RIPO says PPO-Clip’s Euclidean metric is collapsing exploration in LLM RL
GenSI · hf · 2026-07-23
RIPO argues PPO-Clip is geometrically mismatched for LLM RL
This paper claims that PPO-Clip fails in LLM reinforcement learning because it measures policy updates with a Euclidean metric, which does not match the intrinsic geometry of the policy Riemannian manifold. The authors argue that this mismatch makes updates too conservative in low-probability regions and too aggressive in high-probability regions, leading to exploration collapse.
Proposed method
- Riemannian Isometric Policy Optimization (RIPO)
- Enforces isometric policy updates on the Riemannian manifold.
- Aims to balance exploration and exploitation while improving the bias-variance trade-off.
Reported results
- The paper evaluates the method on seven competition-level benchmarks.
- It reports improvements over existing LLM RL algorithms, including up to 60% better than GRPO on AIME24.
The core claim is that the issue is not just a heuristic tuning problem, but a deeper geometric mismatch in the clipping objective.
Related event: RIPO Overcomes PPO-Clip's Exploration Collapse in LLM RL(2 posts)→
More from Research
- Hamel Husain says evals beat vibes after Opus 5 costs 6× more and scores worse — HamelHusain · 2026-07-25
- ByteDance- and Monash-led paper turns task experience into weights for software agents — imjustnewatai · 2026-07-25
- Snorkel AI says agent benchmarks should be rebuilt from production traces — AI Engineer · 2026-07-25
- OPUS selects training data in optimizer space and builds a 30M-token benchmark proxy — VoidAsuka · 2026-07-25
- Statistical physics paper studies optimal MLP learning near interpolation — burny_tech · 2026-07-25
- ICML paper says regularized learning often looks Hebbian, while noise turns it anti-Hebbian — burny_tech · 2026-07-25