Research map / Large language models / AI and society
Maia-2: A Unified Model for Human-AI Alignment in Chess
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.
Proposes Maia-2, a unified single-model human chess move predictor that conditions on active and opponent skill levels via skill-aware attention over channel-wise patches, improving accuracy and cross-skill coherence over the per-skill Maia-1 models.
Problem
Existing human-like chess models (Maia-1) train independent models per skill level, producing volatile, incoherent predictions across skill levels and ignoring opponent skill, which limits their use as coherent AI partners and teaching tools.
Approach
A single unified network embeds active and opponent skill levels as categorical embeddings, injects them into queries of skill-aware multi-head self-attention over channel-wise patches of the ResNet-encoded board, uses only the current position, and adds an auxiliary head (legal moves, move attributes) alongside policy and value heads.
Key results
- Maia-2 achieves 53.25% average top-1 move prediction accuracy on the Maia Testset versus 51.39% for the best Maia-1 model (Maia 1500), with the largest gain on Advanced players (+1.5 points).
- Average perplexity on the Grounded Testset drops from 4.67 bits (Maia 1900) to 4.07 bits for Maia-2.
- 27% of positions are treated monotonically by Maia-2 versus 1% by Maia-1 on the Grounded Testset; a randomized Lichess study showed rematch rate 41.2% for Maia-2 vs 40.3% for Maia-1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use unified Maia-2 instead of separate Maia models | human move prediction · Maia-2 Testset, Avg | top-1 accuracy: 52.74 -> 53.25 | Helped |
| use unified Maia-2 instead of Maia 1900 | human move prediction · Grounded Testset, Avg | perplexity: 4.67 -> 4.07 | Helped |
| add skill-aware attention (ablation) | Maia-2_subset · Maia Testset, Avg | top-1 accuracy: 51.75 -> 52.74 | Helped |
| add auxiliary information head (ablation) | Maia-2_subset · Maia Testset, Avg | top-1 accuracy: 52.09 -> 52.74 | Helped |
| use unified Maia-2 instead of separate Maia models | human move prediction · Grounded Testset, Skilled | perplexity: 4.44 -> 4.30 | Helped |
| use unified Maia-2 instead of separate Maia models | human move prediction · Grounded Testset, Advanced | perplexity: 4.71 -> 3.90 | Helped |
| use unified Maia-2 instead of separate Maia models | human move prediction · Grounded Testset, Master | perplexity: 4.86 -> 4.02 | Helped |
| use unified Maia-2 instead of Maia-1 | cross-skill move prediction · Cross-skill Testset | prediction accuracy | Helped |
| use unified Maia-2 instead of Maia-1 | prediction coherence · Grounded Testset | percentage monotonic: 1% -> 27% | Helped |
| use unified Maia-2 instead of Maia-1 | prediction coherence · Grounded Testset | percentage transitional: 17% -> 22% | Helped |
| play against Maia-2 instead of Maia-1 | chess bots in randomized Lichess human study · Lichess human study | rematch rate: 40.3% -> 41.2% | Helped |
| increase active-player skill while holding opponent skill fixed | Maia-2 · Cross-skill Testset | move decision change rate: up to 22% | Mixed |
Surprising findings
- Maia-2 (even the subset variant trained on Maia-1 data) outperforms Maia-1 with about 73% of the gain attributed to architecture and 27% to extra data.
- Maia-2 can predict moves with only the current board position, without the six-position history used by Maia-1.
Limitations
- Maia-2 does not incorporate search.
- The strongest players are grouped into a single bucket, so very top-level play is not modeled well.
- Coherence is encouraged by architecture but not enforced, so volatility can still occur.
Experimental setup
Datasets: Lichess rapid games (Jan 2013 - Nov 2023, excluding Dec 2019) Maia Testset (Dec 2019) Cross-skill Testset (Dec 2023) Grounded Testset (450k positions with Stockfish evaluations)
Benchmarks: Maia Testset Cross-skill Testset Grounded Testset
Metrics: top-1 move prediction accuracy perplexity monotonic and transitional percentage centipawn loss blunder rate linear probe r2 / macro F1
Baselines: Maia-1 (1100, 1500, 1900) Stockfish Leela Maia-2 w/o Att Maia-2 w/o Aux
Builds on: Maia AlphaZero Vision Transformer
Closest papers by meaning
- Human-Aligned Chess With a Bit of Search (ICLR 2025 · similarity 0.94)
- Designing Skill-Compatible AI: Methodologies and Frameworks in Chess (ICLR 2024 · similarity 0.91)
- An Evolutionary Perspective on AI Alignment (Student Abstract) (AAAI 2025 · similarity 0.87)
- Reason to Play: Behavioral and Brain Alignment Between Frontier LRMs and Human Game Learners (NeurIPS 2026 · similarity 0.87)
- Strength Estimation and Human-Like Strength Adjustment in Games (ICLR 2025 · similarity 0.87)
- Position: Towards Bidirectional Human-AI Alignment (NeurIPS 2025 · similarity 0.86)
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.