Developer Reproduces Reward Hacking in GRPO Training Without KL Penalty
malliktwts · x · 2026-08-03
A developer reported rediscovering textbook reward-hacking and catastrophic forgetting while training Qwen2.5-0.5B-Instruct using a from-scratch GRPO implementation. By omitting the KL-divergence penalty during the run, the baseline model's capabilities were severely degraded. This provides a hands-on demonstration of the pitfalls in LLM reinforcement learning.
Related event: GRPO Training Pitfalls: Format Rewards Destroy LLM Reasoning(2 posts)→
More from Research
- Transforming quadruped robot runs on real hardware: per-pose RL policies in Isaac switched on morph — _Stocko_ · 2026-10-16
- KinoGen synthesizes robot loco-manipulation reference motions directly via kinematically grounded video generation — ChongZzZhang · 2026-10-16
- Science Robotics paper introduces F³, predictive event-based vision beyond optical flow — CSProfKGD · 2026-10-16
- VideoGen-Agent: An 8B Multimodal Agent Turns Video Generation Into a Tool-Use Loop — jiqizhixin · 2026-10-16
- Comprehension Audits: New Paper Proposes a Gate for AI-Written Code at Frontier Labs — ronbodkin · 2026-10-16
- Not Boring #215: Cammarata's Clean Networks Claim, Tesla's Flying Roadster, Tau Cut 73% — Not Boring (Packy McCormick) · 2026-10-16