Skip to content

Thought Communication in Multiagent Collaboration

πŸ•’ Published (v1): 2025-01-01 Β· Source: NeurIPS Β· Venue: NeurIPS 2025 Β· link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

πŸ’¬ Ask ChatGPT✦ Ask Claude

TL;DR

ThoughtComm proposes replacing natural-language message passing in multi-agent LLM systems with direct exchange of latent representations ("thoughts") extracted from agent hidden states. The approach is grounded in a nonparametric identifiability theory guaranteeing recovery of both shared and private latent thoughts. A sparsity-regularized autoencoder + prefix-adapter pipeline operationalizes the theory, yielding consistent accuracy and consensus gains over language-based baselines.

Problem

LLM-based multi-agent systems communicate exclusively through natural language tokens or token embeddings, which are lossy, ambiguous, and indirect reflections of agents' internal reasoning. This bottleneck causes vague message specification and inter-agent misalignment, limiting collective intelligence beyond what strong single-agent baselines already achieve.

Method

ThoughtComm formalizes agent responses as generated by an unknown invertible function f mapping latent thoughts Z_t to model states H_t = f(Z_t). A sparsity-regularized autoencoder (ℓ₁ on the Jacobian J_fΛ†) inverts this mapping to extract latent thoughts Ẑ_t from concatenated agent hidden states. The Jacobian's non-zero pattern B(J_fΛ†) identifies which latent dimensions belong to which agents (shared vs. private). Each agent receives a personalized latent vector constructed by agreement-based reweighting (Ξ±_j counts how many agents share each thought dimension). A lightweight adapter g maps this vector into a prefix of length m that is prepended to the agent's token embeddings for the next generation round. Training uses a reconstruction loss plus a semantic-similarity loss to keep injected prefixes linguistically coherent. The autoencoder and adapter are task-agnostic and pretrained once on 500 examples.

Key Contributions

  • Identifiability theory (Thms. 1–3): Proves that shared thoughts, private thoughts, and the full thought-agent incidence structure are recoverable up to permutation in a general nonparametric setting, using only β„“β‚€ sparsity on the Jacobian β€” no auxiliary variables or parametric assumptions required.
  • ThoughtComm framework: End-to-end pipeline (autoencoder β†’ structural routing β†’ prefix adapter) that operationalizes the theory for open-source LLMs.
  • Modality-agnostic design: Theory extends to any observable arising from a hidden generative process, not limited to text.
  • Efficiency: Only the autoencoder and adapter are trained; overhead scales with embedding dimension, not LLM parameter count (cost is identical for 70B vs. 405B LLaMA).

Results

  • On MATH (level-3), Qwen-3-1.7B: ThoughtComm achieves 93.0% vs. 75.8% (Multiagent Finetuning) and 43.6% (single answer) β€” +17.2pp absolute over SOTA, +113.3% relative over single answer.
  • On GSM8K, Qwen-3-1.7B: 85.0% vs. 84.2% (Multiagent Finetuning) and 67.4% (single answer).
  • Average across all five models and both benchmarks: +19.06% over Multiagent Finetuning, +67.23% over single answer.
  • ThoughtComm consensus scores consistently exceed Multiagent Finetuning across all models.
  • With more debate rounds (2β†’6 rounds, LLaMA-3-8B on MATH): Multiagent Finetuning accuracy degrades; ThoughtComm accuracy and consensus both improve.
  • Prefix length sensitivity (m ∈ {1,4,8,16}): performance fluctuations under 5% across all models, including near-optimal performance at m=1.
  • Synthetic MCC experiments: ThoughtComm exceeds the identifiability threshold across all 8 dimensionality settings (128–1024).

Limitations

  • Requires access to LLM hidden states (model states), making it incompatible with closed-source APIs (GPT-4, Claude, Gemini).
  • Evaluated only on math reasoning benchmarks (MATH, GSM8K); generalization to other task types (coding, planning, open-ended QA) is not empirically demonstrated.
  • Only open-weight models of 0.6B–8B scale are tested; very large models (70B+) are discussed theoretically for efficiency but not benchmarked.
  • The extension to non-text modalities (images, audio) is stated as theoretically valid but not empirically validated.
  • 500-example fine-tuning split may not reflect performance under data-scarce or domain-shift conditions.

Relevance to Agentic AI / LLM Agents

ThoughtComm directly attacks a core bottleneck in multi-agent LLM architectures: the dependence on natural-language communication channels that introduce ambiguity and misalignment between agents. By grounding agent-to-agent information exchange in disentangled latent representations with provable identifiability, it offers a principled path toward tighter inter-agent coordination than any token-level or embedding-level communication scheme. The modular, model-agnostic design (small adapter, fixed backbone) makes it practically deployable as a communication layer on top of existing multi-agent debate or workflow systems. This work is directly relevant to researchers building agentic systems that rely on multi-agent debate, role specialization, or collaborative reasoning, as it provides both theory and implementation for a fundamentally new communication substrate.