Research map / Large language models / AI and society
Generating Creative Chess Puzzles
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.
Generates creative chess puzzles by benchmarking generative architectures trained on Lichess puzzles, then RL-fine-tuning an autoregressive transformer with engine-search-based uniqueness and counter-intuitiveness rewards plus diversity filtering.
Problem
Generative AI struggles to produce creative, aesthetic, and counter-intuitive outputs; chess puzzles that are unique, counter-intuitive, and novel are rare (about 2.1% of Lichess unique puzzles) and no standard definition of chess puzzle creativity exists.
Approach
Defines measurable proxies for chess puzzle creativity (uniqueness, counter-intuitiveness via shallow-vs-deep engine search statistics, novelty via board/PV distance and entropy, aesthetic themes), uses them as RL outcome rewards with KL, data-mixing and piece regularization, and a diversity filter that prevents entropy collapse.
Key results
- RL raised the unique + counter-intuitive puzzle rate from 0.22% (supervised transformer) to 2.5%, exceeding the best Lichess-trained model (0.4%) and the Lichess baseline (2.14%/2.1%).
- Counter-intuitiveness score (Stockfish critical point + capture material) reached AP 0.7401 on TEST versus 0.5487 for search features alone and 0.4768 for uniqueness alone.
- In an 8-expert rating study, Booklet puzzles scored highest on creativity (2.48), fun (2.56) and counter-intuitiveness (2.12), above book puzzles and Lichess puzzles (1.7 creativity).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply RL with uniqueness and counter-intuitiveness rewards to the Lichess-trained Transformer | Transformer · Generated chess puzzles | Counter-intuitive puzzle rate: 0.22% to 2.5% | Helped |
| Apply RL with diversity filtering | Transformer · Generated chess puzzles | Qualified puzzle rate: 2.5% vs 2.1% | Helped |
| Add Stockfish search-through-time features to search features | Counter-intuitiveness score function · Golden Set TEST | Average precision: 0.5487 -> 0.7401 | Helped |
| Add AlphaZero search-through-time features to Search Features + SF through time | Counter-intuitiveness score function · Golden Set TEST | Average precision: 0.7401 -> 0.7360 | Hurt |
| Use RL (auto-selected) puzzles instead of Lichess puzzles | Chess puzzles · Expert study | Creative: 1.7 $\pm$ 0.2 -> 2.1 $\pm$ 0.37 | Helped |
| Curate generated puzzles into the Booklet instead of using Book puzzles | Chess puzzles · Expert study | Fun: 2.22 $\pm$ 0.3 -> 2.56 $\pm$ 0.24 | Helped |
| Add diversity filtering to RL | Transformer · Generated chess puzzles | Diversity | Helped |
Surprising findings
- Uniqueness rate stayed around 20% during RL even though uniqueness was part of the reward; gains came mainly from counter-intuitiveness.
- Auto-selected RL puzzles were rated lower than book puzzles on realism and difficulty but higher than Lichess puzzles on creativity and fun.
Limitations
- Counter-intuitiveness is defined only via Stockfish and AlphaZero search statistics and can be exploited (reward hacking such as meaningless checks); realism and entropy collapse required extra mechanisms.
- Expert study used only 8 raters and 10 puzzles per source; the booklet was curated and cherry-picked by humans, so the reported creativity is not purely auto-generated.
- Only the autoregressive transformer was RL-trained, and checkpoint selection by test loss may limit the AR model's peak puzzle score.
Experimental setup
Datasets: Lichess Puzzler Lichess games (2024-11) chess book puzzles (Golden Set)
Benchmarks: Lichess Puzzler dataset (generated samples) Golden Set counter-intuitiveness TRAIN/TEST
Metrics: legality rate uniqueness rate counter-intuitive rate puzzle rate average precision board distance PV distance sequence entropy expert ratings (realism, difficulty, creativity, fun, counte
Baselines: Lichess Puzzles Lichess Games MaskGIT Latent Diffusion Masked Diffusion Transformer (AR, supervised) Chess book puzzles
Builds on: PPO Lichess Puzzler methodology AlphaZero MD4 MaskGIT
Closest papers by meaning
- Human-Aligned Chess With a Bit of Search (ICLR 2025 · similarity 0.86)
- Designing Skill-Compatible AI: Methodologies and Frameworks in Chess (ICLR 2024 · similarity 0.85)
- Amortized Planning with Large-Scale Transformers: A Case Study on Chess (NeurIPS 2024 · similarity 0.85)
- Mastering Board Games by External and Internal Planning with Language Models (ICML 2025 · similarity 0.85)
- Implicit Search via Discrete Diffusion: A Study on Chess (ICLR 2025 · similarity 0.84)
- GAVEL: Generating Games via Evolution and Language Models (NeurIPS 2024 · similarity 0.84)
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.