acceptodds
Desk rejected at ICLR 2027

What Drives Successful Latent Communication? Tracing Sender Information Through Latent Working Memory

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Latent multi-agent systems can communicate through internal working memory rather than explicit text. In KV-cache relays, however, the transmitted working memory can contain both states produced while the sender processes its private input and states generated later during latent reasoning. Successful whole-relay communication therefore leaves a key attribution question unresolved: which source actually governs the receiver's sender-specific behavior? We study this behavioral provenance attribution problem using paired decision worlds with mutually exclusive sender-private requirements under the same public context. Our Causal Provenance Decomposition (CPD) first verifies sender-specific communication with native relays, then crosses the requirement source of sender-input states and sender-generated states while keeping both components available. Across Qwen3-4B and Llama-3.1-8B-Instruct on 150 paired worlds per Airline and Retail domain, receiver decisions under provenance conflict follow the sender-input source in approximately 63–85% of outputs, versus 1–10% for the sender-generated source. The resulting source-following margin is 52.7–84.0 percentage points and remains positive under conservative reassignment of all unmatched outputs. These results show that successful whole-relay communication does not imply comparable following of the generated-state donor under provenance conflict, motivating latent relays that transfer sender-private information without direct access to sender-input KV states.

Primary area:
foundation or frontier models, including LLMs
Keywords:
latent communication, multi-agent systems, agent communication, counterfactual evaluation, latent representations

Related papers

Discussion (0)

Sign in to comment.