New ABC-GRPO algorithm enhances LLM training stability and performance

Researchers have introduced All-Quadrant Bounded Clipping GRPO (ABC-GRPO), a novel algorithm designed to improve the stability and generalizability of reinforcement learning for large language models. ABC-GRPO addresses an unbounded blind spot in existing Group Relative Policy Optimization (GRPO) methods by applying unconditional clipping across all four quadrants of the likelihood-ratio and advantage space. This new approach demonstrates superior performance on mathematical reasoning tasks using Qwen3 base models, achieving higher scores on Avg@64 and Pass@64 compared to GRPO, SAPO, and dual-clip PPO, while also maintaining higher entropy. The improvements extend to the MATH-500 dataset and out-of-domain code generation tasks on HumanEval. AI

IMPACT Enhances stability and generalizability in LLM training, potentially leading to more robust and capable AI systems.

RANK_REASON The cluster contains a research paper detailing a new algorithm for reinforcement learning with LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• ABC-GRPO
• All-Quadrant Bounded Clipping GRPO
• Chi Liu
• dual-clip PPO
• Group Relative Policy Optimization
• GRPO
• HumanEval
• MATH-500
• Proximal Policy Optimization
• Qwen3

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →