Investigating the Impact of Rationales for LLMs on Natural Language Understanding

2025-10-18

AI Summary

Prior chain-of-thought (CoT) research focuses predominantly on mathematical and commonsense reasoning, overlooking its role in natural language understanding (NLU), leaving a critical gap in understanding CoT’s efficacy and transferability for NLU tasks. Method: We introduce NLURC—the first high-quality, rationale-annotated NLU benchmark—and propose a rationale-enhanced training framework integrating rationale generation, injection, and supervised fine-tuning. Contribution/Results: (1) Activating CoT in large language models yields substantial gains over direct prediction—matching or exceeding the performance of smaller models with tenfold fewer parameters; (2) our method improves zero-shot generalization to unseen NLU tasks; (3) generated rationales achieve high fidelity and strong interpretability, with performance competitive with leading commercial closed-source models. This work provides the first systematic empirical validation of CoT’s effectiveness and cross-task transferability in NLU, establishing a new paradigm for interpretable NLU modeling.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Common-Sense Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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.

Problem

• Investigating rationales' impact on natural language understanding tasks
• Exploring rationale-augmented methods using constructed NLU dataset collection
• Evaluating rationale effectiveness across different model sizes and tasks

Innovation

• Chain-of-thought rationales enhance NLU model performance
• Specialized rationale-augmented training method improves NLU tasks
• Rationale-trained models outperform larger models on unseen tasks