Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Distillation Traps and Guards: A Calibration Knob for LLM Distillability

About

Knowledge distillation (KD) transfers capabilities from large language models (LLMs) to smaller students, yet it can fail unpredictably and also underpins model leakage risks. Our analysis revealed several distillation traps: tail noise, off-policy instability, and, most fundamentally, the teacher-student gap, that distort training signals. These traps manifest as overconfident hallucinations, self-correction collapse, and local decoding degradation, causing distillation to fail. Motivated by these findings, we propose a post-hoc calibration method that, to the best of our knowledge, for the first time enables control over a teacher's distillability via reinforcement fine-tuning (RFT). Our objective combines task utility, KL anchor, and across-tokenizer calibration reward. This makes distillability a practical safety lever for foundation models, connecting robust teacher-student transfer with deployment-aware model protection. Experiments across math, knowledge QA, and instruction-following tasks show that students distilled from distillable calibrated teachers outperform SFT and KD baselines, while undistillable calibrated teachers retain their task performance but cause distilled students to collapse, offering a practical knob for both better KD and model IP protection.

Weixiao Zhan, Yongcheng Jing, Leszek Rutkowski, Dacheng Tao• 2026

Related benchmarks

TaskDatasetResultRank
General Knowledge QACSQA
Accuracy84.1
18
General Knowledge QAMMLU-Pro
Accuracy69.4
18
General Knowledge QASuperGPQA
Average Accuracy33.6
18
Math ReasoningBigMath level 4
Accuracy66.3
18
Math ReasoningBigMath level 5
Average Accuracy43.1
18
Open-ended generationDolly
Skywork Reward V2 Score0.961
18
Open-ended generationVicuna
Skywork Reward V2 Score99
18
Showing 7 of 7 rows

Other info

Follow for update