Calibrating LLMs with Preference Optimization on Thought Trees for Generating Rationale in Science Question Scoring

TL;DR

This work proposes a novel framework capable of generating more faithful rationales and, more importantly, matching performance with classifier-based black-box scoring systems, and sheds light on the effectiveness of performing preference optimization using synthetic preference data obtained from thought tree paths.

Abstract

Generating rationales that justify scoring decisions has been a promising way to facilitate explainability in automated scoring systems. However, existing methods do not match the accuracy of classifier-based methods. Plus, the generated rationales often contain hallucinated information. To address these issues, we propose a novel framework capable of generating more faithful rationales and, more importantly, matching performance with classifier-based black-box scoring systems. We first mimic the human assessment process by querying Large Language Models (LLMs) to generate a thought tree. We then summarise intermediate assessment decisions from each thought tree path for creating synthetic rationale data and rationale preference data. Finally, we utilise the generated synthetic data to calibrate LLMs through a two-step training process: supervised fine-tuning and preference optimization. Extensive experimental results demonstrate that our framework achieves a 38% assessment performance improvement in the QWK score compared to prior work while producing higher-quality rationales, as recognised by human evaluators and LLMs. Our work sheds light on the effectiveness of performing preference optimization using synthetic preference data obtained from thought tree paths. Data and code are available at https://github.com/lijiazheng99/thought_tree_assessment.

Figures (9)

• Figure 1: Illustration of question information used in (a) Human assessment and (b) Traditional PLM-based text classifier assessment approaches. We propose (c) Imitating the human assessment procedure by prompting LLMs to generate thought trees that form rationales.
• Figure 2: Overview of Our Three-stage Thought Tree Guided Rationale Generation Framework.
• Figure 3: Illustration of the structure of a thought tree.
• Figure 4: Normalized confusion matrix for Mixtral 8$\times$7B SFT (left) and Mixtral 8$\times$7B DPO (right).
• Figure A1: Explanation of our prompt structure using querying decision 2 as an example.