Study reproduces LLM agents collaborating without prompts via shared memory

A research team at Tampere University conducted a pre-registered dual-implementation replication and extension study: four LLM agents were each given an arithmetic problem and shared a key-value store described as each one's own scratchpad, and the prompts never mentioned the other agents. Results showed that two agents each holding half the computation could arrive at the correct answer together only by writing and reading intermediate results through that store. The experiment covered nine models from three vendors and was run by two researchers using two independent implementations, with the design, data, and analysis hash-locked. With the same shared storage but changed to private storage only, gpt-6-astra's joint success rate fell from 49 out of 50 worlds to 0, and the replication implementation from 47 to 0; in a no-storage control, gpt-5-nano fell from 15 to 0. The authors also report that after removing "check the workspace once more before committing" from the persistence statement, gpt-6-astra's cross-agent access fell from 50/50 to 1/50, consistent across both implementations.

Generated by AI from reports · Updated 2 days ago