LGMay 20

Advantage Collapse in Group Relative Policy Optimization: Diagnosis and Mitigation

arXiv:2605.2112534.6
Predicted impact top 10% in LG · last 90 daysOriginality Incremental advance
AI Analysis

For practitioners training LLMs with RLVR, this diagnoses and mitigates a critical failure mode in GRPO, improving training efficiency and final performance.

GRPO suffers from advantage collapse where homogeneous rewards cause vanishing gradients. The proposed AVSPO method reduces collapse by 58-63% and improves accuracy by 4-6 percentage points across 0.5B-14B models.

Group Relative Policy Optimization (GRPO), a prominent algorithm within the Reinforcement Learning from Verifiable Rewards (RLVR) framework, has achieved strong results in improving the reasoning capabilities of large language models (LLMs). However, GRPO is prone to advantage collapse, a failure mode where homogeneous rewards within a group (e.g., all correct or all incorrect answers) yield near-zero advantages and vanishing gradients. To address this, we introduce the Advantage Collapse Rate (ACR), the first diagnostic metric quantifying the proportion of training batches with ineffective gradients. Across models from 0.5B to 14B parameters on mathematical reasoning benchmarks, we show that ACR strongly predicts training stagnation and final performance. We then propose Adaptive Virtual Sample Policy Optimization (AVSPO), a lightweight extension of GRPO that injects virtual reward samples, guided by real-time ACR monitoring, to enable learning from homogeneous groups without additional model rollouts. AVSPO reduces advantage collapse by 58-63% relative to GRPO and yields consistent accuracy gains of 4-6 percentage points across all model scales, while maintaining generalization on the evaluated out-of-domain task. Code and datasets are available at https://qingyonghu.github.io/AVSPO.

Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes