Knowledge Distillation for Reasoning

Distilling chain-of-thought reasoning into smaller models

Performed knowledge distillation of large language models into smaller models while preserving chain-of-thought and multi-step reasoning capabilities. Analyzed generalization, calibration, and faithfulness of reasoning traces across model scales.

Technologies: PyTorch, Hugging Face, LLaMA 3
Duration: Nov 2025 - Dec 2025

Objectives

  • Transfer complex reasoning abilities from large models to more efficient smaller models
  • Preserve the interpretability and correctness of reasoning chains
  • Analyze trade-offs between model size and reasoning quality

Findings

  • Successfully distilled reasoning capabilities with minimal performance degradation
  • Identified key factors affecting reasoning faithfulness across model scales
  • Developed metrics for evaluating reasoning trace quality

This research contributes to making advanced AI reasoning more accessible and deployable on resource-constrained devices.