AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

XRPO: Pushing the Limits of GRPO with Targeted Exploration and Exploitation

Udbhav Bamba, Minghao Fang, Yifan Yu, Haizhong Zheng, Fan Lai

ICML 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add ICL seeding for zero-reward prompts (ablation)XRPO · Qwen3-1.7B, 4 math benchmarks avgcons@32: 43.33 -> 47.50Improved
add novelty-guided advantage sharpening (ablation)XRPO · Qwen3-1.7B, 4 math benchmarks avgpass@1: 34.35 -> 37.81Improved
add hierarchical rollout planning over uniform allocation (ablation)XRPO · Qwen3-1.7B, 4 math benchmarks avgpass@1: 35.10 -> 37.81Improved
use XRPO instead of GSPOGSPO · Qwen3-1.7B, AIME'25pass@1: 33.23 -> 35.72Improved
use XRPO instead of GRPOGRPO · Llama-3.2-3B, BRUMO'25pass@1: 4.79 -> 3.54Hurt
use XRPO instead of GRPOGRPO · Llama-3.2-3B, avg of 3 benchmarkspass@1: 4.17 -> 4.48Improved
add test-time ICL on top of XRPO (ablation)XRPO · Qwen2.5-7B-Instructpass@1: 12.05 -> 11.77Hurt
use strong embedding retrieval for ICL vs no ICL (ablation)ICL seeding · DAPO-Math-17k, Qwen2.5-7Baccuracy: 18.6 -> 24.5Improved
use random ICL retrieval vs no ICL (ablation)ICL seeding · DAPO-Math-17k, Qwen3-1.7Baccuracy: 37.7 -> 38.12Improved
truncate 20% of ICL exemplars (ablation)ICL seeding · Qwen3-1.7Baccuracy: 41.5 -> 39.0Hurt
use uncertainty-based allocation instead of difficulty-based CurESXRPO · Qwen3-1.7B, avg of 3 benchmarkscons@32: 32.22 -> 45.56Improved
combine XRPO with GSPOGSPO · Qwen3-1.7B, HMMT'25pass@1: 21.2 -> 22.8Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy