acceptodds
Under review as a conference paper at ICLR 2027

InnoComm: Learning Decision-Relevant Latent Messages for Multi-Agent Language Models

Abstract

When evidence is distributed across large language model agents, a communication interface must preserve the information needed for receiver decisions under a real byte budget. We introduce InnoComm, a sender-local latent codec that maps locally contextualized evidence into 128-dimensional slots, serializes them as INT8 packets, and delivers them to a set-consistent language-model receiver. The codec is trained to improve receiver decisions through task supervision, frozen-base behavior matching, and packet-receiver co-adaptation rather than full activation reconstruction. On two annotation-derived candidate-decision tasks, the uncompressed interface exhibits stable content utility across three receiver seeds on Development, with a separate component confirmation on a held-out FEVEROUS population. In single-receiver-seed Development experiments, a fixed-reader four-slot packet compresses the full FP16 hidden-bank payload by approximately to a mean of 1,428.872 application-layer bytes per request while lowering NLL by 0.044406 relative to the full base. Under a tighter 146-byte-per-sender budget, co-adapting the shared codec and memory adapter reduces NLL by (95% CI ) relative to a matched codec-only continuation, while preserving content effects, base utility, and tested permutation consistency. A strong adapted text interface remains ahead by NLL at the same mean byte budget. Together, these results show that co-adapting the shared packet and receiver interface improves compact latent communication under an actual wire budget.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Latent Communication, Multi-Agent Language Models, Decision-Relevant Representations, Learned Compression, Byte-Constrained Inference

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_uYrqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies byte-budgeted latent communication between LLM agents that each hold part of the evidence for a multiple-choice decision. InnoComm compresses each sender's hidden states into a few 128-dimensional INT8 slots that a frozen 7B receiver reads through a memory adapter, and it can also co-train that adapter with the codec. On two decision tasks built from FEVEROUS and MuSiQue, co-adaptation lowers NLL by 0.043 relative to a matched codec-only continuation at 146 bytes per sender.

Critical strengths
  1. Rigorous accounting of communication cost. All interfaces are measured with realized application-layer bytes through an executable serialization path.
Critical weaknesses
  1. The central premise is not supported, because text dominates the risk–byte frontier. At 397 bytes/request the best latent packet loses to byte-matched text and to hard-capped head-truncated text with fewer bytes (both with an adapted text receiver). And native full text beats every latent endpoint. Given this, the paper does not show any regime in which one would choose InnoComm over sending (truncated) text.

  2. Results on real multi-agent settings are absent, so generality is unclear.

Questions for authors
  1. In what realistic deployment regime do latent packets beat byte-matched text? Can you show an experiment where InnoComm is on the Pareto frontier?

  2. How does InnoComm compare with existing latent-communication methods and context-compression methods under the same byte accounting?

  3. Does the approach transfer to open-ended multi-agent tasks, and can the codec generalize across tasks without re-training?

Reviewer_4bT2OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies multi-sender communication with a frozen LLM receiver. Each sender holds a public query and a private report, and the receiver chooses among candidate decisions. A frozen Qwen2.5-7B encoder produces contextual states for each report. A learned packetizer maps these states to 128-dimensional slots, which travel as real INT8 bytes in a deterministic content-based order. A shared decoder maps the slots to memory vectors for the receiver's adapter, and alignment trains only the codec and an adapter copy. The tasks are candidate-decision problems derived from FEVEROUS and MuSiQue annotations. At one slot, or 146 bytes per sender, shared alignment lowers pooled NLL by about 0.043 relative to matched codec-only training. Adapted head-truncated text under the same cap reaches lower NLL (0.654 vs 0.696) with fewer mean bytes.

Critical strengths

S1. The paper counts the budget in actual transmitted bytes. A sender packet holds a 14-byte header, a 4-byte scale and 128 INT8 bytes per slot. One slot therefore costs exactly 146 bytes after real serialization and deserialization. Content-based canonical ordering makes reception independent of arrival order. Null-input, natural-replacement and trained no-evidence controls test content use. Separate permutation tests check consistency.

S2. The alignment result uses a matched codec-only continuation as its comparison. This isolates the effect of receiver co-adaptation. At one slot, the paired source-bootstrap gain is about 0.043 with CI [0.020, 0.064], which excludes zero. The paper also reports evidence that limits a broader reading: the fixed-1 alignment interval crosses zero, and adapted text outperforms the latent interface at the same cap.

Critical weaknesses

W1. The manuscript needs substantial editing and reads like AI-generated text with insufficient human review. The Introduction cites no prior work, making it difficult to assess novelty against the closest learned communication and latent interface approaches. Conceptual citations appear mostly in Related Work. The figure organization makes it hard to follow how training connects to deployment and to locate the main comparisons among the diagnostics. NLL values, effects and interval endpoints are repeatedly reported to six decimals, for example 0.042625 with 95% CI [0.019817, 0.063856]. These extra digits add little given the reported uncertainty and make the dense tables harder to read.

W2. The final aligned and codec-only endpoints are selected and evaluated on the same Development sources, and the paper states this openly. The 5,000 source-cluster bootstrap resamples capture sampling variation given those endpoints, but they do not capture the selection step. The three receiver seeds and separate 512-source FEVEROUS confirmation concern the earlier uncompressed H-Local configuration. They do not test the final compressed, aligned endpoints.

Questions for authors

Q1. Suppose the already selected aligned and codec-only endpoints are frozen and evaluated once on source-disjoint data at the 146-byte cap, with source-level pairing. Does the gain over codec-only training persist with an interval that excludes zero?

Q2. Adapted head-truncated text reaches lower NLL than the final latent interface (0.654 vs 0.696) while using fewer mean bytes. What practical advantage does the latent interface offer in this setting? Could you compare the final aligned model and this text baseline on sender encoding cost, receiver latency and peak memory on the same examples and hardware?

Reviewer_V88mOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes InnoComm, a latent communication framework for multiagent language models that compresses local evidence from individual senders into compact INT8 packets. The framework uses a learned packetizer and shared decoder, and jointly optimizes the communication codec and receiver memory adapter while keeping the LLM backbone frozen. Experiments on candidate decision tasks derived from FEVEROUS and MuSiQue demonstrate substantial compression compared with full hidden state transmission and improved negative log likelihood through receiver alignment.

Critical strengths
  1. Well-motivated problem formulation. The paper addresses the relevant problem of learning compact, decision-relevant representations for communication among LLM agents under explicit bandwidth constraints.

  2. Explicit communication protocol. The proposed method implements actual INT8 serialization and deserialization, accounting for packet headers, quantization metadata, and transmission overhead. This makes the communication-cost evaluation more concrete than comparisons based solely on latent dimensionality.

  3. Controlled receiver-alignment experiments. The paper compares receiver-aligned training against a matched codec-only continuation and includes several content-specificity controls. These experiments provide some evidence that receiver adaptation improves downstream decisions while preserving information from local reports.

Critical weaknesses
  1. Limited methodological novelty. The proposed framework primarily combines established techniques, including attention-based latent compression, INT8 quantization, task-oriented supervision, knowledge distillation, and receiver-side adapter tuning. While their integration into a communication pipeline is useful, the paper does not introduce a fundamentally new communication mechanism or learning principle. The main contribution appears to be an empirical observation that jointly adapting the codec and receiver improves performance.

  2. The central packet-compression and receiver-alignment results are reported on the development set, which is also used for checkpoint and configuration selection. This raises concerns about potential model-selection bias, as the reported bootstrap confidence intervals do not account for uncertainty introduced by the selection procedure. Although the paper includes an independent FEVEROUS confirmation set, it is used only to validate the uncompressed interface rather than the proposed packet-compression and receiver-alignment methods. Furthermore, the evaluation is restricted to annotation-derived binary or ternary candidate-decision tasks, without experiments on open-ended generation, genuine multi-hop question answering, or realistic multi-agent collaboration.

  3. Despite achieving substantial compression relative to full FP16 hidden-state transmission, the proposed method fails to outperform a strong adapted text communication baseline under comparable communication budgets. Specifically, the aligned one-slot packet achieves an NLL of 0.696435 at 396.512 bytes per request, whereas the adapted hard-cap text baseline achieves a lower NLL of 0.653758 while using only 366.904 bytes per request.

  4. Insufficient analysis of the learned communication mechanism.The paper provides limited insight into why receiver alignment improves decision performance and how the learned latent messages encode decision-relevant information. Although the reported ablations demonstrate performance differences, they do not adequately isolate the contribution of joint codec-receiver optimization from the general benefits of additional receiver adaptation. In particular, an adapter-only training baseline with the codec frozen would help determine whether jointly updating both components is necessary.

Questions for authors
  1. In Table 7, "RA codec-only shared" selects Epoch 0. Furthermore, is there an ablation that freezes the pre-trained B codec and only updates the receiver memory adapter? Without this baseline, how can we distinguish the benefit of joint codec-receiver co-adaptation from the benefit of simply unfreezing the 3.67M-parameter receiver adapter?
  2. Given that the adapted hard-cap text baseline achieves a significantly lower NLL with fewer bytes without requiring sender-side LLM encoding, in what practical scenarios would the proposed latent packet interface be preferred over budget-adapted text?
  3. In Table 13 and Figure 4A, why does the receiver-aligned shared model degrade when increasing the communication budget from 1 slot to 2 slots and 4 slots ?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.