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.
This is the LOTUS model.
from_pretrained loads the weights only — the looped padded architecture needs the LOTUS code
(
code repo).
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:
1python scripts/eval.py \
2 --model_id yingfanbot/gsm-lotus-llama3b \
3 --datasets gsm8k --n_looped_iters 6 --c_thought 25 --bf16
Yields 70.05% on GSM8k-Aug (verified by loading this repo's safetensors directly).
1@article{fan2026bridging,
2 title={Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers},
3 author={Fan, Ying and Svete, Anej and Lee, Kangwook},
4 journal={arXiv preprint arXiv:2606.31779},
5 year={2026}
6}
