AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Learning to Reason as Action Abstractions with Scalable Mid-Training RL

Shenao Zhang, Donghan Yu, Yihao Feng, Bowen Jin, Zhaoran Wang, John Peebles, Zirui Wang

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Scaling RL with verifiable rewards · 3 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
mid-train with RA3 instead of next-token predictionLlama-3.2-1B · code benchmarks avgpass@1: 25.3 -> 29.8Improved
mid-train with RA3 instead of next-token predictionQwen2.5-1.5B · code benchmarks avgpass@1: 41.7 -> 46.6Improved
mid-train with RA3 instead of next-token predictionLlama-3.1-8B · code benchmarks avgpass@1: 47.7 -> 48.9Improved
mid-train with RA3 instead of next-token predictionLlama-3.2-1B · code benchmarks avgpass@5: 40.3 -> 43.1Improved
mid-train with RA3 instead of next-token predictionLlama-3.1-8B · MBPPpass@1: 48.6 -> 48.0Hurt
mid-train with RA3 instead of next-token predictionLlama-3.1-8B · MBPP+pass@1: 51.1 -> 53.2Improved
add mid-training (NTP) before RLVRbase models · RLVR post-trainingperformanceImproved
mid-train with RA3 instead of NTP before RLVRRLVR with GRPO · RLVR post-trainingconvergence speed and asymptotic performanceImproved
use small KL penalty c in latent policy (ablation)RA3 on Qwen · Qwen mid-trainingredundant reasoning at every stepHurt
use large KL penalty c=0.2 (ablation)RA3 on Qwen · Qwen mid-traininglatent behavior: policy outputs only act, RA3 degenerates to NTPHurt
use default penalty c=0.05 (ablation)RA3 on Qwen · Qwen mid-trainingreasoning frequency: reasoning at fewer than 40% of linesImproved
fine-tune on reasoning-bootstrapped data instead of raw data (ablation)RA3 M step · mid-trainingcross-entropy lossImproved

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