Research map / Large language models / Agents, code and math
Generalized Planning for the Abstraction and Reasoning Corpus
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.
Casts ARC object-centric tasks as generalized planning problems in a PDDL-based DSL with external functions, solving them with the PGP(v) planner that synthesizes planning programs with pointers, outperforming ARGA and the Kaggle first-place solver on 160 tasks.
Problem
Solving Abstraction and Reasoning Corpus (ARC) tasks, which require fluid intelligence and abstraction that pure machine learning methods struggle with.
Approach
Model each ARC task as a generalized planning problem: a PDDL domain with object-centric graph abstractions and external functions, plus ARC-specific pruning of action schemes, predicates and pointer arguments, and a pixel-based heuristic, to synthesize planning programs with pointers and loops.
Key results
- GPAR solves 81/160 (50.63%) object-centric ARC tasks on the test set, versus 64/160 (40.00%) for the Kaggle first-place solution and 57/160 (35.62%) for ARGA.
- GPAR achieves 53.75% training accuracy (86/160), the highest overall.
- GPAR reaches 62.90% test accuracy on recolor tasks and 34.33% on augmentation tasks, while all solvers stay below 50% on augmentation in training and testing.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPAR instead of the strongest baseline | ARC solvers · ARC 160 object-centric tasks (all), training | accuracy: 49.38% -> 53.75% | Helped |
| Use GPAR instead of the strongest baseline | ARC solvers · ARC 160 object-centric tasks (all), testing | accuracy: 40.00% -> 50.63% | Helped |
| Use GPAR instead of the strongest baseline | ARC solvers · ARC object-centric movement, training | accuracy: 67.74% -> 64.52% | Hurt |
| Use GPAR instead of the strongest baseline | ARC solvers · ARC object-centric movement, testing | accuracy: 54.84% -> 61.30% | Helped |
| Use GPAR instead of the strongest baseline | ARC solvers · ARC object-centric recolor, training | accuracy: 40.32% -> 66.13% | Helped |
| Use GPAR instead of the strongest baseline | ARC solvers · ARC object-centric recolor, testing | accuracy: 45.16% -> 62.90% | Helped |
| Use GPAR instead of the strongest baseline | ARC solvers · ARC object-centric augmentation, training | accuracy: 52.24% -> 37.31% | Hurt |
| Use GPAR instead of the strongest baseline | ARC solvers · ARC object-centric augmentation, testing | accuracy: 31.34% -> 34.33% | Helped |
Surprising findings
- GPAR's augmentation-class performance is below 50% for all planners, so shape transformations remain hard for DSL-based imperative programs.
- Some tasks GPAR solves in training fail to generalize to test, e.g., ambiguity between 4- and 8-connected abstractions.
Limitations
- Weak on movement tasks needing dynamic node attributes (center, corner, area) and large numeric parameters.
- Restrictions make the solver incomplete; solutions can be ambiguous and overfit training instances.
Experimental setup
Datasets: ARC object-centric subset (160 tasks)
Benchmarks: ARC 160 object-centric tasks
Metrics: training accuracy testing accuracy
Baselines: ARGA Kaggle ARC first-place solution
Builds on: PGP(v) ARGA PDDL means-ends analysis
Closest papers by meaning
- RLAD: Training LLMs to Discover Abstractions for Solving Reasoning Problems (ICLR 2026 · similarity 0.86)
- ACPBench: Reasoning About Action, Change, and Planning (AAAI 2025 · similarity 0.86)
- Generalized Planning in PDDL Domains with Pretrained Large Language Models (AAAI 2024 · similarity 0.85)
- Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification (ICML 2025 · similarity 0.85)
- On the self-verification limitations of large language models on reasoning and planning tasks (ICLR 2025 · similarity 0.85)
- Planning Anything with Rigor: General-Purpose Zero-Shot Planning with LLM-based Formalized Programming (ICLR 2025 · 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.