Research map / Architectures and efficiency / Attention and transformers
Evidence of Learned Look-Ahead in a Chess-Playing Neural Network
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.
Shows that Leela Chess Zero's policy network internally represents future optimal moves via activation patching, attention-head analysis, and bilinear probes, giving evidence of learned look-ahead.
Problem
Whether neural networks learn look-ahead or search algorithms "in the wild" rather than only collections of simple heuristics, studied in the policy network of a chess engine.
Approach
Exploits Leela's transformer structure (one token per chessboard square) to localize future-move information via activation patching on automatically generated corruptions, identify attention heads that move information backward and forward in time, and train a bilinear probe that reads the optimal move two turns ahead.
Key results
- Patching the third-move target square at layer 10 reduces log odds of the correct move by 1.88 ± 0.04, versus 0.55 ± 0.01 for the maximum over other squares.
- Ablating a single attention entry in L12H12 (query first-move target, key third-move target) reduces log odds of the correct move by more than 1.5 in over 10% of puzzles.
- A bilinear probe predicts the third move's target square with (92 ± 1)% accuracy after layer 12, versus 15 ± 2% for a randomly initialized Leela.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Predict the third-move target square from Leela activations using a bilinear probe | Leela T82-768x15x24h · Filtered Lichess puzzles, test split | accuracy: (15 ± 2)% -> (92 ± 1)% | Helped |
| Patch corrupted-state activations at the third-move target square in layer 10 (ablation) | Leela T82-768x15x24h · Filtered Lichess puzzles | mean log-odds reduction: $1.88 \pm 0.04$ | Hurt |
| Patch the largest-effect other square in layer 10 (ablation) | Leela T82-768x15x24h · Filtered Lichess puzzles | mean maximum log-odds reduction: $0.55 \pm 0.01$ | Hurt |
| Zero the L12H12 attention entry from the third-move target to the first-move target (ablation) | Leela T82-768x15x24h · Filtered Lichess puzzles | log-odds reduction: more than 1.5 in more than 10% of puzzles | Hurt |
| Zero all 4095 other L12H12 attention entries at once instead of the third-to-first target entry (ablation) | Leela T82-768x15x24h · Filtered Lichess puzzles | log-odds reduction | No clear effect |
| Zero attention entries out of the third-move target in matching piece-movement heads (ablation) | Leela T82-768x15x24h · Puzzles with principal variation longer than 3 moves | top-move log-odds reduction: at least 1.5 in 60% of puzzles | Hurt |
| Ablate piece-movement heads for other piece types instead of the third-move piece type (ablation) | Leela T82-768x15x24h · Puzzles with principal variation longer than 3 moves | model performance | No clear effect |
| Ablate information flow on a random square instead of the third-move target (ablation) | Leela T82-768x15x24h · Puzzles with principal variation longer than 3 moves | model performance | No clear effect |
| Finetune Leela to behave the same with or without past board-state history, then omit history | Leela T82-768x15x24h · Puzzle solving and games against the original model | playing strength | No clear effect |
| Use different past-board histories (zeros, repeated current state, or synthesized history) (ablation) | Original Leela T82-768x15x24h · Board-state inputs | policy output | Mixed |
Surprising findings
- Second-move squares (opponent response) are not similarly important despite the authors' expectation, and effects are much stronger on puzzles where first and second move targets coincide (83% of data).
- Zeroing one attention entry in L12H12 has a larger effect than ablating all 4095 other L12H12 attention weights together.
Limitations
- No precise description of how look-ahead is implemented; only single-line look-ahead is tested, not comparison of multiple lines (search).
- Analysis focuses on unusually complex puzzle states, not the full input distribution.
- Results are on puzzles where the dataset's filtering favors sacrifices (83% overlapping first and second move targets).
Experimental setup
Datasets: Lichess puzzles dataset (900k, filtered to 22.5k)
Benchmarks: filtered Lichess puzzles (22.5k)
Metrics: change in log odds of best move under activation patching probe accuracy for predicting third-move target square log odds reduction from attention ablation
Baselines: Little Demon 2 randomly initialized Leela probe baseline ruoss2024grandmasterlevel network
Builds on: AlphaZero activation patching Leela Chess Zero
Closest papers by meaning
- The Algorithm Is Not the Behavior: Learned Priors Override Look-Ahead in a Chess-Playing Neural Network (NeurIPS 2026 · similarity 0.92)
- Amortized Planning with Large-Scale Transformers: A Case Study on Chess (NeurIPS 2024 · similarity 0.88)
- Enhancing Chess Reinforcement Learning with Graph Representation (NeurIPS 2024 · similarity 0.86)
- Implicit Search via Discrete Diffusion: A Study on Chess (ICLR 2025 · similarity 0.86)
- Chess-World-Model: A 10M-Game Benchmark for Exact State Tracking from Chess Move Sequences (NeurIPS 2026 · similarity 0.86)
- Human-Aligned Chess With a Bit of Search (ICLR 2025 · similarity 0.85)
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.