AIAny
Icon for item

Beyond Euclidean Clipping: Overcoming Exploration Collapse in LLM RL via Riemannian Isometric Policy Optimization

Proposes Riemannian Isometric Policy Optimization (RIPO) to fix exploration collapse in PPO-style RL for LLMs by aligning policy updates with the policy manifold's Riemannian geometry, improving exploration–exploitation balance and optimization stability across competition benchmarks.

Introduction

Most RL fine-tuning pipelines for LLMs rely on PPO-Clip, but this paper argues a subtle geometric mismatch causes exploration collapse: PPO-Clip implicitly measures policy discrepancy with a Euclidean metric while the true policy space has Riemannian geometry. That mismatch makes updates overly conservative in low-probability regions and overly aggressive in high-probability regions, shrinking effective exploration. RIPO enforces isometric updates on the Riemannian manifold to correct this imbalance, yielding a better bias–variance trade-off and more stable optimization.

Key Findings
  • Riemannian correction: Replaces Euclidean discrepancy measures with Riemannian-aware, isometric policy updates so that step sizes reflect intrinsic policy distances rather than parameter-space Euclidean distances — this reduces biased update behavior across probability regions.
  • Exploration–exploitation balance: The isometric update rule prevents premature collapse of exploration in low-probability actions while avoiding excessive updates in high-probability regions, producing more consistent policy improvement.
  • Empirical gains: Across seven competition-level benchmarks, RIPO substantially outperforms prior LLM-RL methods; reported improvements include up to 60% better performance versus GRPO on AIME24, and generally more stable training curves.
  • Stability via bias–variance trade-off: The formulation achieves a favorable bias–variance compromise, which the authors show stabilizes optimization compared with PPO-Clip variants.
Who it's for and trade-offs

Great fit if you are fine-tuning LLMs with RL for tasks where safe, sustained exploration matters (e.g., open-ended generation, multi-step reasoning, or competition-style benchmarks) and you observe collapse or instability with standard PPO-Clip. Look elsewhere if your application cannot tolerate added per-update computation or if you need a plug-and-play replacement with no modification to policy parameterization — RIPO requires estimating and using Riemannian geometry, which can increase implementation complexity and runtime cost. It is also primarily evaluated in LLM RL benchmarks, so adaptations may be needed for other RL domains.

Where it fits

RIPO aims to sit between heuristic PPO fixes and more complex trust-region methods by explicitly addressing the geometry mismatch rather than tuning clipping or learning rates. Compared with PPO-Clip it targets a principled metric correction; compared with methods like GRPO it reports large empirical gains on the tested benchmarks.

Brief method note

The core mechanism enforces isometric updates on the policy manifold so that policy discrepancies are measured and corrected in a geometry-consistent way. That requires computing or approximating local Riemannian metric information and applying update rules that preserve intrinsic distances, which drives the improved exploration behavior and stabilized variance observed in experiments.

Information

  • Websitearxiv.org
  • AuthorsZhicheng Cai, Xinyuan Guo, Hanlin Wu, Mingxuan Wang, Wei-Ying Ma, Ya-Qin Zhang, Hao Zhou
  • Published date2026/07/11

More Items

Studies train-time knowledge injection via hypernetworks that generate fixed LoRA adapters from large fact corpora, empirically characterizing power-law scaling across hypernetwork depth, width, and target model size and reporting improved OOD generalization.

Quantifies active visual observation in multimodal LLMs with ActiveVision, a 17-task benchmark that forces repeated perception rather than one-shot description. Finds frontier MLLMs fail badly (top model 10.6% vs humans 96.1%) and that model-generated vision code does not close the gap.

Performs full-parameter post-training of trillion-parameter MoE DeepSeek-V4 models on an Ascend NPU SuperPOD, using a hierarchical optimization of model parallelism, communication orchestration, and kernel execution to increase Model FLOPs Utilization. Also builds CPT/SFT pipelines with solver-verified synthetic data for Operations Research, reporting strong zero-shot Pass@1 results.