Research map / Reinforcement learning / Imitation learning
Escaping the Verifier: Learning to Reason via Demonstrations
Locke Cai, Max Ryabinin, Ivan Provilkov
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add critic reasoning (ablation) | RARO · DeepMath 1.5B | accuracy: 35.9 -> 41.3 | Improved |
| add relativistic (pairwise) critic (ablation) | RARO · DeepMath 1.5B | accuracy: 36.9 -> 41.3 | Improved |
| add tie option for the critic (ablation) | RARO · DeepMath 1.5B | accuracy: 38.6 -> 41.3 | Improved |
| add replay buffer (ablation) | RARO · DeepMath 1.5B | accuracy: 35.4 -> 41.3 | Improved |
| share one LLM for critic and policy (ablation) | RARO · DeepMath 1.5B | accuracy: 39.4 -> 41.3 | Improved |
| vary critic tie reward from 0.52 to 0.60 (ablation) | RARO · DeepMath 1.5B | validation accuracy: 41.1, 42.0, 40.7 | No clear effect |
| vary policy tie coefficient from 0.20 to 0.60 (ablation) | RARO · DeepMath 1.5B | validation accuracy: 40.8, 40.2, 42.0 | No clear effect |
| vary policy-loss weight 4/5, 2/3, 1/3 (ablation) | RARO · Poetry 1.5B | validation reward: 64.0, 68.8, 66.2 | No clear effect |
| train with RARO instead of SFT on demonstrations | Qwen2.5-1.5B-Instruct · Countdown | accuracy: 40.7 -> 54.4 | Improved |
| train with RARO instead of RL-Logit | Qwen2.5-7B-Instruct · DeepMath | accuracy: 49.3 -> 57.5 | Improved |
| use RARO instead of SFT | Qwen2.5-7B-Instruct · Poetry Writing | win rate: 5.9 -> 25.0 | Improved |
| use RARO instead of RL-Logit on poetry | Qwen2.5-1.5B-Instruct · Poetry Writing | score: 35.0 -> 36.4 | No clear effect |
Closest papers by meaning
- Generative Adversarial Reasoner: Enhancing LLM Reasoning with Adversarial Reinforcement Learning (ICLR 2026 · similarity 0.89)
- Reinforcing General Reasoning Without Verifiers (ICLR 2026 · similarity 0.89)
- $\textbf{Re}^{2}$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving (ICLR 2026 · similarity 0.88)
- Probing to Refine: Reinforcement Distillation of LLM Reasoners via Explanatory Inversion (ICLR 2026 · similarity 0.88)
- The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning (NeurIPS 2025 · similarity 0.88)
- RL Tango: Reinforcing Generator and Verifier Together for Language Reasoning (NeurIPS 2025 · similarity 0.88)
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.
