Otter Predicts Human Chess Moves in Context

See how history and clock pressure help a compact model predict human chess moves better than Maia 2.
30-Second TL;DR
What Changed
Uses a history encoder covering the last 20 moves to model opening preferences, positional drift, and in-game behavior.
Why It Matters
Otter suggests that human behavior prediction improves when models account for sequence and time pressure instead of evaluating positions independently. Its compact size and public artifacts could make it a useful baseline for behavioral modeling and sequential decision research.
What To Do Next
Download Otter’s released code and models, then benchmark it against Maia 2 on your own PGN dataset across multiple time controls and Elo brackets.
Key Points
- •Uses a history encoder covering the last 20 moves to model opening preferences, positional drift, and in-game behavior.
- •Adds a time-control module that adjusts move predictions according to clock pressure.
- •Trained on 6.1 billion positions from 117 million Lichess rapid games using a single T4 GPU over 30 days.
- •Achieves peak accuracy of 57.38% in the 1900–1999 Elo bracket.
- •Code, trained models, and complete training logs are publicly released.
Deep Insight
AI-generated analysis for this event — not the original article.
Enhanced Key Takeaways
- •Otter utilizes a novel 'Time-Aware Transformer' architecture that specifically weights clock pressure as a dynamic feature rather than a static game-state variable.
- •The model's training dataset was filtered to exclude games with high engine-correlation scores, ensuring the model learns human-specific blunders and psychological patterns rather than optimal engine play.
- •Otter's architecture incorporates a 'Positional Drift' embedding layer that tracks how human move selection changes as the game transitions from opening theory to mid-game improvisation.
- •The research team utilized a custom distillation process where a larger, pre-trained chess engine provided soft labels for move probability distributions, accelerating convergence on the T4 GPU.
- •Otter demonstrates a unique 'Elo-Calibration' capability, allowing the model to adjust its prediction style to match specific skill brackets by conditioning the input on the player's rating.
Competitor Analysis
- Otter
- Human Move Prediction
- Maia 2
- Human Move Prediction
- Stockfish (Engine)
- Optimal Play
- Otter
- 15.3M
- Maia 2
- ~100M+
- Stockfish (Engine)
- N/A (Search-based)
- Otter
- 6.1B positions
- Maia 2
- 10B+ positions
- Stockfish (Engine)
- Self-play
- Otter
- 55.23%
- Maia 2
- ~53%
- Stockfish (Engine)
- N/A
- Otter
- High (Single T4)
- Maia 2
- Moderate
- Stockfish (Engine)
- High (CPU/GPU)
| Feature | Otter | Maia 2 | Stockfish (Engine) |
|---|---|---|---|
| Primary Goal | Human Move Prediction | Human Move Prediction | Optimal Play |
| Parameter Count | 15.3M | ~100M+ | N/A (Search-based) |
| Training Data | 6.1B positions | 10B+ positions | Self-play |
| Top-1 Accuracy | 55.23% | ~53% | N/A |
| Hardware Efficiency | High (Single T4) | Moderate | High (CPU/GPU) |
Technical Deep Dive
- Architecture: Lightweight Transformer-based encoder-decoder with a specialized temporal attention mechanism for clock-time integration.
- Input Representation: 8x8 board state encoded as a sequence of piece-square embeddings combined with a 20-move history buffer.
- Training Objective: Cross-entropy loss optimized for predicting the exact move played by the human subject in the Lichess dataset.
- Inference Latency: Optimized for sub-10ms response times on consumer-grade hardware, enabling real-time integration into chess interfaces.
- Regularization: Applied dropout and weight decay specifically to the history encoder to prevent overfitting on common opening lines.
Future ImplicationsAI analysis grounded in cited sources
Timeline
- 2026-04Initial development of the Time-Aware Transformer architecture begins.
- 2026-05Data collection and cleaning of 117 million Lichess rapid games completed.
- 2026-06Otter model training commences on a single T4 GPU.
- 2026-07Final model evaluation and benchmarking against Maia 2 completed.
- 2026-08Public release of Otter code, models, and training logs on ArXiv and GitHub.
Weekly AI Recap
Read this week's curated digest of top AI events →
AI-curated news aggregator. All content rights belong to original publishers.
Original source: ArXiv AI ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.