Student conditioned reasoning traces improve question answering accuracy
Beyond the Answer: Student-Conditioned Trace Construction for Reasoning Distillation
Computation and LanguageMachine Learning
Summary
Sometimes giving a machine just the right answer isn’t enough for it to learn well. This paper shows that providing step-by-step reasoning that's tailored to the machine's own attempts can help it learn better. The authors created a system called ReDiST that improves reasoning by editing and selecting helpful explanations suited to the machine’s current skill level. Tests show this approach helps the machine answer questions more accurately, even on different types of questions than it trained on.
What this means in practice
- •For natural language processing engineers: Improve training of question-answering models by using student-specific reasoning traces rather than only final answers.
- •For ai customer support teams: Enhance automated response systems by distilling explanations that better match the AI’s reasoning capabilities, improving answer reliability.
Authors
Shaoang Li, Yanhang Shi, Yufei Li, Mingfu Liang, Frank Shyu, Sandeep Pandey, Luke Simon, Xi Liu, Jian Li
Abstract
A reasoning trace can accompany a correct answer without providing effective supervision for a particular student. We study how to construct traces that improve learning beyond answer-only supervision by considering both their intermediate content and their accessibility to the target learner. We introduce ReDiST, a student-conditioned trace-construction and distillation framework for evidence-rich reasoning. Starting from the student's own attempts, ReDiST revises answer-correct drafts against the supplied evidence and repairs unsuccessful drafts with additional guidance from the verified answer. A frozen student-matched reader screens and ranks original and edited candidates using answer recovery and verified-answer probability, given the question, evidence, and candidate trace. We distill the selected targets while balancing trace and final-answer supervision with an auxiliary answer loss. Construction and selection occur offline, and inference uses only the distilled student. Experiments on question-answering benchmarks show that ReDiST can improve both source-task and cross-benchmark performance over answer-only, original-trace, and the evaluated trace-selection and rewriting baselines. Across supervision recipes, performance rankings can differ between source and transfer benchmarks, highlighting the importance of evaluating trace teaching utility across distributions.