---
license: apache-2.0
language:
- en
tags:
- text-generation
- causal-lm
- adaptive-reasoning
- hierarchical-reasoning
- hrm
- custom-architecture
- compact-model
datasets:
- CosmicSet-2.0-mini
arxiv: 2605.28919
---

# CosmicFish-HRM

**Paper:** [CosmicFish-HRM: Adaptive Reasoning via Hierarchical Recurrent Mechanisms in Compact Language Models](https://arxiv.org/abs/2605.28919)

**GitHub:** [MistyozAI/CosmicFish-HRM](https://github.com/MistyozAI/CosmicFish-HRM)

CosmicFish-HRM is a compact 82.77M parameter causal language model built around a Hierarchical Reasoning Module (HRM) that dynamically allocates reasoning compute during inference. Rather than applying a fixed number of forward-pass layers to every input, the model iterates through high-level and low-level reasoning cycles and uses a learned halting head to decide when to stop. Harder inputs trigger deeper reasoning trajectories while simpler ones halt early.

Built at Mistyoz AI, Hyderabad.

---

## Architecture

![Architecture](architecture.png)

```
Input Blocks (Transformer) -> HRM Core (H + L levels, variable steps) -> Output Blocks (Transformer) -> LM Head
```

The HRM core maintains two interacting recurrent states operating at different abstraction levels. The high-level module captures slower, more abstract reasoning while the low-level module handles finer-grained local computation. After each reasoning step a lightweight halting head decides whether to continue or stop, conditioned on the mean-pooled high-level state.

**Key components:**

- Grouped-Query Attention (GQA) with 8 query heads and 4 KV heads
- Rotary Positional Embeddings (RoPE)
- SwiGLU feedforward layers
- RMSNorm (pre-norm for I/O blocks, post-norm inside HRM)
- Learned halt/continue Q-head controlling per-input reasoning depth
- Step penalty in the training loss encouraging efficient halting

## Model Specs

| Parameter | Value |
|---|---|
| Total parameters | 82.77M |
| Embedding dimension | 448 |
| Vocabulary size | 50,304 |
| Context length | 512 |
| Input transformer layers | 6 |
| Output transformer layers | 6 |
| HRM H-layers | 4 |
| HRM L-layers | 4 |
| Max HRM steps | 16 |
| Attention heads | 8 (4 KV, GQA) |

## Evaluation

Zero-shot benchmark results:

| Model | HellaSwag | PIQA | WinoGrande |
|---|---|---|---|
| CosmicFish-HRM (82M) | 26.2 | 58.1 | 50.7 |
| GPT-2 Small (117M) | 29.7 | 62.5 | 50.7 |
| OPT-125M | 30.6 | 62.6 | 52.9 |
| Pythia-160M | 29.4 | 62.1 | 52.8 |

At compact scale a portion of the parameter budget is allocated to the HRM reasoning infrastructure rather than raw language modeling capacity, which accounts for the gap versus fixed-depth baselines of similar size. The paper argues this tradeoff becomes more favorable as model scale increases.

## Adaptive Reasoning Behavior

The primary contribution of CosmicFish-HRM is not benchmark accuracy but adaptive compute allocation. The model uses different numbers of reasoning steps depending on input complexity:

| Prompt | Mean HRM Steps |
|---|---|
| "The capital of France is" | 2.78 |
| "Photosynthesis is the process by which plants" | 4.77 |
| "If all roses are flowers and some flowers fade quickly..." | 7.03 |
| "A bat and a ball cost $1.10 in total..." | 8.40 |

Average steps across benchmarks stay well below the 16-step maximum, with high variance across samples, confirming the halting mechanism is input-sensitive rather than collapsing to a fixed depth.

| Benchmark | Mean Steps | Std Dev |
|---|---|---|
| HellaSwag | 3.03 | 6.26 |
| PIQA | 1.87 | 5.13 |
| WinoGrande | 0.95 | 3.78 |
| Overall | 2.68 | 5.95 |

## Usage

This model uses a custom architecture. The model code is included in this repo as `modeling_hrm_cosmicfish.py`.

**Standalone chat script (downloads automatically):**

```bash
pip install torch safetensors huggingface-hub transformers termcolor
python chat.py
```

**Load manually:**

```python
import torch
import json
import tiktoken
from safetensors.torch import load_file
from huggingface_hub import snapshot_download
from modeling_hrm_cosmicfish import HRMCosmicFish, HRMCosmicFishConfig

cache_dir = snapshot_download("MistyozAI/CosmicFish-HRM")

with open(f"{cache_dir}/config.json") as f:
    cfg = json.load(f)

config = HRMCosmicFishConfig(
    vocab_size=cfg["vocab_size"],
    n_embd=cfg["n_embd"],
    block_size=cfg["block_size"],
    n_head=cfg["n_head"],
    n_kv_head=cfg["n_kv_head"],
    n_input_layers=cfg["n_input_layers"],
    n_output_layers=cfg["n_output_layers"],
    hrm_H_layers=cfg["hrm_H_layers"],
    hrm_L_layers=cfg["hrm_L_layers"],
    hrm_H_cycles=cfg["hrm_H_cycles"],
    hrm_L_cycles=cfg["hrm_L_cycles"],
    hrm_max_steps=cfg["hrm_max_steps"],
    dropout=0.0,
)

state_dict = load_file(f"{cache_dir}/model.safetensors")
model = HRMCosmicFish(config)
model.load_state_dict(state_dict)
model.eval()

tokenizer = tiktoken.get_encoding("gpt2")
prompt = "Artificial intelligence is"
tokens = tokenizer.encode(prompt)
idx = torch.tensor(tokens, dtype=torch.long).unsqueeze(0)

with torch.no_grad():
    output = model.generate(idx, max_new_tokens=100, temperature=0.7, top_k=40)

print(tokenizer.decode(output[0].tolist()))
```

---
Pytorch File: [CF.pt](https://drive.google.com/file/d/1He4PAIixuL5EMmzmxV4nq-OLI8xlp15Y/view?usp=sharing)

Pytorch File: [Base.pt](https://drive.google.com/file/d/1Apx898RYOtyDSjd_9IhoIGlTbNYf3N7H/view?usp=sharing)

---

Mistyoz AI, Hyderabad