Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards

Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards. By Keru Chen and four others, from Arizona State University, the University of Houston, and two other institutions. Published on the thirteenth of October. No code repository is linked to it.

When large language model agents work as a team, collaborating well means three behaviors: sharing relevant evidence, asking for information that is missing, and actually folding teammates' contributions into decisions. Training those behaviors is hard. The final outcome of a task says little about the value of any individual message, so a team that succeeds or fails gives almost no signal about which message helped. A useful disclosure and a pointless one can sit in the same winning conversation and receive the same credit.

CoSIR, short for Collaboration via Surprisal-based Information Rewards, is a post-training approach that addresses this with a simple, verifiable environment. The bet is that collaboration behaviors learned there will generalize to more complex tasks the agents have never seen. Verifiable is the important word: in that environment the value of a message can be computed, not guessed.

At its center is a Bayesian evaluator. It scores a piece of shared information by how much it reduces team surprisal, defined as the sum of the negative log probabilities assigned to the correct option, given the evidence each agent has available. A request is scored by the reduction in that same quantity that would follow if the requested information were truthfully provided. Separate rewards then encourage agents to use what they receive, in their private choices and calculations. So every message, whether it shares or asks, gets a score of its own.

The authors post-train Qwen3-4B, Qwen3-8B, and Gemma-3-4B-it through self-play, then test transfer to seven unseen collaboration tasks posed in natural language. CoSIR achieves the highest aggregate score, and its performance on general tasks stays competitive. Behavioral analyses show agents disclosing more of their private evidence, and fewer judged failures to use what teammates said. Those are the very behaviors the rewards were built to encourage, now showing up on tasks the training never contained.

The move here is to replace a distant outcome with a local one. Instead of waiting for the team to win, each message is judged by how much it would sharpen the team's belief in the right answer, which is something a simple environment can verify exactly.