Small language models improve reasoning by sharing hints, preprint finds

A research preprint reports that small language models working together solved more reasoning problems when they shared hints than when they independently generated answers for a vote. In the researchers' TalkMesh system, the most confident agent broadcasts a hint, less confident agents revise their answers, and each keeps a revision only if it scores higher. In one test using Qwen3.5-0.8B on GSM8K, accuracy rose from 0.568 with independent sampling to 0.705 with the mesh. The authors also tested a defense against agents sending false hints: it retained 0.507 accuracy when four of eight agents colluded, compared with 0.000 for majority voting.