chriscowncrow/TinyRecursiveModels
🔍 Explore recursive reasoning with TinyRecursiveModels, a compact 7M parameter neural network achieving high scores on tough tasks without massive resources.
What's novel
🔍 Explore recursive reasoning with TinyRecursiveModels, a compact 7M parameter neural network achieving high scores on tough tasks without massive resources.
Code Analysis
13 files read · 4 roundsA research codebase implementing small (~7M param) non-autoregressive transformers that solve ARC-AGI visual reasoning puzzles through iterative recursive hidden-state refinement with adaptive computation time (ACT) halting and puzzle-specific learned embeddings.
Strengths
Genuinely novel recursive reasoning architecture with truncated BPTT for memory-efficient deep iteration, clever distributed SignSGD for sparse puzzle embeddings, and a well-designed ARC evaluator with pass@K aggregated voting across training steps.
Weaknesses
No test suite exists, significant code duplication across model variants (trm_hier6 hardcodes 6 levels), and the README is deeply misleading—presenting a research codebase as a downloadable consumer application with 'double-click to run' instructions.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
9
Contributors
2
Files
41
Active weeks
3
Repository
Language
Python
Stars
3
Forks
0
License
MIT