Investigating the Impact of Rationales for LLMs on Natural Language Understanding

TL;DR

This work investigates whether rationales, especially chain-of-thought (CoT) outputs, can improve natural language understanding (NLU) tasks. It introduces NLURC, a large Chinese NLU dataset collection with 252,507 rationales across 34 datasets and 11 task types, and develops multiple rationale-augmented training and inference methods. The study finds that CoT inference benefits emerge with larger models, while most rationale-based training methods underperform label-only training, except for the Align approach, which balances label and rationale learning to yield consistent gains. Training on NLURC not only improves performance on seen tasks but also enhances generalization and rationale interpretability on unseen tasks, with smaller models rivaling much larger ones in some cases. Overall, NLURC provides a valuable benchmark and demonstrates that careful integration of rationales can boost NLU performance and explainability, particularly through balanced learning and instruction-tuning on diverse tasks.

Abstract

Chain-of-thought (CoT) rationales, which provide step-by-step reasoning to derive final answers, benefit LLMs in both inference and training. Incorporating rationales, either by generating them before answering during inference, or by placing them before or after the original answers during training - significantly improves model performance on mathematical, symbolic and commonsense reasoning tasks. However, most work focuses on the role of rationales in these reasoning tasks, overlooking their potential impact on other important tasks like natural language understanding (NLU) tasks. In this work, we raise the question: Can rationales similarly benefit NLU tasks? To conduct a systematic exploration, we construct NLURC, a comprehensive and high-quality NLU dataset collection with rationales, and develop various rationale-augmented methods. Through exploring the applicability of these methods on NLU tasks using the dataset, we uncover several potentially surprising findings: (1) CoT inference shifts from hindering NLU performance to surpassing direct label prediction as model size grows, indicating a positive correlation. (2) Most rationale-augmented training methods perform worse than label-only training, with one specially designed method consistently achieving improvements. (3) LLMs trained with rationales achieve significant performance gains on unseen NLU tasks, rivaling models ten times their size, while delivering interpretability on par with commercial LLMs.

Figures (16)

• Figure 1: Human evaluation on rationale quality of different prompt designs. Black dashed lines indicate average score for each dimension.
• Figure 2: Rationale generation prompt for a Paraphrase Detection task (LCQMC). Orange is the original prompt. Blue is input with the golden label. Red is the judge criteria for each label.
• Figure 3: Examples of training methods
• Figure 4: Align-trained 7B model's CoT error type distribution in Stance and RC. w/CS. task.
• Figure 5: Performance of varying label and rationale coefficients in Align training. The X-axis values represent the label coefficient, and 1 corresponds to Label-Only training.