---
license: mit
base_model: meta-llama/Llama-3.2-3B-Instruct
library_name: transformers
pipeline_tag: text-generation
tags: [latent-reasoning, looped-transformer, gsm8k, lotus]
---

# gsm-lotus-llama3b

**LOTUS** (direct LM-head supervision, no CODI) checkpoint fine-tuned from `meta-llama/Llama-3.2-3B-Instruct` on GSM8k-Aug, from the paper
[Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers](https://arxiv.org/abs/2606.31779).
This is the LOTUS model.

- **GSM8K (GSM8k-Aug) test accuracy:** 70.05% (924/1319)
- **Latent config:** K = 6 blocks, c_thought = 25 tokens/block, n_looped_iters = 6
- **Base:** meta-llama/Llama-3.2-3B-Instruct (vocab 128256 -> 128259 for 3 latent tokens)

## Loading
`from_pretrained` loads the weights only — the looped padded architecture needs the LOTUS code
([code repo](https://github.com/yingfan-bot/lotus)).

## Reproduce the number above
Run in the pinned env (torch 2.7 / transformers 4.46.2). This loads the safetensors straight from this
repo and runs the latent loop — no separate checkpoint file needed:
```bash
python scripts/eval.py \
  --model_id yingfanbot/gsm-lotus-llama3b \
  --datasets gsm8k --n_looped_iters 6 --c_thought 25 --bf16
```
Yields 70.05% on GSM8k-Aug (verified by loading this repo's safetensors directly).

## Citation
```bibtex
@article{fan2026bridging,
  title={Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers},
  author={Fan, Ying and Svete, Anej and Lee, Kangwook},
  journal={arXiv preprint arXiv:2606.31779},
  year={2026}
}
```