abstract-reasoning

RAVEN: A Dataset for Relational and Analogical Visual rEasoNing

Evaluation on Logical Reasoning and Abstract Reasoning Challenges

Abstract Spatial-Temporal Reasoning via Probabilistic Abduction and Execution

Learning Perceptual Inference by Contrasting

Usable implementation of Emerging Symbol Binding Network (ESBN), in Pytorch

ACRE: Abstract Causal REasoning Beyond Covariation

Learning Algebraic Representation for Systematic Generalization in Abstract Reasoning

Effective Abstract Reasoning with Dual-Contrast Network

A dual-layer framework enhancing LLM abstract synthesis through standard output expansion and omnidirectional insight generation, with applications in AI design and creative ideation.

ARC-Test-Time-Training (ARC-TTT)

A prompt-level hack for deeper LLM thinking, which applies abstract reasoning principles to direct LLMs to look at paradoxes and edge cases from different angles.

Official package for "A Neural Affinity Framework for Abstract Reasoning." Includes the validated 9-category ARC taxonomy, pre-computed fine-tuning results, and scripts to verify the Compositional Gap.

CausalARC: Abstract Reasoning with Causal World Models

LAteNT: Neuro-symbolic multi-agent system for abstract reasoning. Nine specialized agents coordinate via blackboard architecture implementing Socratic debate. Achieves program synthesis through MDL-scored hypothesis generation, Popperian falsification, and counterfactual causal verification. Near Perfect solve rate on procedural reasoning tasks.

Audited ARC-AGI-2 research on support-conditioned grid rewriting, CompressARC semantics, pairing interventions, target isolation, and exact-grid readouts.For ARC run 2026

An extremely difficult benchmark for LLMs.

Hypothèse de la Cognition Géométrique. Cadre théorique unifiant l'intuition mathématique et la perception esthétique comme des processus de reconnaissance topologique dans des variétés neuronales de haute dimension.

Open-source utilities from my ARC Prize 2026 (ARC-AGI-2) work: a Bearer-auth Kaggle API client for the new KGAT_ tokens, a kernel run poller, and a small ARC DSL.

4.8M-parameter recursive transformer for ARC-AGI-1 (3-D RoPE, dual-state recursion, deep supervision) — and the 15-run ablation programme showing test-time training, not architecture, is what makes it work.

Verified Python programs for a subset of ARC-AGI-2 training tasks, with a focus on executable problem–program pairs for LLM program synthesis.