On Teacher Hacking in Language Model Distillation
Daniil Tiapkin, Daniele Calandriello, Johan Ferret, Sarah Perrin, Nino Vieillard, Alexandre Ramé, Mathieu Blondel
TL;DR
The paper formalizes the notion of teacher hacking in language model distillation by contrasting an oracle-ground-truth distribution with a teacher and a student trained via soft distillation. It introduces golden (oracle–student) and proxy (teacher–student) metrics to track true-objective alignment versus proxy alignment, and uses a controlled, semi-synthetic setup to measure when distillation drifts from the ground truth. Experiments show that offline, fixed data can trigger teacher hacking (proxy improves while golden degrades), whereas online data generation preserves or improves both metrics; data diversity and larger generation budgets further mitigate hacking. The work provides practical guidelines for distillation pipelines to build robust, efficient LMs and highlights the risk of proxy optimization when relying on imperfect teachers. Overall, it contributes a diagnostic framework and actionable strategies to reduce adaptation to imperfect proxies in post-training LM pipelines.
Abstract
Post-training of language models (LMs) increasingly relies on the following two stages: (i) knowledge distillation, where the LM is trained to imitate a larger teacher LM, and (ii) reinforcement learning from human feedback (RLHF), where the LM is aligned by optimizing a reward model. In the second RLHF stage, a well-known challenge is reward hacking, where the LM over-optimizes the reward model. Such phenomenon is in line with Goodhart's law and can lead to degraded performance on the true objective. In this paper, we investigate whether a similar phenomenon, that we call teacher hacking, can occur during knowledge distillation. This could arise because the teacher LM is itself an imperfect approximation of the true distribution. To study this, we propose a controlled experimental setup involving: (i) an oracle LM representing the ground-truth distribution, (ii) a teacher LM distilled from the oracle, and (iii) a student LM distilled from the teacher. Our experiments reveal the following insights. When using a fixed offline dataset for distillation, teacher hacking occurs; moreover, we can detect it by observing when the optimization process deviates from polynomial convergence laws. In contrast, employing online data generation techniques effectively mitigates teacher hacking. More precisely, we identify data diversity as the key factor in preventing hacking. Overall, our findings provide a deeper understanding of the benefits and limitations of distillation for building robust and efficient LMs.