Researchers introduced CoSIR, a post-training method for teaching large language model (LLM) agents to collaborate, in a preprint published on October 13. The method rewards agents for sharing relevant evidence, requesting missing information and using teammates’ contributions in their decisions. The researchers say final task success alone gives limited feedback on which messages helped a team reach the correct answer.

CoSIR uses a verifiable environment in which agents see different parts of a rating table and exchange information to select the best option. A Bayesian evaluator (a method that updates probabilities as evidence arrives) scores shared information by how much it reduces team surprisal, defined in the paper as the sum of negative log probabilities assigned to the correct option given each agent’s evidence. Requests receive scores based on the potential reduction in surprisal if truthful answers are provided, and separate rewards encourage agents to use received information in their private choices. The researchers use self-play reinforcement learning to apply these rewards directly to the underlying models.

The team post-trained Qwen3-4B, Qwen3-8B and Gemma-3-4B-it, then tested transfer on 7 unseen natural-language collaboration benchmarks covering distributed evidence, mathematical reasoning and collaborative decision-making. CoSIR achieved the highest combined score across the 7 tasks among compared methods on all 3 model backbones; on Qwen3-8B, its score was 31.85 points above the base model.

Behavioral analysis found that Qwen3-8B’s disclosure of private facts on HiddenBench rose from 68.5% to 74.3%. In an analysis using MAST across the 3 models, failures to use teammates’ input fell from 26.3% to 22.8%, while failures to seek clarification dropped from 36.4% to 32.0%. The researchers report that performance on GSM8K, MMLU and IFEval remained largely preserved.

Researchers introduced CoSIR, a post-training method for teaching large language model (LLM) agents to collaborate, in a preprint published on October 13. The method rewards agents for sharing relevant evidence, requesting missing information and using teammates’ contributions in their decisions. The researchers say final task success alone gives limited feedback on which messages helped a team reach the correct answer.

CoSIR uses a verifiable environment in which agents see different parts of a rating table and exchange information to select the best option. A Bayesian evaluator (a method that updates probabilities as evidence arrives) scores shared information by how much it reduces team surprisal, defined in the paper as the sum of negative log probabilities assigned to the correct option given each agent’s evidence. Requests receive scores based on the potential reduction in surprisal if truthful answers are provided, and separate rewards encourage agents to use received information in their private choices. The researchers use self-play reinforcement learning to apply these rewards directly to the underlying models.

The team post-trained Qwen3-4B, Qwen3-8B and Gemma-3-4B-it, then tested transfer on 7 unseen natural-language collaboration benchmarks covering distributed evidence, mathematical reasoning and collaborative decision-making. CoSIR achieved the highest combined score across the 7 tasks among compared methods on all 3 model backbones; on Qwen3-8B, its score was 31.85 points above the base model.

Behavioral analysis found that Qwen3-8B’s disclosure of private facts on HiddenBench rose from 68.5% to 74.3%. In an analysis using MAST across the 3 models, failures to use teammates’ input fell from 26.3% to 22.8%, while failures to seek clarification dropped from 36.4% to 32.0%. The researchers report that performance on GSM8K, MMLU and IFEval remained largely preserved.