Tiny Recursive Model (TRM)

A Parameter-Efficient Hierarchical Architecture for Iterative Reasoning

Abstract

This repo implements the Tiny Recursive Model (TRM) — a lightweight recurrent architecture designed for structured reasoning through hierarchical recursion.

We experiment with TRM in both supervised and reinforcement learning settings.

In supervised tasks, TRM achieves strong parameter efficiency (up to 67–87% fewer parameters).

In reinforcement learning, we find that policy gradient methods struggle to fully exploit TRM’s recursive reasoning ability.

Overall, our results suggest that iterative refinement models work best when trained with structured or explicit supervision.

Full experimental analysis: EXPERIMENTAL_RESULTS.md

Key Contributions

1. Parameter Efficiency: TRM achieves competitive accuracy with 67-87% fewer parameters than baseline models on supervised spatial reasoning tasks
2. Architectural Innovation: Hierarchical recursive design with K-cycle iterative refinement and carry mechanisms
3. Empirical Analysis: Systematic comparison across learning paradigms revealing task-specific architectural benefits
4. Open Implementation: Modular codebase enabling replication and extension of results

---

1. Overview

TRM is a hierarchical recursive architecture where representations are iteratively refined through multiple reasoning cycles. The model uses two nested processing levels (H-level and L-level) with a carry mechanism for information persistence.

Implementation Variants

Implementation	Module	Description
Baseline	trm.py	Minimal recursive architecture
RL-Adapted	trm_rl_evns.py	Environment-specific RL implementations
Enhanced	trm_rl_improved.py	Improved training strategies
Full	trm_rl.py	Complete hierarchical TRM with K-cycles
Debug	trm_rl_debug.py	Diagnostic tools

---

2. Repository Structure

Tiny Recursive Model/
├── README.md
├── EXPERIMENTAL_RESULTS.md
├── requirements.txt
├── trm_core/
│   ├── trm.py
│   ├── trm_rl.py
│   ├── trm_rl_debug.py
│   ├── trm_rl_evns.py
│   ├── trm_rl_improved.py
│   ├── trm_rl_optimized.py
│   ├── trm_rl_optimized_updated_sudoku.py
│   └── utils/
│       └── improved_sudoku_env.py
├── experiments/
│   ├── simple_maze_comparison.py
│   └── maze_comparison.py
└── results/
    └── experiments/

---

3. Installation

git clone https://github.com/acharyaanusha/Tiny-Recursive-Model.git
cd Tiny-Recursive-Model
pip install -r requirements.txt

Requirements: Python ≥ 3.9, PyTorch ≥ 2.0, NumPy, Matplotlib, tqdm, gymnasium

---

4. Running Experiments

Supervised Learning:

python experiments/simple_maze_comparison.py  # 8×8 maze
python experiments/maze_comparison.py         # 10×10 maze

Reinforcement Learning:

python trm_core/trm_rl_debug.py  # Diagnostics
python trm_core/trm_rl.py        # Full training

---

5. Architecture

Baseline:

State → Embedding → [Recursive Processing] → Output
                           ↑___feedback____|

Hierarchical TRM:

State → Embedding → H-Level → L-Level → Output
                      ↑______carry______|
                   (repeat K cycles)

Key Features: K-cycle recursion, hierarchical processing, carry mechanism, SwiGLU activation, deep supervision

---

6. Hyperparameters

Maze (5×5): h_dim=64, l_dim=32, k_cycles=3, lr=2e-3, episodes=500

Sudoku (4×4): h_dim=128, l_dim=64, k_cycles=5, lr=1e-3, episodes=1000

---

7. Results

Full analysis: EXPERIMENTAL_RESULTS.md

Supervised Learning

Task	Model	Parameters	Accuracy
Simple Maze (8×8)	CNN	74,850	100%
	TRM	9,922	100% (87% fewer params)
Complex Maze (10×10)	Transformer	~150K	~100%
	TRM	~50K	~100% (67% fewer params)

Reinforcement Learning

Model	Parameters	Success Rate
MLP Baseline	20,356	36.0%
TRM (K=5)	69,038	34.0%
TRM (K=3)	69,038	22.0%

Key Findings

• Supervised: TRM achieves competitive accuracy with 67-87% fewer parameters
• RL: TRM does not outperform MLP despite 3× more parameters
• Hypothesis: Iterative refinement requires structured targets (labels) rather than scalar rewards for effective learning

---

8. Citation

@misc{jolicoeurmartineau2025morerecursivereasoningtiny,
    title   = {Less is More: Recursive Reasoning with Tiny Networks}, 
    author  = {Alexia Jolicoeur-Martineau},
    year    = {2025},
    eprint  = {2510.04871},
    archivePrefix = {arXiv},
    primaryClass = {cs.LG},
    url     = {https://arxiv.org/abs/2510.04871}, 
}

If you use this repository in your research, please cite:

@software{acharya2025trm,
  title  = {Tiny Recursive Model (TRM): A Parameter-Efficient Hierarchical Architecture for Iterative Reasoning},
  author = {Anusha Acharya},
  year   = {2025},
  url    = {https://github.com/acharyaanusha/Tiny-Recursive-Model}
}