Large language model distillation reduces teacher bias without target feedback
Coupled Calibration and Learning: Mitigating Teacher Bias in LLM Distillation without Target-Domain Reward Feedback
Machine LearningArtificial Intelligence
Summary
When a big language model (teacher) helps train a smaller one (student), copying the teacher can pass along its mistakes and biases. This problem gets harder when the smaller model faces new types of questions without feedback on how well it’s doing. The authors propose a method where the teacher’s advice is adjusted using feedback from familiar questions, then used to guide the student on new questions. Over time, the student and teacher improve together, reducing bias even without specific feedback from the new question types.
What this means in practice
- •For language model developers: Train smaller language models more reliably by correcting teacher biases with source feedback even when target feedback is missing.
- •For machine learning engineers: Improve model compression workflows by iteratively calibrating teacher models to produce unbiased training signals for students under domain shifts.
Authors
Haichen Hu, Yuheng Zhang, David Simchi-Levi
Abstract
Large language model (LLM) distillation aims to transfer the capabilities of a powerful teacher to a smaller student. Direct imitation, however, can also transfer the teacher's systematic bias and errors. This challenge is particularly pronounced under covariate shift, when the teacher's reliability on target questions is uncertain and target-domain reward feedback is unavailable. We propose Coupled Calibration and Learning (CCL), an LLM distillation algorithm that couples teacher calibration with student updates through token-level branching, using reward feedback only on source questions. Each iteration calibrates the teacher using source feedback and then uses the calibrated teacher to train the student on target questions. The updated student, in turn, informs subsequent calibration. In an autoregressive policy framework, we prove that the output student's expected average Kullback-Leibler divergence to the oracle student converges to zero at a polynomial rate in the number of iterations. The oracle maximizes the true reference-regularized target reward within the student class, which need not represent the unrestricted optimal policy. Our analysis quantifies the progress of projected student gradient updates while controlling the error in teacher calibration. We further establish a separation from regularized direct matching: its error relative to the oracle student can remain bounded away from zero even when the teacher achieves higher regularized target reward than every student policy. These results demonstrate that LLM distillation can overcome persistent teacher bias and recover the optimal student through coupled calibration and learning, without target-domain reward feedback.