Research map / Multi-agent systems and games / LLM agents
Thought Communication in Multiagent Collaboration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes thought communication, where LLM agents share latent thoughts extracted from their hidden states via a sparsity-regularized autoencoder with identifiability guarantees, and evaluates it on MATH and GSM8K debate settings.
Problem
LLM-based multi-agent systems communicate only via natural language tokens or embeddings, which is lossy and ambiguous, limiting collaboration; the paper asks how to recover shared and private latent thoughts that agents can exchange directly.
Approach
Model agent hidden states as generated by an unknown invertible function of latent thoughts; prove that shared and private thoughts between agent pairs, and the thought-agent incidence structure, are identifiable under sparsity (l0/l1 on the Jacobian); then extract thoughts with a sparsity-regularized autoencoder, route them by agreement-weighted structure, and inject them via prefix adaptation.
Key results
- On MATH with Qwen-3-1.7B, ThoughtComm reaches 93.00% accuracy versus 75.80% for Multiagent Finetuning and 43.60% for single answer.
- On GSM8K with Qwen-3-0.6B, ThoughtComm reaches 75.80% accuracy versus 70.80% for Multiagent Finetuning and 58.20% for single answer.
- On MATH with Llama-3-8B-Instruct, ThoughtComm reaches 45.60% accuracy versus 39.68% for Multiagent Finetuning; on Phi-4-mini, ThoughtComm (74.60%) trails single answer (63.80%) only on... it exceeds Multiagent Finetuning (60.20%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ThoughtComm instead of Multiagent Finetuning | Qwen 3-0.6B multiagent collaboration · MATH | accuracy: 71.20 ± 2.03 -> 85.00 ± 1.60 | Helped |
| Use ThoughtComm instead of Multiagent Finetuning | Qwen 3-0.6B multiagent collaboration · GSM8K | accuracy: 70.80 ± 2.03 -> 75.80 ± 1.92 | Helped |
| Use ThoughtComm instead of Multiagent Finetuning | Qwen 3-1.7B multiagent collaboration · MATH | accuracy: 75.80 ± 1.92 -> 93.00 ± 1.14 | Helped |
| Use ThoughtComm instead of Multiagent Finetuning | Qwen 3-1.7B multiagent collaboration · GSM8K | accuracy: 84.20 ± 1.63 -> 85.00 ± 1.60 | Helped |
| Use ThoughtComm instead of Multiagent Finetuning | Phi-4-mini-instruct (3.84B) multiagent collaboration · MATH | accuracy: 60.20 ± 2.19 -> 74.60 ± 1.95 | Helped |
| Use ThoughtComm instead of Multiagent Finetuning | Phi-4-mini-instruct (3.84B) multiagent collaboration · GSM8K | accuracy: 82.16 ± 1.71 -> 84.20 ± 1.63 | Helped |
| Use ThoughtComm instead of Multiagent Finetuning | LLaMA 3-8B-Instruct multiagent collaboration · MATH | accuracy: 39.68 ± 2.19 -> 45.60 ± 2.23 | Helped |
| Use ThoughtComm instead of Multiagent Finetuning | LLaMA 3-8B-Instruct multiagent collaboration · GSM8K | accuracy: 69.20 ± 2.06 -> 68.40 ± 2.08 | Hurt |
| Use ThoughtComm instead of Multiagent Finetuning | DeepSeek-R1-Distill-Llama-8B multiagent collaboration · MATH | accuracy: 72.40 ± 2.00 -> 82.80 ± 1.69 | Helped |
| Use ThoughtComm instead of Multiagent Finetuning | DeepSeek-R1-Distill-Llama-8B multiagent collaboration · GSM8K | accuracy: 76.80 ± 1.89 -> 80.80 ± 1.76 | Helped |
| Replace natural-language multiagent communication with latent thought communication | multiagent collaboration across the reported LLMs · MATH and GSM8K | consensus | Mixed |
Surprising findings
- Multiagent Finetuning's accuracy declines as debate rounds increase beyond 4 on Qwen-3-1.7B while its consensus keeps rising, indicating premature agreement on wrong answers.
Limitations
- Relies on access to model hidden states, which may not be available for closed-source models.
- Only evaluated on math reasoning benchmarks with small open models (0.6B to 8B) and three agents in debate.
- Identifiability guarantees are asymptotic and rely on assumptions such as invertibility and Jacobian spanning conditions.
Experimental setup
Datasets: MATH GSM8K
Benchmarks: MATH GSM8K
Metrics: accuracy consensus R^2 mean correlation coefficient
Baselines: Single Answer Multiagent Finetuning
Builds on: Multiagent Finetuning prefix tuning multi-agent debate
Closest papers by meaning
- Latent Collaboration in Multi-Agent Systems (ICML 2026 · similarity 0.90)
- MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems (NeurIPS 2025 · similarity 0.89)
- Mental-Models for Multi-Agent Systems (NeurIPS 2026 · similarity 0.89)
- Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action (NeurIPS 2026 · similarity 0.89)
- Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs (ICML 2026 · similarity 0.88)
- Hypothetical Minds: Scaffolding Theory of Mind for Multi-Agent Tasks with Large Language Models (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.