The year/Independent research

Paper 2509.22611

Quantile Advantage Estimation: Stabilizing RLVR for LLM Reasoning

Published
Sep 2025
Research lab
Independent
Citations
10
GitHub
29 stars

01 In brief

Summary

The paper introduces Quantile Advantage Estimation (QAE) to stabilize Reinforcement Learning with Verifiable Rewards (RLVR) for LLM reasoning.

It identifies that mean-baseline methods like GRPO and DAPO cause entropy collapse or explosion, harming performance.

QAE replaces the mean reward baseline with a group-wise K-quantile, creating a two-regime gate: for hard queries (success rate ≤ 1−K), it reinforces rare successes; for easy queries (> 1−K), it targets remaining failures.

This minimal change provably bounds one-step entropy change, preventing both collapse and explosion.

Empirically, QAE stabilizes entropy, sparsifies credit assignment (about 80% of responses get zero advantage), and improves pass@1 on AIME'24, AIME'25, and AMC'23 across Qwen3 models (8B, 14B, 30B-A3B) and methods (DAPO, Clip-Cov, KL-Cov, GSPO), while maintaining pass@16.

The work reframes entropy regulation as a baseline-design problem rather than token-level tuning, offering a drop-in improvement that composes with existing techniques.

02 From the paper

Abstract

Reinforcement Learning with Verifiable Rewards (RLVR) strengthens LLM reasoning, but training often oscillates between {entropy collapse} and {entropy explosion}. We trace both hazards to the mean baseline used in value-free RL (e.g., GRPO and DAPO), which improperly penalizes negative-advantage samples under reward outliers. We propose {Quantile Advantage Estimation} (QAE), replacing the mean with a group-wise K-quantile baseline. QAE induces a response-level, two-regime gate: on hard queries (p <= 1 - K) it reinforces rare successes, while on easy queries (p > 1 - K) it targets remaining failures. Under first-order softmax updates, we prove {two-sided entropy safety}, giving lower and upper bounds on one-step entropy change that curb explosion and prevent collapse. Empirically, this minimal modification stabilizes entropy, sparsifies credit assignment (with tuned K, roughly 80% of responses receive zero advantage), and yields sustained pass@1 gains on Qwen3-8B/14B-Base across AIME 2024/2025 and AMC 2023. These results identify {baseline design} -- rather than token-level heuristics -- as the primary mechanism for scaling RLVR.