AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Learning More with Less: A Dynamic Dual-Level Down-Sampling Framework for Efficient Policy Optimization

Chao Wang, Tao Yang, Hongtao Tian, Yunsheng Shi, Qiyao Ma, XiaotaoLiu, Ting Yao, Wenbo Ding

ICLR 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants

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 sample-level advantage-variance down-sampling (D1S) (ablation)GRPO · 7 math benchmarksPass@8 average: 53.1 -> 54.2Improved
add cross-group selection to D1S (ablation)GRPO · 7 math benchmarks (Qwen2.5-Math-7B)Pass@1 average: 31.3 -> 34.1Improved
add token-level entropy-advantage selection (D2S) without schedule (ablation)GRPO · 7 math benchmarks (Qwen2.5-Math-7B)Pass@8 average: 54.2 -> 54.1No clear effect
add dynamic down-sampling schedule (D3S vs D2S) (ablation)GRPO · 7 math benchmarks (Qwen2.5-Math-7B)Pass@8 average: 54.1 -> 56.8Improved
add D3S to GRPOQwen2.5-Math-7B · 7 math benchmarksPass@1 average: 29.8 -> 34.3Improved
add D3S to GRPOLlama3.1-8B-Instruct · 7 math benchmarksPass@8 average: 32.4 -> 40.2Improved
add D3S to GSPOQwen2.5-Math-7B · 7 math benchmarksPass@1 average: 33.2 -> 35.8Improved
add D3S to GSPOQwen2.5-Math-1.5B · 7 math benchmarksPass@1 average: 33.3 -> 33.3No clear effect
replace PODS reward-variance selection with D3SGRPO · 7 math benchmarks (Qwen2.5-Math-7B)Pass@8 average: 54.4 -> 56.8Improved
remove cross-group selection (D3S-I) on unaligned Llama (ablation)D3S · Llama3.1-8B-Instruct, 7 math benchmarksPass@1 average: 23.8 -> 24.0Improved
add cross-group selection (D1S-C vs D1S) on unaligned Llama (ablation)GRPO · Llama3.1-8B-Instruct, 7 math benchmarksPass@8 average: 40.6 -> 33.3Hurt
add D3S to GRPO on domain-aligned OpenMath2-LlamaGRPO · OpenMath2-Llama3.1-8BPass@1 average: 28.2 -> 28.7Mixed

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