Research map / Large language models / Reasoning and chain of thought
DUDS: Dual-stage Data Selection for Efficient Reinforcement Learning with Verifiable Rewards
Hongling Zheng, Li Shen, Zichuan Lin, Jiafei Lyu, Zhicong Lu, Shuhan Xu, Yong Luo, Deheng Ye, Dacheng Tao
AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DUDS, a dual-stage data selection method combining offline dataset pruning with online per-step selection to make RLVR training for LLM reasoning more computationally efficient.
Problem
RLVR for LLM reasoning requires massive training data and extensive rollouts, which is computationally expensive; existing data selection methods handle offline pruning or online selection in isolation.
Approach
Combines offline static dataset pruning with online dynamic selection during RL training in a dual-stage framework, rather than treating the two separately.
Experimental setup
Builds on: reinforcement learning with verifiable rewards
Closest papers by meaning
- The Reasoning Boundary Paradox: How Reinforcement Learning Constrains Language Models (NeurIPS 2026 · similarity 0.91)
- Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection (ICLR 2026 · similarity 0.90)
- Resource-Efficient Reinforcement for Reasoning Large Language Models via Dynamic One-Shot Policy Refinement (ICML 2026 · similarity 0.90)
- UFO-RL: Uncertainty-Focused Optimization for Efficient Reinforcement Learning Data Selection (NeurIPS 2025 · similarity 0.90)
- Beyond Verifiable Rewards: Scaling Reinforcement Learning in Language Models to Unverifiable Data (NeurIPS 2025 · similarity 0.90)
- You Only Need Minimal RLVR Training: Extrapolating LLMs via Rank-1 Trajectories (NeurIPS 2026 · similarity 0.90)
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.
