Research map / Large language models / Agents, code and math
ARCANA: A Benchmark for Abstraction and Analogical Reasoning
Ruqi Chen, Wei Wang, Zilei Wang
Read the paperCheck your idea against 137,217 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.
Introduces ARCANA, a benchmark for far analogical reasoning between superficially dissimilar domains that share deep relational or structural similarity.
Problem
Few benchmarks explicitly test far analogical reasoning, i.e., mapping relational or structural similarities across superficially dissimilar domains.
Approach
A benchmark whose tasks require far analogical mappings from a source domain to a target domain that differ superficially but share deep relational structure.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use vision-language prompting instead of language-only chain-of-thought (ablation) | Gemini 3 Pro · ARCANA | accuracy (-Typo): 23.92% -> 18.66% | Hurt |
| use vision-language prompting instead of language-only chain-of-thought (ablation) | GPT 5.2 · ARCANA | accuracy (+Typo): 40.19% -> 42.11% (+Typo); 38.28% -> 37.32% (-Typo) | Mixed |
| combine language and vision-language setups (union of tasks solved) (ablation) | LLM baselines · ARCANA | accuracy: GPT 5.2 -Typo 38.28% / 37.32% -> 48.33% | Improved |
| correct analogy generated by the model | LLMs on pulley category · ARCANA pulley category | answer accuracy: GPT 0% vs Gemini 67% | Improved |
| apply ConceptSIR (concept slippage, introspection, reflection) instead of direct prompting | Gemini 3 Pro + GPT 5.2 · ARCANA hard subset (17 tasks) | tasks solved: solves 9/17 tasks | Improved |
| provide only task visualization without matrices to Gemini 3 Pro inside ConceptSIR (ablation) | ConceptSIR · pulley3 task | correct analogy | Hurt |
| increase reflection loop budget from 3 to 15 with GPT 5 (ablation) | ConceptSIR with GPT 5 · bead_chain1 task | solved | No clear effect |
Experimental setup
Benchmarks: ARCANA
Closest papers by meaning
- Distinguishing Performance From Competence in Evaluations of Humanlike Abstract Reasoning (NeurIPS 2026 · similarity 0.90)
- Think Visually, Reason Textually: Vision-Language Synergy in Abstract Reasoning (CVPR 2026 · similarity 0.89)
- Benchmarking Abstract and Reasoning Abilities Through A Theoretical Perspective (ICML 2025 · similarity 0.88)
- ARC Is a Vision Problem! (CVPR 2026 · similarity 0.87)
- MARVEL: Multidimensional Abstraction and Reasoning through Visual Evaluation and Learning (NeurIPS 2024 · similarity 0.87)
- Large Language Models as Analogical Reasoners (ICLR 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.
