New LLM RL paper says PPO-Clip hurts exploration and RIPO lifts AIME24 by 60%
burny_tech · x · 2026-07-25
A new paper argues that PPO-Clip in LLM reinforcement learning has a geometric flaw: it measures policy change in Euclidean terms even though policy updates live on a Riemannian manifold.
The authors say this mismatch suppresses rare but useful actions, causing exploration collapse. Their fix, Riemannian Isometric Policy Optimization (RIPO), gives low-probability actions more room and high-probability actions less, improving stability and reportedly boosting AIME24 performance by up to 60% over GRPO across seven benchmarks.
Related event: RIPO Overcomes PPO-Clip's Exploration Collapse in LLM RL(2 posts)→
More from Research
- AI veteran Thomas Dietterich: stop using cs.AI as your primary arXiv category — tdietterich · 2026-10-27
- arXiv advice from T. Dietterich: wrong primary category will almost surely get your paper on hold — tdietterich · 2026-10-27
- arXiv has 4,000+ papers on hold amid submission surge; moderator shares how to avoid it — tdietterich · 2026-10-27
- Community-trained MOSS Sharpness Slider LoRA: 72 photos, 500 steps, +/-1 focus control — aoleg77 · 2026-10-27
- CoreWeave, You.com and NVIDIA lift BrowseComp 22.9% in 3-week RL post-training sprint — RichardSocher · 2026-10-27
- Open-source decision models sold their own base in Red Alert 2 — then learned until no human could beat them — heathercmiller · 2026-10-27