Token-Level Uncertainty-Aware Objective for Language Model Post-Training

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Liu, Tingkai, Benjamin, Ari S., Zador, Anthony M.
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866917963571920896
author Liu, Tingkai
Benjamin, Ari S.
Zador, Anthony M.
author_facet Liu, Tingkai
Benjamin, Ari S.
Zador, Anthony M.
contents In the current work, we connect token-level uncertainty in causal language modeling to two types of training objectives: 1) masked maximum likelihood (MLE), 2) self-distillation. We show that masked MLE is effective in reducing epistemic uncertainty, and serve as an effective token-level automatic curriculum learning technique. However, masked MLE is prone to overfitting and requires self-distillation regularization to improve or maintain performance on out-of-distribution tasks. We demonstrate significant performance gain via the proposed training objective - combined masked MLE and self-distillation - across multiple architectures (Gemma, LLaMA, Phi) and datasets (Alpaca, ShareGPT, GSM8K), mitigating overfitting while maintaining adaptability during post-training. Our findings suggest that uncertainty-aware training provides an effective mechanism for enhancing language model training.
format Preprint
id arxiv_https___arxiv_org_abs_2503_16511
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Token-Level Uncertainty-Aware Objective for Language Model Post-Training
Liu, Tingkai
Benjamin, Ari S.
Zador, Anthony M.
Computation and Language
Artificial Intelligence
In the current work, we connect token-level uncertainty in causal language modeling to two types of training objectives: 1) masked maximum likelihood (MLE), 2) self-distillation. We show that masked MLE is effective in reducing epistemic uncertainty, and serve as an effective token-level automatic curriculum learning technique. However, masked MLE is prone to overfitting and requires self-distillation regularization to improve or maintain performance on out-of-distribution tasks. We demonstrate significant performance gain via the proposed training objective - combined masked MLE and self-distillation - across multiple architectures (Gemma, LLaMA, Phi) and datasets (Alpaca, ShareGPT, GSM8K), mitigating overfitting while maintaining adaptability during post-training. Our findings suggest that uncertainty-aware training provides an effective mechanism for enhancing language model training.
title Token-Level Uncertainty-Aware Objective for Language Model Post-Training
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2503.16511