Research map / Large language models / Language, safety and interpretability
PuzLM: Solving Jigsaw Puzzles with Sequence-to-Sequence Language Models
Gur Elkin, Ofir I Shahar, Ohad Ben-Shahar
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 |
|---|---|---|---|
| add PCA before k-means (ablation) | PuzLM · JPwLEG-3 | perfect accuracy (%): 67.0 -> 84.5 | Improved |
| keep only border patches in super-token (ablation) | PuzLM · JPwLEG-3 | perfect accuracy (%): 50.3 -> 84.5 | Improved |
| lexicographic ordering of super-tokens (ablation) | PuzLM · JPwLEG-3 | perfect accuracy (%): 35.0 -> 84.5 | Improved |
| clockwise border traversal instead of raster scan (ablation) | PuzLM · JPwLEG-3 | perfect accuracy (%): 80.8 -> 84.5 | Improved |
| add [SEP] token between pieces (ablation) | PuzLM · JPwLEG-3 | perfect accuracy (%): 79.0 -> 84.5 | Improved |
| use lightweight PCA+k-means tokenizer instead of VQ-VAE (ablation) | PuzLM · JPwLEG-3 | perfect accuracy (%): 12.0 -> 84.5 | Improved |
| use lightweight tokenizer instead of TiTok-L (ablation) | PuzLM · JPwLEG-3 | perfect accuracy (%): 20.0 -> 84.5 | Improved |
| increase granularity B beyond 4 (ablation) | PuzLM · JPwLEG-3 | accuracy | Hurt |
| increase vocabulary size k beyond a point (ablation) | PuzLM · JPwLEG-3 | accuracy | No clear effect |
| transformer backbones instead of LSTM/GRU (ablation) | PuzLM · JPwLEG-3 | absolute accuracy | Improved |
| PuzLM vs FCViT on larger puzzles | PuzLM · JPwLEG-5 | perfect accuracy (%): 18.6 -> 32.5 | Improved |
| PuzLM vs FCViT | PuzLM · JPwLEG-3 | perfect accuracy (%): 87.9 -> 84.5 | Hurt |
Closest papers by meaning
- Solving Masked Jigsaw Puzzles with Diffusion Vision Transformers (CVPR 2024 · similarity 0.85)
- ERL-MPP: Evolutionary Reinforcement Learning with Multi-head Puzzle Perception for Solving Large-scale Jigsaw Puzzles of Eroded Gaps (AAAI 2025 · similarity 0.85)
- PuzzleWorld: A Benchmark for Multimodal, Open-Ended Reasoning in Puzzlehunts (ICLR 2026 · similarity 0.84)
- Enigmata: Scaling Logical Reasoning in Large Language Models with Synthetic Verifiable Puzzles (NeurIPS 2025 · similarity 0.83)
- LINGOLY: A Benchmark of Olympiad-Level Linguistic Reasoning Puzzles in Low Resource and Extinct Languages (NeurIPS 2024 · similarity 0.83)
- HardcoreLogic: Challenging Large Reasoning Models with Long-tail Logic Puzzle Games (ICLR 2026 · similarity 0.82)
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.
