AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

RL for Reasoning by Adaptively Revealing Rationales

Mohammad Hossein Amani, Aryo Lotfi, Nicolas Baldwin, Samy Bengio, Mehrdad Farajtabar, Emmanuel Abbe, Robert West

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
Add adaptive backtracking (AdaBack) on top of SFT+GRPOLlama-3 1B · GSM8ktest accuracy: 36.7 -> 43.2Improved
Apply AdaBack directly to the base model instead of GRPO on baseLlama-3 3B · Tensor-2 GSM8ktest accuracy: 0.0 -> 49.2Improved
Use AdaBack on base model instead of SFT+RLLlama-3 1B · Base-7 GSM8ktest accuracy: 14.4 -> 18.4Improved
Add SFT before AdaBack on 3B (ablation)Llama-3 3B · Tensor-2 GSM8ktest accuracy: 49.2 -> 42.2Hurt
Add AdaBack on top of SFT+RL at 3BLlama-3 3B · GSM8ktest accuracy: 72.7 -> 70.7Hurt
Use AdaBack instead of R3 curriculum (ablation)Llama-3 base · DeepScaleRtest accuracy: 1B: 6.6 -> 9.0; 3B: 9.6 -> 10.6Improved
Use AdaBack instead of R3 (ablation)Llama-3 base · GSM8ktest accuracy: 1B: 41.5 -> 39.2; 3B: 74.2 -> 73.3Hurt
Use AdaBack instead of R3 (ablation)Llama-3 base · MATHtest accuracy: 1B: 7.8 -> 9.1; 3B: 19.2 -> 19.1Mixed
Add AdaBack to RL on instruct-tuned or contaminated modelsLlama 3.2 3B-Instruct, Qwen2.5-1.5B · MATH, GSM8kreward/accuracyNo clear effect
Use AdaBack instead of SFT+RL on chain-of-paritiesLlama 3.2 1B · synthetic chain-of-parities (L=16, n=1024)reward: SFT+RL plateau at 0.1; AdaBack learns in under 700 iterationsImproved
Vary reward threshold tau (ablation)AdaBack · GSM8ktraining reward and test accuracyNo clear effect
Use AdaBack instead of standard RL for pass@kLlama3-1B · GSM8k pass@kpass@kImproved

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