Group Entropy-Controlled Policy Optimization
GEPO adjusts RL advantage signals by prompt-group entropy to keep mixed-task LLM training comparable.
The paper argues that heterogeneous task mixes create different entropy regimes under the same policy, making global or token-level entropy control too blunt. GEPO extends GRPO by estimating entropy from existing grouped samples and shaping advantages asymmetrically. It dampens positive advantages in low-entropy groups and negative advantages in high-entropy groups. The authors report consistent gains over GRPO and other entropy-controlled methods across thirteen benchmarks on two base models. HF Daily Papers' note
The paper argues that heterogeneous task mixes create different entropy regimes under the same policy, making global or token-level entropy control too blunt. GEPO extends GRPO by estimating entropy from existing grouped samples and shaping advantages asymmetrically. It dampens positive advantages in low-entropy groups and negative advantages in high-entropy groups. The authors report consistent gains over GRPO and other entropy-controlled methods across thirteen benchmarks on two base models. HF Daily Papers' note
score 4