X. T. Chen

Xi Chen, Haosen Yang, Huicong Zhang et al.

Source-free unsupervised domain adaptation (SFUDA) aims to enable the utilization of a pre-trained source model in an unlabeled target domain without access to source data. Self-training is a way to solve SFUDA, where confident target samples are iteratively selected as pseudo-labeled samples to guide target model learning. However, prior heuristic noisy pseudo-label filtering methods all involve introducing extra models, which are sensitive to model assumptions and may introduce additional errors or mislabeling. In this work, we propose a method called Uncertainty-aware Pseudo-label-filtering Adaptation (UPA) to efficiently address this issue in a coarse-to-fine manner. Specially, we first introduce a sample selection module named Adaptive Pseudo-label Selection (APS), which is responsible for filtering noisy pseudo labels. The APS utilizes a simple sample uncertainty estimation method by aggregating knowledge from neighboring samples and confident samples are selected as clean pseudo-labeled. Additionally, we incorporate Class-Aware Contrastive Learning (CACL) to mitigate the memorization of pseudo-label noise by learning robust pair-wise representation supervised by pseudo labels. Through extensive experiments conducted on three widely used benchmarks, we demonstrate that our proposed method achieves competitive performance on par with state-of-the-art SFUDA methods. Code is available at https://github.com/chenxi52/UPA.

Mingkang Zhu, Xi Chen, Bei Yu et al.

Large reasoning models (LRMs) generate intermediate reasoning traces before producing final answers, yielding strong gains on multi-step and mathematical tasks. Yet aligning LRMs with human preferences, a crucial prerequisite for model deployment, remains underexplored. The statistically correct objective for preference alignment requires marginalizing over reasoning traces, but this computation is intractable in practice. A common workaround optimizes a single sampled trajectory, which introduces substantial gradient variance from stochastic trace sampling. To address this challenge, we frame preference optimization for LRMs through the lens of the bias--variance trade-off and propose Bias--Variance Optimized Preference Optimization (BVPO), a simple, drop-in method that mixes two gradient estimators: a high-variance trace-based estimator and a low-variance empty-trace estimator obtained by disabling reasoning trace generation. Our theory shows that BVPO strictly reduces trace-induced variance for any nontrivial mixture, provides a closed-form choice of the mixing weight that minimizes mean-squared error relative to the true marginal gradient, and under standard smoothness and step-size conditions, tightens classical convergence bounds for stochastic gradient descent. Empirically, BVPO improves alignment over the best baseline by up to 7.8 points on AlpacaEval~2 and 6.8 points on Arena-Hard. Despite being trained only on general conversational data, BVPO also boosts reasoning performance for base models by up to 4.0 points on the average of six math reasoning benchmarks. These results identify variance from trace sampling as a key bottleneck and demonstrate that directly optimizing the bias--variance trade-off yields more stable training and stronger overall performance.