PoEM: predicting RL policy outcomes from new rewards without training
giannis_daras · x · 2026-11-21
A new paper introduces PoEM, a method that predicts what policy RL would learn under a new reward — without actually training on it.
The approach exploits linearity in policy space: by composing models previously trained on other rewards, PoEM estimates RL outcomes for new reward functions, avoiding costly training runs.
More from Research
- Inkling 2 Mini: a 26B looped transformer that rivals 594B models on agentic work — ziqiao_ma · 2026-11-21
- Multi-turn agentic RL on one TPU lifts Gemma 4 E2B from 9.4% to 95% win rate — Saboo_Shubham_ · 2026-11-21
- JapanFold drug discovery platform predicts 1M+ protein structures at $0.0022 per fold — DavidBennett__ · 2026-11-21
- "Would you please look at the data?": giffmana flags bad annotations in FineVision's coco_color — giffmana · 2026-11-21
- Vishal Sikka on reported Riemann Hypothesis proof with Claude: 'an incredible day' — vsikka · 2026-11-21
- NDIF hosts Thinking Machines' Inkling 2 Mini and launches a Looped Lens interpretability challenge — benno_krojer · 2026-11-21