Research map / Architectures and efficiency / Attention and transformers
Chessformer: A Unified Architecture for Chess Modeling
Daniel Monroe, George Eilender, Philip Chalmers, Zhenwei Tang, Ashton Anderson
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 |
|---|---|---|---|
| replace absolute position embedding with GAB at ~4M params (ablation) | Chessformer (Leela distillation) · Leela distillation ablation (searchless) | Elo: 0 -> 83 | Improved |
| replace absolute position embedding with GAB (ablation) | Chessformer (Leela distillation) · Leela distillation ablation (searchless) | policy accuracy (%): 56.6 -> 58.5 | Improved |
| replace absolute position embedding with GAB (ablation) | Chessformer (Leela distillation) · Leela distillation ablation (searchless) | puzzle accuracy (%): 61.0 -> 64.2 | Improved |
| replace absolute position embedding with GAB (ablation) | Chessformer (Leela distillation) · Leela distillation ablation (searchless) | value accuracy (%): 88.7 -> 89.0 | Improved |
| replace relative position bias with GAB (ablation) | Chessformer (Leela distillation) · Leela distillation ablation (searchless) | Elo: 40 -> 83 | Improved |
| use GAB at 2.5M params versus absolute encoding at 3.67M (ablation) | Chessformer (Leela distillation) · Leela distillation ablation (searchless) | Elo: 0 -> -4 | No clear effect |
| replace absolute position embedding with GAB (ablation) | Maia-3 (5M) · Human move matching (5M scale) | policy accuracy (%): 54.7 -> 55.4 | Improved |
| replace relative position bias with GAB (ablation) | Maia-3 (5M) · Human move matching (5M scale) | policy accuracy (%): 54.6 -> 55.4 | Improved |
| use 3M GAB model versus 4.58M absolute-encoding model (ablation) | Maia-3 · Human move matching (5M scale) | policy accuracy (%): 54.7 -> 54.8 | No clear effect |
| add 7 past board positions as input (ablation) | Maia-3-5M · Human move matching, history ablation | move-matching accuracy (%): 54.0 -> 55.4 | Improved |
| increase history from 7 to 31 past positions (ablation) | Maia-3-5M · Human move matching, history ablation | move-matching accuracy (%): 55.4 -> 55.4 | No clear effect |
| replace CNN with Chessformer (Leela-CF) in Leela Chess Zero | Leela Chess Zero · Full engine vs Leela-CNN | Elo: +112 / +111 / +105 Elo at 160k / 320k / 640k | Improved |
Closest papers by meaning
- Amortized Planning with Large-Scale Transformers: A Case Study on Chess (NeurIPS 2024 · similarity 0.91)
- Maia-2: A Unified Model for Human-AI Alignment in Chess (NeurIPS 2024 · similarity 0.91)
- Human-Aligned Chess With a Bit of Search (ICLR 2025 · similarity 0.90)
- Chess-World-Model: A 10M-Game Benchmark for Exact State Tracking from Chess Move Sequences (NeurIPS 2026 · similarity 0.90)
- Designing Skill-Compatible AI: Methodologies and Frameworks in Chess (ICLR 2024 · similarity 0.88)
- Enhancing Chess Reinforcement Learning with Graph Representation (NeurIPS 2024 · 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.
