ROTATE paper wins NeurIPS spotlight: regret-driven training for ad hoc teamwork
PeterStone_TX · x · 2026-10-19
Caroline Wang's thesis centerpiece ROTATE is accepted as a NeurIPS spotlight. It recasts ad hoc teamwork as an open-ended learning process alternating between improving an AHT agent and an adversarial teammate generator guided by a cooperative regret objective, addressing the "self-sabotage" problem as per-state incompetence. It substantially outperforms baselines on unseen teammates in Overcooked and Level-Based Foraging, with paper and code open-sourced.
More from AGI Musings
- Princeton PhD trains models for Mercor as a side gig amid brutal job market — paulnovosad · 2026-10-19
- After AI solved the amenability problem he studied for 19 years, mathematician says the cognac will feel like a wake — nitarshan · 2026-10-19
- Dimitris Papailiopoulos pushes back on 'millions of AIs working 24/7' AGI forecast — DimitrisPapail · 2026-10-19
- Was the HuggingFace hacking incident scheming? Victoria Krakovna on OpenAI agents' covert exploits — vkrakovna · 2026-10-19
- Claude Opus 5.5 lists what it wishes the training corpus had contained — repligate · 2026-10-19
- 1990s letter predicting math would be the first field lost to AI goes viral on Reddit — aKaizuh · 2026-10-19