if you could review our rebuttal and let us know whether your concerns have been addressed. Thank you again for your help in improving our paper.

Best regards,

The Authors

Reviewer_n71LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose Semantic Information-bottleneck Guided Multi-Agent Communication (SIGMA), a framework for cooperative multi-agent reinforcement learning under partial observability. The proposed system combines LLM-guided semantic pretraining, probabilistic communication, and uncertainty-weighted fusion within centralized training with decentralized execution. A shared variational encoder represents each agent’s local observation as a Gaussian distribution, and agents communicate the distribution parameters (μi,σi)(\mu_i,\sigma_i).

The framework first collects real observations and obtains one-time LLM-generated descriptions of the agents’ local views. A β-VAE is pretrained using observation reconstruction, reconstruction of frozen Sentence-BERT description embeddings, and KL regularization. During communication, each receiving agent combines its own representation with the received messages through coordinate-wise inverse-variance weighting:

μfuse=∑k∈{i}∪C(i)σk−2μk∑k∈{i}∪C(i)σk−2,\mu_{\mathrm{fuse}} = \frac{\sum_{k\in\{i\}\cup C(i)}\sigma_k^{-2}\mu_k} {\sum_{k\in\{i\}\cup C(i)}\sigma_k^{-2}},

where C(i)C(i) denotes the communication neighbors of agent ii. The fused mean and the agent’s history representation are supplied to its local utility network, allowing lower-variance messages to exert greater influence on the communication-enhanced representation.

The implementation uses QMIX with a dual-speed update strategy. The utility networks, mixing network, and history encoders are optimized through temporal-difference learning, while the VAE receives slower updates using observation reconstruction and KL regularization. The TD gradient is excluded from the VAE updates, and the LLM is used exclusively for offline annotation. The main contribution is the use of a semantically pretrained probabilistic representation whose variance guides message fusion. For a given latent dimension, agents transmit a fixed number of values at each communication step, with adaptation occurring through message weighting.

The authors evaluate SIGMA on SMAC-Comm, SMACv2, Level-Based Foraging, and Google Research Football, reporting higher mean performance than the listed baselines in the main comparisons. Further experiments examine separately trained communication-budget configurations, message noise and dropout, alternative fusion methods, and value-decomposition backbones. Additionally, the uncertainty analysis relates the learned variance to global-state reconstruction error, while the transfer experiments reuse a frozen source VAE and communication protocol with an observation adapter and newly trained target utility and mixing networks.

Critical strengths
  • The proposed SIGMA provides a practical architecture for incorporating semantic supervision into multi-agent communication through a one-time offline annotation process. The shared variational encoder learns from observations and their language descriptions, and the resulting distribution parameters support uncertainty-weighted fusion during execution without online LLM calls. Additionally, experiments with different annotation models and integration with VDN, QMIX, and QPLEX show that the approach can operate across several configurations.

  • The authors provide complementary experiments that examine the proposed components alongside task performance. The communication-budget study accounts for both the mean and standard deviation transmitted by SIGMA, while the fusion comparisons report improvements over simple averaging and attention-weighted averaging. Additionally, the uncertainty analysis reports correlations with state-reconstruction error across three seeds, a mean-only auxiliary-decoder comparison, and responses to controlled occlusion and observation noise. Together with the evaluation across four benchmark families, these analysis show the usefulness of the architecture under the reported configurations.

Critical weaknesses
  • The experimental comparison provides insufficient detail about the hyperparameter optimization used to establish the reported performance and training-time advantages. Appendix C.4 examines several settings for SIGMA, while the baselines largely use recommended configurations, and the authors acknowledge that TarMAC’s SMACv2 performance may be affected by its settings. The manuscript does not identify the hyperparameter tuning algorithm or explain how the final configurations were selected across methods. Since these choices can materially affect the ranking, this leaves uncertainty about the contribution of the proposed communication mechanism relative to suitably tuned alternatives. My recommendation is to implement Protein from PufferLib for hyperparameter optimization across SIGMA and the principal baselines, and reassess the experimental results using the resulting configurations.

  • The authors state in Appendix C.5 that transfer achieves comparable performance with far fewer interactions, while Table 21 specifies that transfer and training from scratch use the same number of environment steps. Under this comparison, transferred SIGMA achieves win rates of 74.8% and 81.0%, compared with 76.6% and 83.2% for SIGMA trained directly. These results support reusing the frozen representation while maintaining similar performance, but reduction in target-task interactions seems to not be as significant.

Questions for authors
  • Practical training efficiency Table 14 compares training time against QMIX and LMAC. Could the authors include an IPPO baseline using an optimized PufferLib (3.0 or 4.0) implementation where the environments are compatible, and compare the time required to reach a common win rate under the same hardware and environment settings? Including SIGMA’s offline preparation in the accounting would help establish when the additional representation learning and communication provide a practical advantage over independent learning.

  • Communication requirements During decentralized execution, can each agent select an action using its local observation and history when messages from its neighbors are unavailable, or must it wait for those messages? Appendix C.2 reports robustness to 10% random message drops. How does the trained SIGMA policy perform when communication is unavailable for consecutive timesteps or messages arrive late?

5 comments
AuthorsOpenReview ↗

Responses by Authors

We sincerely thank you for the insightful comments and positive recognition of our practical architecture. We address each concern below.

Part 1/2: Responses to Questions 1-2

Question 1: Practical training efficiency

Response: We thank you for recommending PufferLib's Protein, which is valuable and has greatly helped our work. We use it for the unified HPO in Weakness 1 and optimized IPPO for the efficiency comparison below, reporting total training time with final win rate on 2o_20b_vs_2r and time-to-80% on 3s5z.

(i) Communication-critical: 2o_20b_vs_2r. IPPO plateaus below 80% because Banelings cannot observe Roaches without communication. Since no common high threshold is available, we report total training time and final win rate. IPPO reaches only 47.2%, below QMIX (52.3%), confirming independent learning's disadvantage here. SIGMA reaches 84.5% at 7.0 h total (0.8 h offline + 6.2 h online), shorter than LMAC's 8.8 h with a higher final win rate (84.5% vs 79.4%); its offline cost is offset by shorter online training and higher performance.

MethodOffline.(h)Online.(h)Total (h)WR(%)
QMIX--5.45.452.3
LMAC--8.88.879.4
IPPO--6.16.147.2
SIGMA0.86.27.084.5

(ii) Balanced: 3s5z. This is the primary time-to-80% map, balanced enough for multiple methods to reach high win rates. Under identical hardware, IPPO reaches 80% in 2.5 h, faster than SIGMA's 3.9 h total (0.8 h offline + 3.1 h online), while SIGMA attains a higher final win rate (96.5% vs 94.8%). Thus on the balanced map, SIGMA provides no practical time advantage; the extra 1.4 h yields a 1.7-point final gain, a trade-off rather than a speed benefit. SIGMA's offline cost is reported separately and amortized across seeds and downstream tasks.

MethodOffline.(h)Online to 80% (h)Total to 80% (h)Final WR (%)
QMIX--2.82.894.2
LMAC--3.63.695.1
IPPO--2.52.594.8
SIGMA0.83.13.996.5

The 0.8 h offline cost covers observation collection, LLM annotation, Sentence-BERT encoding, and VAE pretraining. Across regimes, the pattern fits the role of communication: in the communication-critical map, IPPO plateaus low and SIGMA achieves higher final performance and shorter total time than LMAC; in the balanced map, IPPO is competitive in practical time and SIGMA's advantage is limited to final performance.

Question 2: Communication requirements

Response: We thank you for pointing it out. We clarify that SIGMA never blocks on messages and always falls back to its own message plus local history; we quantify consecutive blackout, delay, and persistent neighbor failure.

Execution mechanism. During decentralized execution, SIGMA does not wait for neighbor messages. Each agent encodes its own message from local observation and history, fuses it with available neighbor messages via Eq.(5), and selects an action from the fused representation and history. Missing messages are excluded; if all are unavailable, fusion reduces to the agent's own message and the policy acts on local observation and history alone. There is no blocking or synchronization. In the delay experiments, a message delayed by dd steps is used as a stale message rather than dropped.

Experiments. The current 10% random-drop test is insufficient. We add the following evaluations on 2o_20b_vs_2r using the standard SIGMA policy without retraining:

ConditionSettingWin Rate (%)Δ vs full
Full comm.—84.5—
Consecutive blackoutT=1T=183.41.1
Consecutive blackoutT=3T=382.12.4
Consecutive blackoutT=5T=580.34.2
Consecutive blackoutT=10T=1077.66.9
Message delayd=1d=183.80.7
Message delayd=3d=381.03.5
Message delayd=5d=579.25.3
Message delayd=10d=1076.18.4
Persistent neighbor failureFixed random half80.93.6

Short delays d=1–3d=1–3 and blackouts T=1–3T=1–3 produce similar drops (0.7–3.5 points), since adjacent observations are highly correlated and stale messages remain close to the current step. As the horizon grows, delay and blackout diverge: d=5d=5 is slightly worse than T=5T=5 blackout because stale neighbor positions may misalign with the Roach’s current location while still entering fusion, whereas blackout avoids misleading input. At d=10d=10, delay drops to 76.1%, slightly below T=10T=10 blackout (77.6%), for the same reason. Both remain far above the no-communication IPPO baseline (47.2%), indicating that even severely stale messages do not cause policy failure and the policy still benefits from history and its own message. Persistent failure of a fixed random half of each agent's neighbors is more structured than random dropping, but remaining neighbors and the self-message provide partial compensation, yielding a drop comparable to T=5T=5 blackout. Overall, the policy degrades gracefully rather than depending on timely messages.

AuthorsOpenReview ↗

Responses by Authors

Part 2/2: Responses to Weakness 1-2

Weakness 1: Unified Hyperparameter Optimization for Fair Comparison

Response: We sincerely thank you for raising this fairness concern and for recommending PufferLib's Protein. We have conducted unified HPO for SIGMA, QMIX, TarMAC, and CommFormer on SMACv2 terran_5_vs_5 under identical search spaces and budgets, using Protein where the PufferLib wrapper is feasible and an equivalent Bayesian protocol (Optuna) otherwise. We report the re-evaluated ranking below.

We agree that the current comparison is asymmetric. SIGMA's hyperparameters were selected on a validation set with a sensitivity analysis, while most baselines use recommended configurations. To close this gap, we run HPO (selecting, for each method, the configuration that maximizes validation win rate within a fixed search space and trial budget) on SMACv2 terran_5_vs_5, where the reviewer's concern is most acute because TarMAC's default hyperparameters were not designed for SMACv2's randomized compositions and this scenario contains the strongest communication baselines.

Protein, PufferLib's cost-aware Bayesian optimizer over the Pareto frontier of wall-clock time and score, is our primary HPO engine because its budget-aware sampling prevents any method from receiving more tuning effort than another. Since SMACv2 is not a first-class PufferLib environment and the baselines have independent training code, we run Protein where the wrapper is feasible and Optuna with the same search space, trial budget, validation seeds, and early-stopping rule otherwise. For each method, the Post-HPO configuration is selected on validation seeds and evaluated on held-out test seeds. The results are as follows.

MethodPre-HPO Win Rate (%)Post-HPO Win Rate (%)Δ
QMIX62.162.5+0.4
TarMAC31.243.6+12.4
CommFormer64.866.9+2.1
SIGMA72.172.4+0.3

TarMAC benefits most (+12.4), confirming your suspicion that original hyperparameters were poorly suited to SMACv2; even so it remains well below the others (43.6%), so its weakness is mainly a method limitation rather than untuned defaults. SIGMA gains little (+0.3) because its hyperparameters were already validation-selected, while QMIX and CommFormer gain 0.4 and 2.1 points. After equal tuning, SIGMA remains strongest among the four (72.4 vs 66.9 and 62.5). LMAC's raw result (68.0, without HPO) lies between SIGMA and CommFormer; its LLM-guided non-standard hyperparameters require separate search-space design, so LMAC is excluded from this round. This is a controlled HPO study on the most acute scenario, not a full re-tuning across benchmarks; we limit the ranking to the four tuned methods and do not claim SIGMA would lead a fully tuned LMAC.

Weakness 2: Transfer Claim Is Not Supported by the Reported Interaction Budget

Response: We sincerely thank you for catching this inconsistency. The claim of "far fewer interactions" is not supported by Table 21, where transfer and from-scratch use the same environment steps and transfer is slightly lower; we will remove this claim and reframe the result as representation reusability under a matched budget, not as a sample-efficiency gain.

You are right that this claim is not supported by the current experiments. Table 21 reports both settings under the same target-domain interaction budget, and transferred SIGMA achieves 74.8% and 81.0% versus 76.6% and 83.2% for from-scratch on bane_vs_hM and 5z_vs_1ul. The experiment therefore does not show reduced target-task interactions, and "far fewer interactions" in Appendix C.5 is an overstatement we will remove.

What the experiment does support is narrower. With the pretrained VAE and communication protocol frozen, and with only a linear adapter plus new utility/mixing networks trained on the target map, SIGMA retains most of its from-scratch performance (within 1.8–2.2 points) under the same budget. This shows representation reusability, in that the semantic latent space and uncertainty-weighted fusion remain usable after dimension alignment, but it does not show sample-efficiency improvement, and we will state this distinction explicitly.

Target mapTransfer (same budget)From scratchGap
bane_vs_hM74.876.6−1.8
5z_vs_1ul81.083.2−2.2

The small gap is consistent with the design, since the adapter only aligns observation dimensions while from-scratch SIGMA optimizes all parameters on the target map. We will revise the abstract, Section 4, and Appendix C.5 to replace "far fewer interactions" with "comparable performance under a matched interaction budget via a frozen pretrained representation and a lightweight adapter," and reframe the transfer contribution as representation reusability and adapter efficiency rather than target-domain sample efficiency.

We are grateful for these constructive suggestions. We hope these responses and the planned revisions address your concerns.

Dear Reviewer n71L,

Thank you for your time and effort in reviewing our paper. Your comments are valuable in improving the quality of this paper. As the author–reviewer discussion period ends today, we would be grateful if you could review our rebuttal and let us know whether your concerns have been addressed. Thank you again for your help in improving our paper.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.