ROTATE, an open-ended training method for generalizable teamwork agents, wins NeurIPS 2026 Spotlight
PeterStone_TX · x · 2026-10-21
ROTATE, a paper on learning generalizable coordination behaviors for multi-agent teamwork, has been accepted to NeurIPS 2026 as a Spotlight. It frames coordination generalization as an open-ended learning problem: the system generates challenging teammates to improve generalization to unseen partners. The latest version also advances the theory, characterizing the self-sabotage problem as per-state incompetence and viewing ROTATE as imposing a competence constraint on learning — explaining why the method works. Paper and code are open-source; work led by Caroline Wang and Arrasy Rahman with Peter Stone's team involved.
More from Research
- After TikTok child-safety alarms, a NeurIPS paper proposes TRACE to evaluate youth restrictions — Stefania_druga · 2026-10-21
- Anthropic's unreleased bio model ran 300 sub-agents, picked 211 drug candidates, 17 bound — ZeroStateReflex · 2026-10-21
- Matrix-Game 4.0: real-time playable game worlds on a single RTX 5090 via engine-video split — Jihai Zhang · 2026-10-21
- Lafa scales continuous flow LMs to 8B, hitting 2.54x AR throughput via self-speculative decoding — Arnav Chavan · 2026-10-21
- SWIFT: Alibaba and UMich paper cuts diffusion training cost via weak-denoiser target extrapolation — alibabagroup · 2026-10-21
- Berkeley correction training cuts diffusion exposure bias, high-quality samples without guidance — UCBerkeley · 2026-10-21