Research map / Large language models / Language, safety and interpretability
Communicating Activations Between Language Model Agents
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 activation communication (AC), where one LM's intermediate-layer activation is grafted into another LM's forward pass via simple functions (sum, mean, replace, or a learned linear map), outperforming natural-language debate at lower compute.
Problem
Natural-language communication between LM agents is expensive in inference compute and discards information in output logits; the paper asks whether LM agents can communicate more efficiently through internal activations.
Approach
Pause agent B at layer j, replace or combine its last-token activation with agent A's layer-k last-token activation via a non-learned function f (sum, mean, replace) or a task-agnostic learned linear map, then continue B's forward pass; this needs only a partial forward pass of A and no task-specific parameters or data.
Key results
- On the coordination games with LLaMA-3.2-3B, AC (replace) reaches 78.0% on Countries and 90.0% on Tip Sheets versus 69.0% and 74.3% for NL communication.
- AC consistently beats both single-model baselines on the seven reasoning benchmarks, with up to 27.0% improvement over NLD on six of seven datasets, using less than 1/4 the compute.
- AC outperforms NLD on 48 of 57 MMLU datasets (average 62.7 vs 60.7).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use activation communication instead of natural-language debate | LLaMA-3.2-3B + LLaMA-3.1-8B · Reasoning benchmarks (six of seven) | accuracy: up to 27.0% improvement over NLD | Helped |
| Add a learned linear projection W before activation replacement (ablation) | AC · Reasoning benchmarks (four of seven) | accuracy: 2.6-50.0% over vanilla AC | Mixed |
| Add chain-of-thought prompting for A before grafting its response activation (ablation) | AC · Reasoning benchmarks | accuracy | No clear effect |
| Train W in-distribution on the GSM8K train set (ablation) | AC (W) · GSM8K | accuracy: 64.0 -> 78.0 | Helped |
| Apply activation addition to all sequence tokens instead of only the last token (ablation) | AC · Reasoning benchmarks | accuracy | Hurt |
| Use replacement rather than sum or mean to combine activations (ablation) | AC · Countries and Tip Sheets | accuracy | Mixed |
| Swap the communicating models so the smaller model performs the full forward pass (ablation) | AC · Biographies and GSM8K | accuracy: 81.5±0.0 and 61.0±4.8 | Hurt |
Surprising findings
- Replacing only the last-token activation outperforms summing across all tokens, and applying the operation to all tokens reduces performance.
- Activation communication works across model families (Qwen, Gemma, LLaMA) without training the linear mapping, despite different tokenizers and training data.
Limitations
- Without the learned mapping, requires aligned activation embedding spaces between the two models.
- Requires access to model embeddings, so it does not work with black-box API models.
- Communicated activations are less interpretable than natural language.
Experimental setup
Datasets: C4 Biographies GSM8k MMLU
Benchmarks: Countries Tip Sheets Biographies GSM8k MMLU (High School Psychology, Formal Logic, College Biology, Professional Law, Public Relations; full MMLU in appendix)
Metrics: accuracy exact match bootstrap 95% confidence interval fact-support score
Baselines: Natural Language Debate single model silent (no communication) single-agent skyline CIPHER
Builds on: multiagent debate CIPHER CALM Lewis signaling games
Closest papers by meaning
- Let Models Speak Ciphers: Multiagent Debate through Embeddings (ICLR 2024 · similarity 0.88)
- KVComm: Enabling Efficient LLM Communication through Selective KV Sharing (ICLR 2026 · similarity 0.88)
- Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs (ICML 2026 · similarity 0.88)
- Latent Collaboration in Multi-Agent Systems (ICML 2026 · similarity 0.88)
- When LLMs Develop Languages: Symbolic Communication for Efficient Multi-Agent Reasoning (ICML 2026 · similarity 0.88)
- Benefits and Limitations of Communication in Multi-Agent Reasoning (ICLR 2026 · similarity 0.88)
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.