zbloss/HRM-maze-30x30-hard

HRM Maze 30x30 Hard

A Hierarchical Reasoning Model (HRM) trained to solve hard 30×30 maze navigation problems using hierarchical processing and adaptive computation.

Model Details

Model Description

This is a Hierarchical Reasoning Model checkpoint fine-tuned specifically for solving hard maze pathfinding problems on 30×30 grids. The model employs a two-level hierarchical architecture inspired by human cognition, with high-level (H) modules for abstract route planning and low-level (L) modules for detailed navigation decisions. It uses Adaptive Computation Time (ACT) with Q-learning based halting to dynamically allocate computational resources.

The model processes maze grids up to 30×30 (900 tokens) and predicts optimal navigation paths through complex maze environments.

• Developed by: Sapient Inc.
• Model type: Hierarchical Reasoning Model (HRM)
• Language(s): Symbolic reasoning (maze navigation symbols)
• License: Apache 2.0
• Original checkpoint: sapientinc/HRM-checkpoint-maze-30x30-hard

Model Sources

• Repository: transformers
• Paper: Hierarchical Reasoning Model
• Original Repository: HRM GitHub

Uses

Direct Use

This model is designed for solving hard maze navigation problems. It can:

• Find optimal paths through complex 30×30 maze environments
• Navigate mazes with multiple obstacles and dead ends
• Process partial maze representations and predict navigation sequences
• Demonstrate hierarchical planning strategies for spatial reasoning tasks

Downstream Use

The model can be used as:

• A component in game AI and procedural content generation
• A baseline for research in hierarchical spatial reasoning
• An example of applying neural networks to pathfinding and navigation problems
• A planning module in robotics and autonomous navigation research

Recommendations

Users should be aware that:

• The model is specialized for maze pathfinding and should not be used for general spatial reasoning tasks
• Input must be properly formatted as grid representations with the 6-token vocabulary
• Inference time may vary due to the adaptive computation mechanism
• The model is optimized for hard difficulty mazes and may be over-engineered for simple mazes

How to Get Started with the Model

import torch
from transformers import HrmForCausalLM

# Load the model
model = HrmForCausalLM.from_pretrained("zbloss/HRM-maze-30x30-hard")
model.eval()

# Prepare a maze grid (e.g., 20x20 = 400 tokens)
# Vocabulary: 0-5 representing different maze elements
# (e.g., 0=empty, 1=wall, 2=start, 3=goal, 4=path, 5=visited)
maze_grid = torch.randint(0, 6, (1, 400))  # Example 20x20 maze
puzzle_ids = torch.zeros(1, dtype=torch.long)

# Run inference
with torch.no_grad():
    outputs = model(input_ids=maze_grid, puzzle_identifiers=puzzle_ids)

# Get predictions
predictions = torch.argmax(outputs.logits, dim=-1)
print(f"Predicted navigation path: {predictions}")
print(f"Q-halt: {outputs.q_halt_logits[0]:.4f}")
print(f"Q-continue: {outputs.q_continue_logits[0]:.4f}")

Training Details

Training Data

The model was trained on a dataset of hard difficulty 30×30 maze environments. These mazes feature:

• Complex layouts with multiple branching paths
• Dead ends requiring backtracking
• Long optimal paths requiring multi-step planning
• Variable start and goal positions

Training Procedure

The model uses a hierarchical architecture with:

• High-level (H) module: 4 transformer layers for abstract route planning
• Low-level (L) module: 4 transformer layers for detailed navigation decisions
• H-cycles: 2 high-level reasoning cycles for strategic planning
• L-cycles: 2 low-level computation cycles per H-cycle for tactical moves
• ACT mechanism: Q-learning based adaptive halting with max 16 steps

Training Hyperparameters

• Training regime: bfloat16 mixed precision
• Architecture: 4 H-layers, 4 L-layers, 8 attention heads
• Hidden size: 512
• Intermediate size: 1536
• Max position embeddings: 900 (supports up to 30×30 grids)
• Vocabulary size: 6 (maze navigation symbols)

Model Architecture

Technical Specifications

Component	Value
Total Parameters	27,270,658 (27.3M)
Model Size	109.09 MB
Vocabulary Size	6
Hidden Size	512
Intermediate Size	1536
H-level Layers	4
L-level Layers	4
Attention Heads	8
H-cycles	2
L-cycles	2
Max Halting Steps	16
Max Grid Size	30×30 (900 tokens)
Position Encoding	RoPE (Rotary Position Embeddings)
Activation	SwiGLU

Model Architecture and Objective

The Hierarchical Reasoning Model (HRM) features:

1. Two-level Hierarchical Processing:

   • H-level (High-level): Performs slow, abstract route planning and strategic navigation
   • L-level (Low-level): Executes fast, detailed navigation decisions and obstacle avoidance

2. Adaptive Computation Time (ACT):

   • Q-learning based halting mechanism
   • Dynamically determines when sufficient computation has been performed
   • Allows variable computational depth based on maze complexity
   • More complex mazes with longer paths trigger more reasoning cycles

3. Recurrent Carry State:

   • Maintains H and L hidden states across reasoning cycles
   • Enables iterative refinement of navigation strategies
   • Supports backtracking and path correction

4. Positional Encoding:

   • RoPE (Rotary Position Embeddings) for position-aware attention
   • Critical for spatial reasoning in grid-based environments
   • Supports up to 900 positions (30×30 grids)

Compute Infrastructure

Software

• Framework: PyTorch with transformers library
• Precision: bfloat16
• Format: Safetensors

Performance

The model is designed to solve hard difficulty mazes on 30×30 grids, demonstrating:

• Multi-step planning capabilities for long navigation sequences
• Ability to recognize and avoid dead ends
• Strategic backtracking when necessary
• Hierarchical decomposition of complex navigation problems

Citation

BibTeX:

@article{wang2025hierarchical,
  title={Hierarchical Reasoning Model},
  author={Wang, Guan and Li, Jin and Sun, Yuhao and Chen, Xing and Liu, Changling and Wu, Yue and Lu, Meng and Song, Sen and Yadkori, Yasin Abbasi},
  journal={arXiv preprint arXiv:2506.21734},
  year={2025}
}

APA:

Wang, G., Li, J., Sun, Y., Chen, X., Liu, C., Wu, Y., Lu, M., Song, S., & Yadkori, Y. A. (2025). Hierarchical Reasoning Model. arXiv preprint arXiv:2506.21734.

More Information

This checkpoint is a converted version of the original HRM checkpoint from sapientinc/HRM-checkpoint-maze-30x30-hard, formatted for use with the HuggingFace transformers library.

For more details about the HRM architecture and training methodology, see:

• Paper: https://arxiv.org/abs/2506.21734
• Original Implementation: https://github.com/sapientinc/HRM

Example Use Cases

1. Game AI: Intelligent maze navigation in video games
2. Path Planning Research: Baseline for hierarchical planning algorithms
3. Robotics: Inspiration for hierarchical navigation strategies
4. Education: Demonstrating neural approaches to classic AI problems

Model Card Contact

For questions or issues with this converted checkpoint, please open an issue in the transformers repository.