Research map / Large language models / Agents, code and math
Subgoal Search For Complex Reasoning Tasks
Konrad Czechowski, Tomasz Odrzygóźdź, Marek Zbysiński, Michał Zawalski, Krzysztof Olejnik, Yuhuai Wu, Łukasz Kuciński, Piotr Miłoś
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 |
|---|---|---|---|
| use subgoal generator (BF-kSubS) instead of low-level-action BestFS | best-first search · INT proof length 15, graph size 50 | success rate: 0.09 -> 0.38 | Improved |
| use subgoal generator instead of low-level actions at larger budget | best-first search · INT proof length 15, graph size 400 | success rate: 0.47 -> 0.91 | Improved |
| use subgoal generator on Sokoban 20x20 | best-first search · Sokoban 20x20, graph size 50 | success rate: 0.02 -> 0.46 | Improved |
| use subgoal generator on Sokoban 12x12 large budget | best-first search · Sokoban 12x12, graph size 5000 | success rate: 0.83 -> 0.93 | Improved |
| use subgoal generator on Rubik's Cube | best-first search · Rubik's Cube | success rate: less than 10% -> nearly 100% | Improved |
| use MCTS with subgoals instead of AlphaZero | MCTS · INT proof length 15 | success rate | Improved |
| increase subgoal distance k (ablation) | BF-kSubS · INT and Rubik's Cube | success rate | Mixed |
| increase subgoal distance k up to 10 (ablation) | BF-kSubS · Sokoban | success rate | Improved |
| use subgoals with k=4 instead of single steps on noisy value function (ablation) | best-first search · Grid world sigma=10 | success rate: 0.142 -> 1 | Improved |
| use subgoals with k=4 under highest value noise (ablation) | best-first search · Grid world sigma=20 | success rate: 0.006 -> 0.983 | Improved |
| use subgoals with k=4 under low value noise (ablation) | best-first search · Grid world sigma=3 | success rate: 0.999 -> 1 | No clear effect |
| move k=4 steps ahead instead of 1 step along solution path (ablation) | value estimates · INT | probability value decreases: 0.32 -> 0.02 | Improved |
Closest papers by meaning
- HardcoreLogic: Challenging Large Reasoning Models with Long-tail Logic Puzzle Games (ICLR 2026 · similarity 0.88)
- PUZZLES: A Benchmark for Neural Algorithmic Reasoning (NeurIPS 2024 · similarity 0.86)
- PuzzleWorld: A Benchmark for Multimodal, Open-Ended Reasoning in Puzzlehunts (ICLR 2026 · similarity 0.86)
- Enigmata: Scaling Logical Reasoning in Large Language Models with Synthetic Verifiable Puzzles (NeurIPS 2025 · similarity 0.85)
- Boosting of Thoughts: Trial-and-Error Problem Solving with Large Language Models (ICLR 2024 · similarity 0.85)
- THOUGHT PROPAGATION: AN ANALOGICAL APPROACH TO COMPLEX REASONING WITH LARGE LANGUAGE MODELS (ICLR 2024 · 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.
