Hint-before-Solving: A Framework to Effectively Utilizing Inherent Knowledge of Large Language Model
Abstract: Large Language Models (LLMs) have recently showcased remarkable generalizability in various domains.
Despite their extensive knowledge, LLMs still face challenges in efficiently utilizing encoded knowledge to develop accurate and logical reasoning. To mitigate this problem, we introduced the Hint-before-Solving framework (HinSo), which guides the model in generating hints (e.g., specific knowledge or key ideas) for solving the problem before the step-by-step solution.
Our studies involving 5 LLMs across 7 datasets of mathematical and commonsense reasoning results indicated that introducing hints before problem-solving can significantly enhance the performance of CoT.
To investigate whether LLMs can learn the HinSo pattern and improve their generalization ability, we constructed two large-scale and high-quality training datasets, HST-S and HST-L, containing 7.5k and 75k samples, respectively. The experimental results of supervised fine-tuning (SFT) showed that, under the same settings, the performance of model trained on the HinSo-formatted data improved significantly compared to CoT-formatted data, with a performance increase of 5.1% and 5.6% on the GSM8K, respectively.
We make our code and dataset publicly available at \url{https://github.com/sfhff216/hsp}.
Paper Type: Long
Research Area: Resources and Evaluation
Research Area Keywords: Language Modeling, NLP Applications, Generation
Contribution Types: NLP engineering experiment, Data resources
Languages Studied: English
Submission Number: 5828
Loading