Searcharxiv⌕ Search

arXiv subjects

Xueyu Zhou

Publications and source records attributed to Xueyu Zhou.

6 recordsLinked to original sources

Bootstrapped On-Policy Distillation

On-policy distillation (OPD) improves language models by distilling teacher predictions on student-generated trajectories, while on-policy self-distillation (OPSD) removes the need for an external teacher by using a reference-conditioned copy of the model as the teacher. Despite its great early improvements, we find that OPSD does not remain consistently effective under continued training and can eventually suffer substantial performance degradation. We study this training behavior and link the degradation to a growing mismatch between the evolving reference-free student and the frozen privileged teacher. In the early stage of OPSD, the privileged teacher remains closely aligned with the student and provides useful guidance for improving the student's generation trajectories. However, as training progresses, the growing distributional mismatch between the privileged teacher and reference-free student can cause the teacher to provide misleading supervision, ultimately degrading student performance. This phenomenon motivates us to first leverage privileged reference supervision to acquire useful knowledge and then transfer it to a fresh student. We introduce Bootstrapped On-Policy Distillation (BOPD), a two-stage framework that first uses OPSD to internalize task-relevant information from reference-conditioned supervision into an intermediate model, and then distills this reference-free intermediate model into a fresh copy of the initial model. Experiments across models show that BOPD improves mathematical reasoning and code generation beyond both the initial models and the constructed OPSD teachers. These results suggest that privileged reference supervision can be more effectively exploited by first transferring its task-relevant signal into a task-adapted model and then distilling from that model without reference conditioning.

cs.LG↗

Self-Privileged Fine-Tuning

Supervised fine-tuning (SFT) is a key component of language model post-training, enabling effective task adaptation and serving as an important foundation for subsequent reinforcement learning. Recent methods seek to improve SFT by introducing token-level weights into the training objective, where these weights are typically derived from a model's output distribution at each token position. To better understand how such weighting schemes affect SFT, we analyze their optimization objectives and theoretically characterize the target distributions they induce. Our analysis shows that dynamic weighting methods, such as Dynamic Fine-Tuning (DFT), induce a heavily sharpened distribution, whereas static weighting methods fit a target distribution proportional to the product of the data distribution and a frozen prior. These findings suggest that static weighting provides a more stable training target, while its effectiveness critically depends on the quality of the prior distribution. Building on this observation, we propose Self-Privileged Fine-Tuning (SPFT), which uses a frozen copy of the initial model conditioned on privileged information (such as ground truth for mathematical questions) as a prior to provide guidance for token-level weighting during fine-tuning. Through a Bayesian decomposition, we further show that SPFT reweights the standard static prior according to how informative each target token is for predicting the ground-truth outcome, thereby assigning larger weights to more predictive tokens. Experiments on mathematical reasoning and code generation show that SPFT improves over standard SFT and other baselines across the main settings.

cs.LG↗

Representation-based Masked Diffusion Model

Masked Diffusion Models (MDMs) have emerged as a compelling paradigm for language modeling, offering the capability for efficient parallel text generation. However, existing parallel sampling methods typically update multiple masked tokens independently and ignore the complex mutual dependencies among the masked tokens. This independent updating mechanism lacks global coordination and might lead to incoherent outputs. To address this limitation, we propose Representation-based Masked Diffusion Model (RMDM), a framework that leverages the text representation to explicitly encode global semantics and help to parallel update tokens more precisely. Specifically, we first encode text into a continuous semantic space using a pretrained encoder and learn an invertible transformation that normalizes the representation distribution to a Gaussian prior, facilitating efficient sampling during generation. Conditioned on this latent semantic representation, we train a masked diffusion model to learn the conditional text distribution, where the representation serves as global semantic guidance to coordinate parallel token updates and faithfully approximate the target distribution. Empirical results demonstrate that RMDM significantly improves generation quality, particularly in aggressive few-step sampling regimes.

cs.CL↗

DOS: Dependency-Oriented Sampler for Masked Diffusion Language Models

Masked diffusion language models (MDLMs) have recently emerged as a new paradigm in language modeling, offering flexible generation dynamics and enabling efficient parallel decoding. However, existing decoding strategies for pre-trained MDLMs predominantly rely on token-level uncertainty criteria, while largely overlooking sequence-level information and inter-token dependencies. To address this limitation, we propose Dependency-Oriented Sampler (DOS), a training-free decoding strategy that leverages inter-token dependencies to inform token updates during generation. Specifically, DOS exploits attention matrices from transformer blocks to approximate inter-token dependencies, emphasizing information from unmasked tokens when updating masked positions. Empirical results demonstrate that DOS consistently achieves superior performance on both code generation and mathematical reasoning tasks. Moreover, DOS can be seamlessly integrated with existing parallel sampling methods, leading to improved generation efficiency without sacrificing generation quality.

cs.CL↗

Fair Sufficient Representation Learning

The main objective of fair statistical modeling and machine learning is to minimize or eliminate biases that may arise from the data or the model itself, ensuring that predictions and decisions are not unjustly influenced by sensitive attributes such as race, gender, age, or other protected characteristics. In this paper, we introduce a Fair Sufficient Representation Learning (FSRL) method that balances sufficiency and fairness. Sufficiency ensures that the representation should capture all necessary information about the target variables, while fairness requires that the learned representation remains independent of sensitive attributes. FSRL is based on a convex combination of an objective function for learning a sufficient representation and an objective function that ensures fairness. Our approach manages fairness and sufficiency at the representation level, offering a novel perspective on fair representation learning. We implement this method using distance covariance, which is effective for characterizing independence between random variables. We establish the convergence properties of the learned representations. Experiments conducted on healthcase and text datasets with diverse structures demonstrate that FSRL achieves a superior trade-off between fairness and accuracy compared to existing approaches.

stat.ML↗

Transfer Learning through Enhanced Sufficient Representation: Enriching Source Domain Knowledge with Target Data

Transfer learning is an important approach for addressing the challenges posed by limited data availability in various applications. It accomplishes this by transferring knowledge from well-established source domains to a less familiar target domain. However, traditional transfer learning methods often face difficulties due to rigid model assumptions and the need for a high degree of similarity between source and target domain models. In this paper, we introduce a novel method for transfer learning called Transfer learning through Enhanced Sufficient Representation (TESR). Our approach begins by estimating a sufficient and invariant representation from the source domains. This representation is then enhanced with an independent component derived from the target data, ensuring that it is sufficient for the target domain and adaptable to its specific characteristics. A notable advantage of TESR is that it does not rely on assuming similar model structures across different tasks. For example, the source domain models can be regression models, while the target domain task can be classification. This flexibility makes TESR applicable to a wide range of supervised learning problems. We explore the theoretical properties of TESR and validate its performance through simulation studies and real-world data applications, demonstrating its effectiveness in finite sample settings.

stat.ML↗