Tiny Recursive Model (TRM)

From Helix Project Wiki
Revision as of 18:13, 8 October 2025 by Steve Helix (talk | contribs) (Created page with "= Tiny Recursive Model (TRM) = thumb|right|400px|Tiny Recursive Model diagram (from Samsung SAIL Montréal) '''Tiny Recursive Model (TRM)''' is a compact neural architecture proposed by Alexia Jolicoeur-Martineau at Samsung SAIL Montréal that demonstrates advanced reasoning capabilities despite being ~10,000x smaller than most large language models. == Overview == TRM is designed around the principle of **recursi...")
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)

Tiny Recursive Model (TRM)

File:TRM Architecture.png
Tiny Recursive Model diagram (from Samsung SAIL Montréal)

Tiny Recursive Model (TRM) is a compact neural architecture proposed by Alexia Jolicoeur-Martineau at Samsung SAIL Montréal that demonstrates advanced reasoning capabilities despite being ~10,000x smaller than most large language models.

Overview

TRM is designed around the principle of **recursive self-refinement**. Unlike typical LLMs that predict responses token-by-token in a single pass, TRM generates an initial draft, constructs an internal “latent reasoning” state (scratchpad), and then improves its own answer over up to 16 iterations. This mimics meta-cognition in a compact form.

Key Features

  • **Tiny Model Size**: ~7M parameters
  • **Deep Reasoning Loop**: Recursively improves answers across time steps
  • **Superior Test Accuracy**:
 * 45% on ARC-AGI-1
 * 8% on ARC-AGI-2
 * Outperforms Gemini 2.5 Pro, DeepSeek R1, and O3-mini on these benchmarks
  • **Minimal Training Data**: Trained on only ~1000 examples

Architecture Summary

  • Two small networks: a **Reasoning MLP** and a **Prediction MLP**
  • Reasoning loop feeds latent variables into the Prediction MLP
  • Uses self-attention + scratchpad updates per reasoning step
  • Final output selected after convergence or max iteration (n=16)

Comparison to LLMs

Feature TRM Large Language Models
Parameters ~7M 70B–175B+ Reasoning Style Iterative, recursive Single-pass prediction ARC-AGI Performance ↑ Better on logic tasks ↓ Weaker generalization Memory Footprint Extremely low High Explainability Easier to trace steps Often opaque Use Case Fit Logic puzzles, verifiable computation Open-ended language, wide context

Relevance to Helix

TRM offers a feasible lightweight module for tasks that require:

  • Deterministic verification
  • Local reasoning on edge devices
  • Constrained-resource environments
  • Transparent reasoning chains (auditable steps)

Possible integration points:

Resources

License

MIT – freely reusable with attribution

Status

Template:Status