Beyond the Answer: Student-Conditioned Trace Construction for Reasoning Distillation

TL;DR

ReDiST is a framework that generates personalized reasoning traces by refining student attempts using evidence and verified answers, then selects the most effective traces via a frozen reader model. It improves learning by aligning trace content and structure with student-level understanding, outperforming answer-only and baseline trace methods on multiple benchmarks.

Why It Matters

Personalized reasoning traces can significantly enhance model learning, especially in educational and reasoning tasks where understanding the thought process matters. This approach enables better knowledge transfer across domains by tailoring instruction to individual learners' needs.

Key Findings

• ReDiST improves both source-task and cross-benchmark performance compared to answer-only and original-trace baselines.
• Trace selection using answer recovery and verified-answer probability leads to more effective distillation than unguided or static trace methods.
• Performance rankings vary across benchmarks, indicating that trace utility depends on data distribution and student-specific learning patterns.