New method improves language model training with unbiased top-k token use
Unbiased Top-$k$ Estimation for On-Policy Distillation
Computation and LanguageMachine Learning
Summary
Training smaller language models to learn reasoning from bigger ones often involves comparing probabilities of different words. Using only a few top candidate words speeds up this training but can make the learning biased and less accurate. The authors propose a method that keeps the speed advantage but fixes this bias by including the actually chosen word in the training calculation. This change helps smaller models learn better from big models without extra cost.
What this means in practice
- •For ai platform engineers: Improve the efficiency and accuracy of training smaller language models by using an unbiased estimator that balances computation and learning quality.
- •For natural language processing teams: Enhance model distillation pipelines to better transfer reasoning capabilities from large teacher models to smaller student models without sacrificing accuracy.
Authors
Linjian Meng, Siyuan Gan, YuHan Li, Xiran Wang, Ziyang Ding, Ditang Gou, Yiming Wu, Zhen Zhao
Abstract
On-policy distillation (OPD) is becoming an important component of large language model (LLM) post-training for transferring the reasoning capability of a strong teacher LLM to a weaker student LLM. OPD trains the student by minimizing the reverse KL divergence between the teacher and the student via rollouts generated by the student's policy. However, estimating the gradient of the reverse KL divergence in OPD remains a challenge. Using only the sampled token from the student-generated rollout is computationally cheap but provides limited distributional supervision, which will degrade accuracy. In addition, using the full vocabulary provides complete distributional supervision but is computationally expensive. Therefore, recent works propose Top-$k$ OPD (TK-OPD) that use selected top-$k$ tokens, which provides richer distributional supervision than sampled-token estimation at substantially lower computational cost than full-vocabulary estimation. Unfortunately, using only the selected top-$k$ tokens induces bias, leading to accuracy degradation, as the probability mass outside the selected top-$k$ tokens is discarded. To address the bias of TK-OPD, we propose Tail-Corrected Top-$k$ On-Policy Distillation (TT-OPD). It preserves the advantages of TK-OPD, including rich distributional supervision and low computational cost, while providing an unbiased estimator of the gradient of the reverse KL divergence. The key insight of TT-OPD is to use not only the selected top-$k$ tokens, but also the sampled token from the student-generated rollout, thereby recovering the discarded probability mass in expectation, avoiding the bias. Experimental results demonstrate that TT-OPD significantly outperforms other tested OPD variants.