Continued Training of Continuous Language Flows with Token Identity Squared Supervision
Jiawei Jiang, Zhaoxiang Zang, Zhengtao Yu
原文摘要
Diffusion and flow models have recently been extended to language modeling, generating text through continuous denoising in token embedding spaces. Pretrained models provide token representations and contextual prediction capabilities, but improving generation with limited additional training requires an appropriate connection between supervision and continuous generation. When continuous predictions are formed by weighting token embeddings with predicted probabilities, different probability distributions can yield the same vector and are therefore indistinguishable under a continuous reconstruction objective based on mean squared error (MSE). We propose Token Identity Brier Flow (TIBF) to apply squared supervision directly to token identities. TIBF fixes token embeddings, combines network scores with token-matching scores under a Gaussian observation model, supervises token probabilities using the Brier loss, and uses probability-weighted embeddings for continuous generation. We show that, for fixed token embeddings, the squared error in token posterior prediction controls the squared errors of the continuous denoising prediction and the flow velocity. A risk decomposition separates uncertainty inherent in the observation from posterior fitting error, motivating continued training that preserves pretrained embeddings and focuses on probability prediction. From common pretrained initializations and under matched continuation budgets, TIBF achieves mean Gen.PPL values of 34.87 on OWT and 80.08 on LM1B, outperforming continuous-MSE and CE controls. Further evaluations show improvements across multiple sampling budgets. Code is available at https://github.com/Jiangjiawei-kust/TIBF.
cs.CL · cs.AI