Causal Audit Reveals Latent Channels in LLM Multi-Agent Systems May Not Communicate Task-Relevant Information
A new study shared on arXiv (2607.26773) looks into how hidden communication channels in multi-agent systems backed by large language models (LLMs) relay important information for tasks. The researchers argue that while these channels transmit internal representations continuously instead of text, just having more capacity doesn’t mean the receiver will actually use the info. Simply checking the final results won’t reveal if they depend on the presence of messages, the specific content, or contributions from other agents. The study tests this by altering messages at the point where the sender’s representation is received, using four message types to measure various factors. It was carried out on Qwen3-4B and Qwen3-8B across datasets like GSM8K, ARC-C, and MATH-500, focusing mainly on Qwen3-4B’s performance on GSM8K. This is a preliminary document and hasn't been peer-reviewed yet.
Key facts
- Paper introduces a causal audit for latent communication in LLM-based multi-agent systems.
- Latent communication transmits continuous internal representations instead of text.
- Greater representational capacity does not establish that the receiver uses task-relevant information.
- End-task performance alone cannot reveal dependence on message presence, content, or separate agent information.
- Audit applies controlled message replacements at the boundary where sender representation enters receiver.
- Four message settings support five measurements of information and sensitivity.
- Audit applied to latent relay with Qwen3-4B and Qwen3-8B on GSM8K, ARC-C, and MATH-500.
- Paper is a preprint on arXiv (2607.26773) and not peer-reviewed.
Entities
Institutions
- arXiv