Coupled Calibration and Learning algorithm for LLM distillation
Coupled Calibration and Learning: Mitigating Teacher Bias in LLM Distillation without Target-Domain Reward Feedback
This research provides a new method for improving LLM distillation, which is a key technique for creating smaller, more efficient AI models. The algorithm is particularly useful for applications where target-domain reward feedback is unavailable.
A new algorithm called Coupled Calibration and Learning (CCL) is proposed to mitigate teacher bias in large language model (LLM) distillation.
It works by coupling teacher calibration with student updates through token-level branching, using reward feedback only on source questions.
Each iteration calibrates the teacher with source feedback and then trains the student on target questions with the calibrated teacher.
The updated student informs subsequent calibration.The algorithm is proven to converge to the optimal student at a polynomial rate without needing target-domain reward feedback.
Coupled Calibration and Learning: Mitigating Teacher Bias in LLM Distillation without Target-Domain Reward Feedback
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.
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include LLM Distillation, prompt engineering.
Source attribution: curated from arXiv cs.AI (@Haichen Hu, Yuheng Zhang, David Simchi-Levi).