This is tomg-group-umd/huginn_swa_75_7_ema_0.9_merge a weight-averaged checkpoint our base model that improves on GSM8k performance. All other details are as the main model.

How to Use
Advanced Usage
Model Summary
Limitations
Technical Details
License
Citation

Downloading and Using the Model

Load the model like this:

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("tomg-group-umd/huginn-0125", torch_dtype=torch.bfloat16, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("tomg-group-umd/huginn-0125")

Modifying the Model's Depth at Test Time:

By providing the argument num_steps, the model will execute a forward pass with that amount of compute:

python

input_ids = tokenizer.encode("The capital of Westphalia is", return_tensors="pt", add_special_tokens=True).to(device)
model.eval()
model.to(device)

model(input_ids, num_steps=32)

The model has about 1.5B parameters in non-recurrent code, 0.5B parameters in the embedding, and 1.5B recurrent parameters, so, as a guideline, the number of materialized parameters is num_steps * 1.5B + 2B. Playing with this parameter is what makes this model interesting, and different from fixed-depth transformers! The model is trained to accept an arbitrary number of steps. However, using fewer than 4 steps will result in very coarse answers. If given enough context to reason about, benchmarks show the model improving up to around num_steps=64. Beyond that, more steps generally do not hurt, but we see no further improvements.

Inference

The model was trained with bfloat16-mixed precision, so we recommend using bfloat16 to run inference (or AMP bfloat16-mixed precision, if you really want). All benchmarks were evaluated in pure bfloat16.

Sampling

The model can be used like a normal HF model to generate text with KV-caching working as expected. You can provide num_steps directly to the generate call, for example:

model.eval()
config = GenerationConfig(max_length=256, stop_strings=["<|end_text|>", "<|end_turn|>"], 
                          use_cache=True,
                          do_sample=False, temperature=None, top_k=None, top_p=None, min_p=None, 
                          return_dict_in_generate=True,
                          eos_token_id=65505,bos_token_id=65504,pad_token_id=65509)


input_ids = tokenizer.encode("The capital of Westphalia is", return_tensors="pt", add_special_tokens=True).to(device)
outputs = model.generate(input_ids, config, tokenizer=tokenizer, num_steps=16)

Note: num_steps and other model arguments CANNOT be included in the GenerationConfig, they will shadow model args at runtime.

Chat Templating

The model was not finetuned or post-trained, but due to inclusion of instruction data during pretraining, natively understand its chat template. You can chat with the model like so

messages = []
messages.append({"role": "system", "content" : You are a helpful assistant."}
messages.append({"role": "user", "content" : What do you think of Goethe's Faust?"}
chat_input = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
print(chat_input)
input_ids = tokenizer.encode(chat_input, return_tensors="pt", add_special_tokens=False).to(device)

model.generate(input_ids, config, num_steps=64, tokenizer=tokenizer)

KV-cache Details

The model requires its own KV-cache implementation HuginnDynamicCache, otherwise the KV-caches of later calls to the recurrent block will overwrite the earlier ones. The current implementation will always try to inject this Cache implementation, but that may break with huggingface updates. If you do not use generate, but implement your own generation, use a pattern like this:

python

# first step:
past_key_values = None
outputs = model(input_ids=input_ids, use_cache=True, past_key_values=past_key_values)
past_key_values = outputs.past_key_values # Should be an instance of HuginnDynamicCache
# next step
outputs = model(input_ids=input_ids, use_cache=True, past_key_values=past_key_values)

Advanced Features

Per-Token Adaptive Compute

When generating, you can also a variable amount of compute per-token. The model is not trained for this, so this is a proof-of-concept, that can do this task zero-shot. You can pick between a few sane stopping rules, entropy-diff, latent-diff,kl and argmax-stability, via criterion=kl. The exit threshold can be modified via exit_threshold=5e-4. We suggest using kl for interesting exits and argmax-stability for conservative exits. Note that using these variables overrides the default generation function. Not all arguments that are valid for the normal generate call are valid here. To make this more explicit, you can also directly call generate_with_adaptive_compute:

python

from transformers import TextStreamer
streamer = TextStreamer(tokenizer)

model.generate_with_adaptive_compute(input_ids, config, num_steps=64, tokenizer=tokenizer, streamer=streamer,
                                     continuous_compute=False, criterion="kl", exit_threshold=5e-4, cache_kwargs={"lookup_strategy": "latest-m4"})

Your cache strategy should be set to "latest-m4" if using adaptive compute.

KV-cache Sharing

To reduce KV cache memory requirements, the model can be run with fewer KV-caches, with later iterations in the recurrence overwriting earlier caches. To use this feature, set the cache argument lookup_strategy to include compress-s16 (where the last number determine the size of the cache).

model.generate_with_adaptive_compute(input_ids, config, num_steps=64, tokenizer=tokenizer, streamer=streamer,
                                     continuous_compute=False, cache_kwargs={"lookup_strategy": "compress-s16"})

You can combine this per-token adaptive compute. In that case your lookup strategy should be latest-m4-compress-s16.

Warmstart / Continuous CoT

At each generation step, the recurrence can be warmstarted with the final state from the previous token by setting continuous_compute=True, like so

model.generate_with_adaptive_compute(input_ids, config, num_steps=64, tokenizer=tokenizer, streamer=streamer, continuous_compute=True)

Model Summary

The model is primarily structured around decoder-only transformer blocks. However these blocks are structured into three functional groups, the prelude $P$, which embeds the input data into a latent space using multiple transformer layers, then the core recurrent block $R$, which is the central unit of recurrent computation modifying states $\mathbf{s} \in \mathbb{R}^{n \times h }$, and finally the coda $C$, which un-embeds from latent space using several layers and also contains the prediction head of the model.

Given a number of recurrent iterations $r$, and a sequence of input tokens $\mathbf{x} \in V^n$ these groups are used in the following way to produce output probabilities $\mathbf{p} \in \mathbb{R}^{n \times |V|}$.

$$\mathbf{e} = P(\mathbf{x})$$

$$\mathbf{s}0 \sim \mathcal{N}(\mathbf{0}, \sigma^2 I{n\cdot h})$$

$$\mathbf{s}i = R(\mathbf{e}, \mathbf{s}{i-1}) ; \textnormal{for} ; i \in \lbrace 1, \dots, r \rbrace$$

$$\mathbf{p} = R(\mathbf{s}_r)$$ where $\sigma$ is the standard deviation of the initial random state. Given an init random state $\mathbf{s}0$, the model repeatedly applies the core block $R$, which accepts the latent state $\mathbf{s}{i-1}$ and the embedded input $\mathbf{e}$ and outputs a new latent state $\mathbf{s}_i$. After finishing all iterations, the coda block processes the last state and produces the probabilities of the next token.

Please refer to the paper for benchmark performance on standard benchmarks.

Limitations

Our checkpoint is trained for only 47000 steps on a broadly untested data mixture with a constant learning rate. As an academic project, the model is trained only on publicly available data and the 800B token count, while large in comparison to older fully open-source models such as the Pythia series, is small in comparison to modern open-source efforts such as OLMo, and tiny in comparison to the datasets used to train industrial open-weight models.

Technical Specifications

This model was trained on 21 segments of 4096 AMD MI-250X GPUs on the OLCF Frontier Supercomputer in early December 2024. The model was trained using ROCM 6.2.0, and PyTorch 2.6 nightly pre-release 24/11/02. The code used to train the model can be found at https://github.com/seal-rg/recurrent-pretraining.

License

This model is released under the apache-2.0 licence.

Citation

@article{geiping2025scaling,
      title={Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach}, 
      author={Jonas Geiping and Sean McLeish and Neel Jain and John Kirchenbauer and Siddharth Singh and Brian R. Bartoldson and Bhavya Kailkhura and Abhinav Bhatele and Tom Goldstein},
      year={2025},
      eprint={2502.},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

You can also find the paper at https://huggingface.co/papers/2502.05171.

Contact

Please, feel free to contact us with any questions, or open an discussion thread on Hugging Face.

tomg-group-umd/huginn_swa_75_7_ema_0.9_merge

Huginn_swa_75_7_ema_0.9_merge

Table of Contents

Downloading and Using the Model

Modifying the Model's Depth at Test Time:

Inference

Sampling

Chat Templating

KV-cache Details

Advanced Features

Per-Token Adaptive Compute

KV-cache Sharing

Warmstart / Continuous CoT

Model Summary

Limitations

Technical Specifications

License

Citation

Contact

tomg-group-umd/
huginn_swa_75_7_ema_0.9_merge