Existing next-word prediction schemes suffer from factual and reasoning hallucinations. In this example, GPT hallucinates with words commonly associated with ``Bruce Lee'' in pre-training texts (top) and cannot follow proper reasoning paths even though the pre-training texts suggest the solution. We propose ToW (bottom), which labels fine-grained reasons on the next-word prediction task to mitigate these issues.
In this work, we explore a novel training-time data-augmentation method called thought-of-words (ToW), which injects fine-grained thoughts directly into the next-word prediction task and teaches the model to understand how the observed next word is related to previous contexts.
For more details, please refer to our paper!
If you find it useful for your research and applications, please cite our paper using this BibTeX:
@misc{xu2024towthoughtswordsimprove,
title={ToW: Thoughts of Words Improve Reasoning in Large Language Models},
author={Zhikun Xu and Ming Shen and Jacob Dineen and Zhaonan Li and Xiao Ye and Shijie Lu and Aswin RRV and Chitta Baral and Ben Zhou},
year={2024},
publisher={arXiv:2410.16235},
}