AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Compute as Teacher: Turning Inference Compute Into Reference-Free Supervision

Dulhan Jayalath, Shashwat Goel, Thomas Foster, Parag Jain, Suchin Gururangan, Cheng Zhang, Anirudh Goyal, Alan Schelten

ICML 2026 · Reinforcement learning › RL from feedback · approach: Scaling RL with verifiable rewards · 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
train with CaT (RL on synthesized pseudo-reference)Llama 3.1 8B · HealthBenchrubric score: 0.29 -> 0.38Improved
train with CaTLlama 3.1 8B · MATH-500accuracy: 0.42 -> 0.56Improved
train with CaT instead of using inference-time synthesisQwen 3 4B · HealthBenchrubric score: 0.42 -> 0.43Improved
train with CaT instead of using inference-time synthesisQwen 3 4B · MATH-500accuracy: 0.81 -> 0.83Improved
use RL with rubric rewards instead of SFT on pseudo-references (ablation)CaT, Llama 3.1 8B · HealthBenchrubric score: 0.28 -> 0.38Improved
use self-proposed rubrics instead of model-as-judge reward (ablation)CaT, Llama 3.1 8B · HealthBenchrubric score: 0.18 -> 0.38Improved
use physician rubrics instead of self-proposed rubrics (ablation)CaT, Llama 3.1 8B · HealthBenchrubric score: 0.38 -> 0.40No clear effect
condition rubric generation on pseudo-reference (ablation)CaT, Llama 3.1 8B · HealthBenchrubric score: 0.34 -> 0.38Improved
use synthesis instead of Self-BoN for reference estimationQwen 3 4B · HealthBench (inference)rubric score: 0.40 -> 0.44Improved
use synthesis instead of majority voteQwen 3 4B · MATH-500 (inference)accuracy: 0.81 -> 0.86Improved
use synthesis instead of Self-BoNLlama 3.1 8B · MATH-500 (inference)accuracy: 0.54 -> 0.53No clear effect
condition synthesis on 8 rollouts instead of 1 (ablation)synthesis, Qwen 3 4B · MATH-500accuracy: 0.79 -> 0.85Improved

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