acceptodds
Withdrawn from ICLR 2027

UniCom: Unifying Explicit and Implicit Communication for LLM Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Effective collaboration among large language model (LLM) agents hinges on the fidelity of their communication protocols, which currently diverge into explicit communication in textual space and implicit communication in latent space. This dichotomy inherently imposes a fundamental tension between interpretability and expressiveness, wherein textual communication sacrifices high-dimensional reasoning nuances for human readability, whereas latent communication preserves rich semantics at the cost of rendering the underlying reasoning logic opaque and unauditable. To reconcile this tension, we present UniCom, a unified agentic communication protocol that integrates the complementary strengths of textual and latent modalities within a single coherent paradigm. At its core, UniCom features a lightweight dual-stream adapter, in which a shared encoder first maps the hidden states of heterogeneous senders into a unified communication space, and two parallel decoders then disentangle the encoded representations into discrete textual tokens and continuous latent semantics. In this way, UniCom naturally embodies the communication duality, delivering an interpretable preview for human inspection alongside an expressive payload for receiver inference. Experiments on five standard benchmarks under two heterogeneous agent scales show that UniCom consistently outperforms both text-only and latent-only communication baselines, revealing its potential to unify agent-to-human and agent-to-agent communication.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Agents, Multi-Agent Systems, Latent Communication

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_mhm7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes UniCom, a unified communication framework that combines explicit textual previews with implicit latent communication among heterogeneous LLM agents. UniCom maps sender hidden states into a shared communication space and integrates them into the receiver through prefix adaptation and dynamic parameter adaptation.

Critical strengths
  1. The heterogeneous-agent evaluation is a strength. The paper considers multiple sender/receiver capability configurations with different Qwen and LLaMA backbones, rather than evaluating only one fixed model pair.
  2. The visual aid is effective.
Critical weaknesses
  1. The writing quality is a major concern. Many sentences are excessively long and combine multiple ideas, which makes the paper unnecessarily difficult to follow, especially in the introduction and preliminary.
  2. The method introduces several additional components beyond dual-stream communication, including prefix adaptation and payload-conditioned dynamic LoRA. Table 3 shows that Prefix Adaptation itself contributes a large portion of the gain. It is therefore unclear how much of the improvement comes from the explicit-implicit unification versus stronger receiver-side adaptation.
  3. Hyperparameter sensitivity for latent length kk and number of senders nn is useful, but the balancing coefficients λexp\lambda_{\mathrm{exp}} and λrec\lambda_{\mathrm{rec}} in Eq. (10) are not analyzed. Since these directly control the tradeoff among task performance, textual anchoring, and latent reconstruction, sensitivity to these coefficients would be important.
Questions for authors
  1. UniCom is presented as a unified communication protocol, but the communication adapters are trained separately for each downstream task. This limits the claim of a general communication interface. Can the learned communication space transfer across tasks without retraining?
  2. Is the explicit stream necessary at inference time for performance, or is its main role only as a training-time regularizer and human-readable preview?
Reviewer_eLxVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors argue that text communication between LLM agents loses information, while latent communication is not interpretable. To address this, the authors propose UniCom, which maps the senders’ hidden states into a shared space and decodes them through two streams. One stream produces a text preview for humans, and the other feeds the receiver through a soft prompt plus LoRA weights generated by a hypernetwork. In detail, only the adapters are trained, with the backbones frozen. On five benchmarks with two heterogeneous agent scales, UniCom outperforms both text-only and latent-only baselines.

Critical strengths
  1. Combining text and latent communication in one protocol is a meaningful direction, and the setting with heterogeneous senders (Qwen and Llama) is practical.

  2. The results are promising, as UniCom is best in every cell of Table 1 and in all four sender-receiver configurations (Table 2), and the ablations show that each component helps.

Critical weaknesses
  1. The comparison to text based methods is not fair. The UniCom receiver sees the same text rationales as TextCom plus the latent payload, and its adapters are trained per task with the task loss, while TextCom has no training at all. So the gain may come mainly from task-specific fine-tuning of the receiver, not from the latent channel.

  2. Some claims do not match the setup. Interpretability is the main motivation, but the text preview is never evaluated, and it is trained only to reconstruct the senders’ rationales, which the receiver already sees. Also, Figure 3 reports UniCom as 2.4× faster than TextCom, but UniCom also consumes the sender rationales, which are pre-generated and cached.

Questions for authors

Does the latency in Figure 3 include generating the sender rationales that UniCom also uses, and how does TextCom perform with a static LoRA or prefix trained on the same task data?

Reviewer_G7TfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper mainly focuses on the interpretability problem of latent communication between LLM agents. To handle different senders with different representations, UniCom maps their latent messages into a shared communication space. The main idea is to decode these shared representations into readable text previews while retaining a latent path for the receiver. Textual supervision also regularizes the latent representations. To strengthen the messages' influence across layers, UniCom adds message-conditioned LoRA alongside soft prefixes at the receiver. Experiments on five benchmarks under two model scales report improvements over text-only and latent-only communication.

Critical strengths
  1. The paper focuses on a meaningful problem. Pure latent communication is difficult for humans to interpret and may suffer from semantic drift. Combining latent messages with readable textual previews is a meaningful direction for improving agent communication.

  2. The shared communication space is a useful design for heterogeneous agents. Model-specific encoders and decoders connect different hidden spaces through a common interface. This reduces the need to learn a separate mapping for every sender–receiver pair.

Critical weaknesses
  1. The interpretability claim is not sufficiently evaluated. Although UniCom generates textual previews, the paper does not directly evaluate their quality or whether they faithfully reflect the latent messages. The ablation on textual supervision only shows its effect on task performance. From this result, it is still hard to know whether the generated previews are readable and faithful. Therefore, how much the method achieves its claimed interpretability is still unclear.

  2. It is unclear how much of the gain comes from latent reasoning and how much comes from communication. UniCom combines latent reasoning with latent communication, but the ablations fail to separate their contributions. Requiring senders to support latent reasoning also limits direct use with existing LLMs.

  3. The empirical support for the claimed gains is limited. Improvements over the best baselines are small, and each benchmark uses only 164–500 evaluation questions. For example, the gain over Interlat on HumanEval in the small setting (26.2 vs. 24.4) amounts to about 3 more solved problems out of 164. Table 1 does not show how the scores vary across runs, so it is hard to tell whether these gains are consistent.

  4. The presentation could be easier to follow. Although the method is described in detail, the later discussion of dual-stream decoding and receiver adaptation could be clearer. In particular, it would help to explain how the textual decoder relates to latent reception and how prefix adaptation and dynamic LoRA work together during training and inference.

Questions for authors
  1. The paper motivates dynamic LoRA by suggesting latent-prefix information may weaken across layers. However, textual messages are also introduced only at the input. Could the authors clarify the motivation for this design? Does the improvement come from conditioning multiple layers or from additional information in the latent messages?
Reviewer_eUNHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

UniCom proposes a communication architecture for heterogeneous LLM agents. Sender-specific encoders map latent reasoning states into a shared representation, which supports a textual preview decoder and two receiver adaptation mechanisms: continuous input prefixes and message-conditioned LoRA updates. Communication modules are trained using task, rationale reconstruction, and latent reconstruction losses, with the backbones frozen during adapter training. Experiments cover five benchmarks and two model configurations, with additional component ablations and efficiency analyses. The paper reports the highest accuracy among the evaluated methods in all ten main task/configuration settings.

Critical strengths
  • A useful modular architecture for heterogeneous communication. Combining a shared message representation with receiver-specific prefixes and generated low-rank updates is a practical design. Keeping the backbones frozen during adapter training could make the approach useful when full model retraining is undesirable.

  • Consistent reported improvements with informative component ablations. UniCom leads all ten cells in Table 1, with gains of approximately 0.8–2.4 percentage points over the strongest listed baseline in each cell. Table 3 also shows that removing textual supervision, reconstruction, prefix adaptation, or parameter adaptation reduces performance. These results motivate further investigation of the combination.

Critical weaknesses

The evaluated protocol and comparison conditions are insufficiently clear. Figure 2 and Algorithm 1 describe latent communication with an optional generated textual preview. However, Appendices D/E state that UniCom receives the same cached sender rationales as TextCom, and Appendix I explicitly places these rationales in the receiver prompt. Thus, the evaluated system appears to consume original textual rationales in addition to latent messages. It is unclear whether the latent baselines receive equivalent textual context and whether TextCom receives comparable task-specific adaptation.

Moreover, Appendix I restricts textual communication to short hints: senders cannot disclose answers, perform arithmetic/symbolic calculations, or write code. This limits the conclusions that can be drawn about text communication generally. The efficiency explanation also attributes speedups to avoiding sender text generation, although the appendix says these texts are used during evaluation. The comparison needs a consistent accounting of information, supervision, and end-to-end computation.

The novelty and interpretability claims exceed the evidence. HyLaT (Mou et al., May 2026) already proposes hybrid latent–text communication, while TFlow (Bao et al., May 2026) converts sender states into transient receiver LoRA updates. Neither is discussed. Interlat also already includes text-derived plan alignment. UniCom’s specific combination may be useful, but its contribution should be distinguished from these precedents.

The claimed auditability is also not demonstrated. Rationale reconstruction does not ensure that the preview captures the information driving receiver predictions. Table 3 establishes the utility of auxiliary supervision, not preview faithfulness. Appendix H shows sender rationales rather than clearly identified outputs of the preview decoder. Its HumanEval example explicitly includes sorting and returning in the textual rationale, so that example does not establish that these requirements were lost during textual communication.

Questions for authors

rationales, generated previews, latent messages, or a combination? Please reconcile Figure 2/Algorithm 1 with Appendices D, E, and I. Does the reported latency include generating all textual rationales used at inference, latent generation, and parameter generation?

Comparative evidence: How does UniCom differ from HyLaT and TFlow? Can you provide a controlled comparison against unrestricted textual communication and an Interlat-style baseline with the same textual context, training data, and comparable adaptation capacity?

Faithfulness evidence: Can you show actual outputs of the preview decoder and evaluate whether they reflect receiver-relevant message content? A controlled message substitution or intervention study would help distinguish informative previews from plausible reconstructions that omit causally important latent information.

Reproducibility and uncertainty: Please specify training datasets/splits, sender initialization, decoder architecture, adapter parameter counts, and optimization settings. Table 5 uses only 336 MMLU and 239 C-Eval examples, and several improvements correspond to roughly four additional correct answers. Please report paired uncertainty estimates and clarify whether Table 1 reports single runs or averages.

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

Related papers

Discussion (0)

Sign in to comment.