AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Reinforcement Learning via Self-Distillation

Jonas Hübotter, Frederike Lübeck, Lejs Deen Behric, Anton Baumann, Marco Bagatella, Daniel Marta, Ido Hakimi, Idan Shenfeld, Thomas Kleine Buening, Carlos Guestrin, Andreas Krause

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
replace GRPO scalar advantages with SDPO self-teacher advantagesRLVR post-training · LiveCodeBench v6, Qwen3-8Baccuracy (%): 41.2 -> 48.8Improved
use SDPO instead of GRPO with scalar rewards onlyQwen3-8B and Olmo3-7B-Instruct · science Q&A and tool use (aggregate)final accuracy (%): 64.1 -> 68.8Improved
use SDPO instead of GRPO for generation lengthscience Q&A, Olmo3-7B-Instruct Chemistry · Chemistryresponse length: 7x reductionImproved
use logit-level instead of token-level or sequence-level SDPO credit assignment (ablation)SDPO · LiveCodeBench v6, Qwen3-8BaccuracyImproved
sequence-level SDPO (rich feedback without dense credit) versus GRPO (ablation)SDPO · LiveCodeBench v6, Qwen3-8BaccuracyImproved
scale model size from small to large Qwen3 (ablation)SDPO versus GRPO · LiveCodeBench v6, Qwen3 0.6B to 8BaccuracyImproved
use SDPO on the weakest model Qwen2.5-1.5B (ablation)SDPO · LiveCodeBench v6, Qwen2.5-1.5BaccuracyHurt
add environment output to own-solution feedback for the self-teacher (ablation)SDPO · LiveCodeBench v6, Qwen3-8Bstudent accuracy (%): 42.6 -> 48.3Improved
include the student's original attempt in the teacher prompt (ablation)SDPO · LiveCodeBench v6, Qwen3-8Bstudent accuracy (%): 48.3 -> 44.5Hurt
use a regularized teacher (trust-region or EMA) instead of unregularized current teacher (ablation)SDPO · LiveCodeBench v6, Qwen3-8Bbest accuracy (%): 36.1 -> 50.6 (trust-region)Improved
update the teacher via trust-region or EMA instead of freezing it (ablation)SDPO · LiveCodeBench v6, Qwen3-8Bbest accuracy (%): 48.8 -> 50.6 (trust-region)Improved
add GRPO advantages to SDPO (lambda 0.9) (ablation)SDPO · LiveCodeBench v6, Qwen3 sizesaccuracyMixed

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