arXiv AI By Jiamian Wang, Samyadeep Basu, Koustava Goswami, Tong Yu, Zhiqiang Tao

Spurious Advantage Hidden in GRPO

Read the original on arXiv AI →

Group Relative Policy Optimization (GRPO) assigns a magnitude to each rollout based on within‑group reward statistics, rewarding rollouts that reach the correct answer through reasoning. However, the same magnitude can be high for rollouts that reach the answer by guessing, creating a spurious advantage that misleads the policy toward guess‑like behaviors. The paper identifies three scenarios where this occurs—bounded‑answer tasks, open‑answer sets with bounded sub‑cases, and search agents with many paths to the same answer—and proposes SIGNBALANCE, a composition‑free magnitude that preserves the verifier sign, uses a global scale, and restores zero‑mean balance via stop‑gradient per‑class rescaling, matching GRPO on open‑answer math and improving on bounded‑answer math and search agents.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.

Hugging Face Trending Papers
Sep 3

Spurious Advantage Hidden in GRPO

The paper identifies a flaw in Group Relative Policy Optimization (GRPO) called the spurious advantage, where the advantage estimator can incorrectly reward rollouts that reach correct answers by guessing rather than reasoning. This issue appears in bounded-answer tasks, open-answer sets with bounded sub-cases, and search agents with many paths to the same answer. The authors propose SIGNBALANCE, a new magnitude estimator that removes composition, maintains verifier sign, applies a global scale, and uses stop‑gradient per‑class rescaling to restore zero‑mean balance, achieving comparable or better performance on math and search benchmarks.

arXiv AI
Sep 4

F-GRPO: Don't Let Your Policy Learn the Obvious and Forget the Rare

The paper introduces F-GRPO, a method that addresses the issue of reinforcement learning policies overfitting to common trajectories while neglecting rare correct ones. By deriving the probability of prompt‑local tail‑miss events and proposing a difficulty‑aware scaling coefficient inspired by Focal loss, the authors show that down‑weighting high‑success sampled groups can improve performance. Experiments on categorical simulations, Maze, and large language models (Qwen2.5‑7B) demonstrate that F‑GRPO raises average math pass rates and out‑of‑distribution performance without increasing group size or computational cost.

By Daniil Plyusov, Alexey Gorbatovski, Boris Shaposhnikov, Viacheslav Sinii, Alexey Malakhov, Daria Korotyshova, Daniil Gavrilov