NeurIPS paper: calibrated probability talk beats action talk for multi-agent collaboration
Aaroth · x · 2026-10-21
A new NeurIPS paper from UPenn, MIT, and Cornell studies how two agents holding different private information should communicate.
- Key finding: exchanging only proposed actions discards decision-relevant uncertainty — two agents may prefer the same action for different reasons. Communicating calibrated beliefs (probabilities) over actions is theoretically and empirically strictly more powerful for information aggregation.
- Experiments: validated in a collaborative maze-solving task with trained transformers and pretrained LLM agents; probability communication wins when exchanged predictions preserve calibrated uncertainty.
- Failure mode: biased probabilities erase the advantage. The authors propose a post-hoc conversational calibration intervention that consistently improves decision-making in both toy transformers and pretrained LLMs.
- Takeaway: probabilistic communication helps not merely by transmitting more information, but because calibrated uncertainty gives other agents a reliable object to update on.
More from coding & agent
- Laude Institute launches Slingshots FOUR: 14 projects across 8 top institutions — a1zhang · 2026-10-21
- Research agent Épi sets 120 records, narrows n-queens constant interval 84x — dair_ai · 2026-10-21
- Compaction hooks in practice: keep last n messages verbatim, summarize the rest for full agent continuity — MoonL88537 · 2026-10-21
- The blank prompt box hides real operational burden: context, workflow and supervision fall on users — _AustinCalvert_ · 2026-10-21
- DeepMind paper: self-evolving agents forget as they learn — fastest learner GEPA also breaks old tasks most — rohanpaul_ai · 2026-10-21
- arXiv: Are Self-Evolving Agents Continual Learners? Plasticity comes at the cost of stability — rohanpaul_ai · 2026-10-21