Thought Communication in Multiagent Collaboration
Yujia Zheng, Zhuokai Zhao, Zijian Li, Yaqi Xie, Mingze Gao, Lizhu Zhang, Kun Zhang
NeurIPS 2025 Spotlight
cs.LG, cs.AI, cs.MA
2025-10-24
ThoughtComm recovers shared and private latents from hidden states and injects them as prefixes. With three agents, Qwen-3-1.7B reaches 93% on MATH versus 43.6% solo.
LLM multi-agent systems still talk in natural language. Language is lossy, ambiguous, and indirect. Failure analyses by Cemri and others trace many coordination breakdowns to vague messages and misalignment, both products of that channel. Machines do not have vocal cords, so a more direct channel is in principle available.
Prior MAS work has changed topology, workflows, and even token-level collaboration, including exchanging embeddings. The medium is still language. This paper asks whether agents can transmit the thoughts themselves.
Each agent's model state before a communication round (the hidden representation of its last generated token) is treated as generated from a vector of latent thoughts by an unknown invertible map. The Jacobian's sparsity pattern says which thoughts influence which agents. Under a nonparametric setup with no auxiliary variables, sparsity regularization on that Jacobian identifies shared thoughts and private thoughts for any pair of agents, and recovers the global thought-to-agent incidence structure up to permutation.
ThoughtComm implements this. A sparsity-regularized autoencoder maps concatenated hidden states to estimated thoughts. A Jacobian mask routes relevant dimensions to each agent and reweights them by how many agents share each dimension. An adapter turns the result into a prefix prepended to the next-round token embeddings. The autoencoder only reconstructs hidden states; the adapter only keeps injected text fluent. Both pieces are largely task-agnostic and can be pretrained once.
The main protocol uses three agents and two debate rounds, with 500 problems to train the communication module and 500 to evaluate. Baselines are a single model and Multiagent Finetuning, which fine-tunes the full LLM.
MATH is the sharpest table:
| Base | Single | Multiagent FT | ThoughtComm |
| Qwen-3-0.6B | 45.80 | 71.20 | 85.00 |
| Qwen-3-1.7B | 43.60 | 75.80 | 93.00 |
| Phi-4-mini 3.84B | 63.80 | 60.20 | 74.60 |
| Llama-3-8B | 36.20 | 39.68 | 45.60 |
| DeepSeek-R1-Distill-Llama-8B | 42.60 | 72.40 | 82.80 |
On Qwen-3-1.7B the jump is 49.4 points over the single model and 17.2 over Multiagent FT. The paper's aggregate claim is +67.23% relative to single-answer and +19.06% relative to that SOTA. GSM8K mostly follows, except Llama-3-8B at 68.40 versus FT's 69.20.
On synthetic mixes, the sparse autoencoder recovers shared and private blocks; the unregularized baseline does not. Across eight setups with dimension 124 to 1024, MCC clears the usual identifiability threshold. Stretching debate from 2 to 6 rounds drops FT accuracy while consensus still rises (empty agreement). ThoughtComm lifts both. Prefix length from 1 to 16 moves scores by under five points.
Identifiability from causal representation learning is wired into MAS communication. Training cost scales with embedding width, not parameter count: Llama-3-70B and 405B share a 16,384-d embedding, so ThoughtComm's extra training does not grow with the LLM. Closed-source models do not expose hidden states, and the authors say so.
The experiments need open-weight hidden states. A closed API would have to recover thoughts from embeddings of observed text; that variant is not run. Evaluation is math only, up to 8B, with a fixed three-agent two-round protocol. Llama-3-8B losing on GSM8K is a reminder that latent injection is not free. Identifiability is a statistical guarantee. It does not mean a recovered dimension names a human concept.