Research map / Large language models / Instruction and fine-tuning
Critique-Guided Distillation for Robust Reasoning via Refinement
Berkcan Kapusuzoglu, Supriyo Chakraborty, Zain Sarwar, Chia-Hsuan Lee, Sambit Sahu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train student to refine answers conditioned on teacher critiques (CGD) instead of generating critiques (CFT) | LLaMA3.1-8B Instruct · Math reasoning avg (Group 1) | accuracy: 41.5 -> 46.9 | Improved |
| train with CGD instead of CFT | S1.1-3B · Math reasoning avg (Group 1) | accuracy: 38.9 -> 46.1 | Improved |
| use CGD instead of CFT on instruction-following | LLaMA3.1-8B Instruct · IFEval | accuracy: 55.6 -> 76.1 | Improved |
| add critique conditioning during training (ablation) | CGD · 8 math and reasoning benchmarks | accuracy: up to 114% relative drop without critique | Improved |
| use specific relevant critiques instead of a generic string (ablation) | CGD · AIME24 (10K subset) | accuracy: 7.8 -> 12.2 | Improved |
| use irrelevant critiques instead of generic ones (ablation) | CGD · AMC23 (10K subset) | accuracy: 26.7 -> 18.3 | Hurt |
| mix 50/50 correct and incorrect student answers in training data (ablation) | CGD · Average accuracy, 25K data | accuracy: 36.8 -> 39.9 | Improved |
| raise learning rate from 1e-6 to 5e-6 (ablation) | CFT · Average over 6 benchmarks | accuracy: 39.3 -> 30.1 | Hurt |
| raise learning rate from 1e-6 to 5e-6 (ablation) | CGD · Average over 6 benchmarks | accuracy: 44.8 -> 42.2 | Hurt |
| apply CGD to LLaMA3.1-8B Instruct | LLaMA3.1-8B Instruct · AIME 2024 greedy | Pass@1: 3.3% -> 16.7% | Improved |
| swap LLaMA3.3-70B teacher for S1.1-32B teacher (ablation) | CGD on LLaMA3.1-8B Instruct · WebInstruct, math and general avg | accuracy: Math avg 46.9 -> 48.0; General avg 36.7 -> 35.7 | Mixed |
| use weaker S1.1-32B teacher instead of Claude Sonnet 3.7 (ablation) | CGD on Qwen2.5-Math-7B · Math avg | accuracy: 50.4 -> 49.0 | No clear effect |
Closest papers by meaning
- Decoding the Critique Mechanism in Large Reasoning Models (NeurIPS 2026 · similarity 0.90)
- Critique-Coder: Enhancing Coder Models by Critique Reinforcement Learning (ICLR 2026 · similarity 0.89)
- Collaborative Reasoning Distillation via Cross-Feedback and Coherent Curation (NeurIPS 2026 · similarity 0.89)
- Memorize Theorems, Not Instances: Probing SFT Generalization through Mathematical Reasoning (NeurIPS 2026 · similarity 0.89)
- Reasoning-Trace Collapse: Evaluating the Loss of Explicit Reasoning During Fine-Tuning (NeurIPS 2026 · similarity 0.89)
- Hint Tuning: Less Data Makes Better Reasoners (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
