Reinforcement learning research, RLHF/RLVR developments, agent training — from arXiv cs.LG/cs.AI, HuggingFace, Latent Space, Interconnects, and HN.
The curator's five highest-signal stories, merged across sources.
FERPO regularizes continuous-control policy updates with forward entropy so a learned critic's action gradients do not collapse the policy.
→ Online RL that climbs a critic's action gradient is unstable once the critic is wrong. The entropy term is the stabilizer.
A study of when intrinsic rewards actually produce exploration, and when they only look like they do.
→ Most intrinsic-reward bonuses are added on faith. This paper separates the ones that explore from the ones that just reshape the return.
DexPolicy schedules exploration for dexterous manipulation when demonstrations already cover part of the trajectory.
→ Hands waste samples exploring poses the demo already showed. Schedule the exploration and the policy spends its budget on contact.
AMD is buying World Labs for $8.2 billion. Atlas, the lab's reconstruction model, is the stated reason.
→ The largest world-model acquisition on the board. The buyer is a chip company, so the bet is inference of 3D state, not a chatbot.
Runway's GWM Worlds 2 turns video generation into a real-time interactive simulation via WorldPrompt.
→ A world model you can interrupt is a different object from a video model you can only replay.
Reinforcement learning papers from arXiv cs.LG, cs.AI, and Hugging Face Daily Papers.
InterEvolve searches reward programs at test time for humanoid loco-manipulation instead of freezing one reward.
LOCI is a spatial linear memory for streaming world models, so the state does not have to sit in a context window.
A world model that learns which cues to store, instead of keeping a fixed memory of the last observations.
DexPolicy schedules exploration for dexterous manipulation when demonstrations already cover part of the trajectory.
Newsletters and HN picks on RL, RLHF, and agent training.
FERPO regularizes continuous-control policy updates with forward entropy so a learned critic's action gradients do not collapse the policy.
A study of when intrinsic rewards actually produce exploration, and when they only look like they do.
Faynt scales competitive Super Smash Bros. Melee policies instead of stopping at toy self-play.
An advantage operator that stays stable when the whole RL update runs under fully homomorphic encryption.
AMD is buying World Labs for $8.2 billion. Atlas, the lab's reconstruction model, is the stated reason.
Runway's GWM Worlds 2 turns video generation into a real-time interactive simulation via WorldPrompt.