AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Hybrid Reinforcement: when reward is sparse, better to be dense

Leitian Tao, Ilia Kulikov, Swarnadeep Saha, Tianlu Wang, Jing Xu, Sharon Li, Jason E Weston, Ping Yu

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Scaling RL with verifiable rewards

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
stratified hybrid reward (HERO) vs RM-onlyQwen3-4B-Base · hard-to-verify eval, trained on easyavg accuracy: 54.6 -> 66.3Improved
stratified hybrid reward (HERO) vs verifier-onlyQwen3-4B-Base · hard-to-verify eval, trained on easyavg accuracy: 57.1 -> 66.3Improved
HERO vs verifier-only on hard-to-verify training dataQwen3-4B-Base · easy-to-verify eval, trained on hardavg pass@1: 47.4 -> 56.8Improved
verifier-only training on hard-to-verify data vs SFT cold startQwen3-4B-Base · hard-to-verify avgavg accuracy: 47.1 -> 47.4Hurt
HERO vs baselinesOctoThinker-8B-Hybrid-Base · easy and hard avg, trained on hardavg pass@1: 34.6 -> 41.0Improved
replace naive reward combination with stratified normalization (ablation)HERO · Easy-to-verify avgavg pass@1: 58.7 -> 62.0Improved
replace naive reward combination with stratified normalization (ablation)HERO · HardVerify-Mathaccuracy: 61.4 -> 73.2Improved
add variance-aware reweighting (ablation)HERO · HardVerify-Mathaccuracy: 69.4 -> 73.2Improved
add variance-aware reweighting (ablation)HERO · Easy-to-verify avgavg pass@1: 60.8 -> 62.0Improved
scale reward model from 7B to 72B (ablation)HERO · Easy-to-verify avgavg pass@1: 62.0 -> 62.8No clear effect
scale reward model from 7B to 72B (ablation)HERO · HardVerify-Mathaccuracy: 73.2 -> 71.4Hurt
use dense reward on negative group (keep sparse positives) (ablation)HERO · verifiable and hard-to-verify tasksaccuracy: 59.4 -> 61.4 (verifiable); 62.2 -> 68.4 (hard)Improved

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