Saved in:
| Main Authors: | , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2602.08220 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866911500241731584 |
|---|---|
| author | Zeng, Boyi Hao, Yiqin Li, He Song, Shixiang Song, Feichen Wang, Zitong Huang, Siyuan Xu, Yi He, ZiWei Wang, Xinbing Lin, Zhouhan |
| author_facet | Zeng, Boyi Hao, Yiqin Li, He Song, Shixiang Song, Feichen Wang, Zitong Huang, Siyuan Xu, Yi He, ZiWei Wang, Xinbing Lin, Zhouhan |
| contents | Scaling large language models by increasing parameters and training data is increasingly constrained by limited high-quality corpora and rising communication costs. This work explores an alternative axis: increasing per-token computation without expanding parameters, by internalizing latent Chain-of-Thought (CoT) into pretraining. We propose Pretraining with Token-Level Adaptive Latent CoT (adaptive latent CoT), where the model generates a variable-length latent CoT trajectory before emitting each token -- allocating longer trajectories to difficult tokens and shorter (or even zero) trajectories to easy ones. Importantly, this behavior emerges naturally from one-stage pretraining on general text and reduces computation in both training and inference via token-wise adaptive halting. Experiments with Llama architectures show that adaptive latent CoT consistently improves language modeling perplexity and broad downstream accuracy, even with fewer training FLOPs than prior recurrent baselines. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_08220 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Pretraining with Token-Level Adaptive Latent Chain-of-Thought Zeng, Boyi Hao, Yiqin Li, He Song, Shixiang Song, Feichen Wang, Zitong Huang, Siyuan Xu, Yi He, ZiWei Wang, Xinbing Lin, Zhouhan Computation and Language Scaling large language models by increasing parameters and training data is increasingly constrained by limited high-quality corpora and rising communication costs. This work explores an alternative axis: increasing per-token computation without expanding parameters, by internalizing latent Chain-of-Thought (CoT) into pretraining. We propose Pretraining with Token-Level Adaptive Latent CoT (adaptive latent CoT), where the model generates a variable-length latent CoT trajectory before emitting each token -- allocating longer trajectories to difficult tokens and shorter (or even zero) trajectories to easy ones. Importantly, this behavior emerges naturally from one-stage pretraining on general text and reduces computation in both training and inference via token-wise adaptive halting. Experiments with Llama architectures show that adaptive latent CoT consistently improves language modeling perplexity and broad downstream accuracy, even with fewer training FLOPs than prior recurrent baselines. |
| title | Pretraining with Token-Level Adaptive Latent Chain-of-Thought |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2602.08220 |