LOTUS trained on natural-language chain-of-thought, fine-tuned from meta-llama/Llama-3.2-3B-Instruct, from the paper Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers. Direct step-aligned supervision through the base LM head (no aux decoder, no CODI); the latent blocks are supervised against natural-language reasoning steps rather than formatted math expressions.

GSM8K (GSM8k-Aug) test accuracy: 68.54% (904/1319)
Latent config: K = 6 blocks, c_thought = 50 tokens/block (NL steps are ~2.7x longer than math-expr, so c=50 covers ~98.8% of steps), 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).

Reproduce the number above

Run in the pinned env (torch 2.7 / transformers 4.46.2) — note --c_thought 50 for this NL model:

bash

python scripts/eval.py \
  --model_id yingfanbot/gsm-lotus-llama3b-nl \
  --datasets gsm8k --n_looped_iters 6 --c_thought 50 --bf16

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}
}