Multi-effort RL reward design: R = S − λ_e·C to push up the whole Pareto frontier
carlesgelada · x · 2026-09-12
Carles Gelada shares his team's multi-effort RL setup aiming to raise the entire Pareto frontier: from first principles they derive the reward R = S − λe·C, with S ∈ {0,1} for solve/fail, C the rollout cost, and λe an effort-specific cost penalty. Later thread posts explain how to pick λe.
Related event: RL Reward Design: Choosing λ_e to Lift the Whole Pareto Frontier(3 posts)→
More from Research
- Just Action Model: decoupling what separates VLA from world-action models — Yonghao Yu · 2026-10-28
- HABIT-Bench: Best agent memory system scores just 33.61% at inferring user habits — SJTU · 2026-10-28
- CodeCap-Bench: executable code beats natural language as video captions — Jiafu Tang · 2026-10-28
- Omni-InteractScope benchmarks proactive interaction in continuous omnimodal streams — Qwen · 2026-10-28
- Mistral releases Timestral time-series foundation models, tops GIFT-Eval — mistralai · 2026-10-28
- NVIDIA's ThinkWorld learns robot skills from unlabeled human videos via chain-of-dynamics reasoning — nvidia · 2026-10-28