Research map / Large language models / AI and society
Human-Aligned Chess With a Bit of Search
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Trains a GPT-2-initialized Transformer on Lichess blitz game sequences to predict human moves, pondering time, and outcomes, then uses time-adaptive MCTS for skill-calibrated, human-aligned chess play.
Problem
Chess AI engines are strong but not human-aligned: they cannot match the skill levels of human partners or model human behaviors such as move choice, pondering time, and resignation.
Approach
Model chess as a language over UCI move tokens with soft Elo control tokens, with heads for next-move policy, human pondering time, and game-outcome value; then run MCTS with a number of rollouts proportional to predicted human pondering time.
Key results
- Allie-Policy reaches 55.7% human move-matching accuracy vs 51.6% for Maia* and 53.7% for GPT-3.5.
- Allie-Adaptive-Search achieves mean skill calibration error of 49 Elo (max 95) against humans from 1100 to 2500 Elo, versus 146 for Maia* and 127 for Allie-Policy.
- Allie resigns in 86.4% of positions where humans resign, with a 0.1% false positive rate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add time-adaptive MCTS to Allie-Policy | Allie · Online games against humans across skill levels | mean skill calibration error (Elo): 127 -> 49 | Helped |
| Use time-adaptive rather than fixed-budget MCTS | Allie-Search · Online performance vs. 2500-rated humans | performance Elo: 2318 -> 2528 | Helped |
| Add time-adaptive MCTS to the move-prediction policy | Allie-Policy · Lichess move prediction, all moves | move-matching accuracy: 55.7 -> 55.9 | Helped |
| Train on 50% of the dataset (ablation) | Allie · Move-matching accuracy across skill levels | move-matching accuracy: -1.0% | Hurt |
| Train on the full dataset for 50% of the steps (ablation) | Allie · Move-matching accuracy across skill levels | move-matching accuracy | No clear effect |
| Reduce model size to 124M parameters (ablation) | Allie · Move-matching accuracy across skill levels | move-matching accuracy: -1.2% | Hurt |
| Increase model size to 774M parameters (ablation) | Allie · Move-matching accuracy across skill levels | move-matching accuracy: +0.3% | No clear effect |
Surprising findings
- Halving the training data causes overfitting only on the value prediction objective, while doubling model size gives only a 0.3% accuracy gain, suggesting the setup is data-constrained.
Limitations
- Training data is dominated by fast blitz games; slower time controls are not modeled.
- Players reported late-game blunders and long pondering in forced positions.
- Pondering-time predictions skew lower than human times.
Experimental setup
Datasets: Lichess blitz games 2022 randomly generated chess games
Benchmarks: Lichess blitz move-matching evaluation set random-game legal move test large-scale online games vs. humans (7,483 games, 2,412 players)
Metrics: top-1 move-matching accuracy top move validity pondering time Pearson correlation resignation TPR/FPR skill calibration error (Elo) performance Elo
Baselines: Maia GPT-3.5-turbo-instruct Allie-Search (non-adaptive MCTS) Allie-Greedy Allie-Policy
Builds on: AlphaZero GPT-2 Maia
Closest papers by meaning
- Maia-2: A Unified Model for Human-AI Alignment in Chess (NeurIPS 2024 · similarity 0.94)
- Designing Skill-Compatible AI: Methodologies and Frameworks in Chess (ICLR 2024 · similarity 0.91)
- Amortized Planning with Large-Scale Transformers: A Case Study on Chess (NeurIPS 2024 · similarity 0.88)
- Strength Estimation and Human-Like Strength Adjustment in Games (ICLR 2025 · similarity 0.87)
- Mastering Board Games by External and Internal Planning with Language Models (ICML 2025 · similarity 0.87)
- The Algorithm Is Not the Behavior: Learned Priors Override Look-Ahead in a Chess-Playing Neural Network (NeurIPS 2026 · similarity 0.87)
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.