Research map / Large language models / Instruction and fine-tuning
How Reasoning Evolves from Post-Training Data: An Empirical Study Using Chess
Lucas Dionisopoulos, Nicklas Majamaki, Prithviraj Ammanabrolu
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 RL after SFT on Best Move data | SFT6 [BM] · Predict Move | % legal moves: 44% -> 93% | Improved |
| add RL after SFT on Best Move - All data | SFT8 [BM - All] · Predict Move | % trivial moves: 3% -> 14% | Improved |
| add RL to Best Line SFT checkpoint | SFT7 [BL] · Predict Move | % trivial moves: 2% -> 64% | Hurt |
| add all data types to Best Move SFT mix (ablation) | SFT6 [BM] -> SFT8 [BM - All] · Predict Move | % trivial moves: 31% -> 14% | Improved |
| add all data types to Best Move SFT mix (ablation) | SFT6 [BM] -> SFT8 [BM - All] · Best Move task | accuracy after RL: 37% -> 41% | Improved |
| use multitask RL (all four tasks) instead of Predict Move only (ablation) | SFT2 [RSA] vs SFT1 [RSPM] · Predict Move | average normalized rank after RL: 0.52 -> 0.60 | Improved |
| add Verbalized Alpha-Beta Pruning SFT data instead of Best Move data (ablation) | SFT3 [VABP] vs SFT6 [BM] · Predict Move | % legal moves after RL: 93% -> 67% | Hurt |
| scale SFT tokens from 15M to 60M and RL samples from 8k to 16k (ablation) | SFT8 [BM - All] -> SFT8 XL · Predict Move | average normalized rank after RL: 0.71 -> 0.82 | Improved |
| chain SFT9 XL dataset after SFT8 XL checkpoint (ablation) | SFT8 XL · Legal Moves task | IoU after RL: 0.79 -> 0.87 | Improved |
| add RL to SFT checkpoint | SFT8 XL · Referenced move accuracy | % legal referenced moves: 48% -> 83% | Improved |
| RL on Best Move checkpoint rather than Best Line checkpoint (ablation) | SFT6 [BM] · reasoning faithfulness | faithfulness | Hurt |
| add RL to Qwen2.5-7B average reasoning quality | SFT8 XL · Average reasoning quality (gpt-oss-120b judge) | score: 5.0 -> 2.2 | Hurt |
Closest papers by meaning
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models (ICML 2026 · similarity 0.91)
- RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs (ICLR 2026 · similarity 0.90)
- Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection (ICLR 2026 · similarity 0.90)
- ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models (NeurIPS 2025 · similarity 0.90)
- Quagmires in SFT-RL Post-Training: When High SFT Scores Mislead and What to Use Instead (ICLR 2026 · similarity 0.89)
- Beyond Accuracy: Dissecting Mathematical Reasoning for LLMs Under Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
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.
