arXiv subjects

Xiran Wang

Publications and source records attributed to Xiran Wang.

8 recordsLinked to original sources

arXiv2026

When Teacher Guidance Misleads: Reward-Aligned On-Policy Distillation

On-policy distillation (OPD) has recently emerged as a popular post-training paradigm for large language models (LLMs), providing an efficient way to transfer the knowledge and capabilities of teacher models into student models. However, teacher guidance on student-generated prefixes is not always reliable. Training should optimize the model to generate responses that are more likely to be correct, or equivalently, to get higher outcome rewards. But during OPD, the teacher model may provide guidance that discourages the student from moving toward correct trajectories or moves the student toward incorrect ones, which is misaligned with outcome reward. Such misaligned guidance is unreliable, as it would mislead the optimization process and ultimately degrade model performance. To mitigate misaligned teacher guidance, we propose Reward-Aligned On-Policy Distillation (RA-OPD). The key insight is to keep only trajectories whose induced updates move the student toward correct trajectories or discourage the student from moving toward incorrect ones. Specifically, for each sampled trajectory, RA-OPD checks whether its trajectory-level distillation return is consistent with its outcome reward and then filters out the misaligned trajectories. RA-OPD selects more reliable trajectories to improve student model performance without requiring additional computational cost. We evaluate RA-OPD on math and code benchmarks using models from the Qwen3 family and the DeepSeek-R1 family. Across seven math benchmarks and three code benchmarks, RA-OPD significantly outperforms standard OPD and other tested OPD variants.

cs.AI↗

arXiv2026

Learning from Hard Prompts: Difficulty-aware Advantage Amplification in Dynamic Sampling

Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO) is a prominent variant of Group Relative Policy Optimization (GRPO). DAPO introduces several improvements over GRPO. Among these, Dynamic Sampling contributes the most to DAPO's accuracy gains relative to GRPO. To improve accuracy, Dynamic Sampling enhances training stability by eliminating zero policy gradients from zero advantages. Specifically, it avoids such zero gradients by filtering out prompts where sampled responses are either entirely correct or incorrect. However, our theoretical analysis shows that Dynamic Sampling decrease training efficiency as it cannot effectively utilize hard-to-sample correct responses on hard prompts. Formally, it asymmetrically amplifies the advantages of distinct responses to the same prompts. On hard prompts, incorrect responses undergo greater amplification than correct ones. This leads the model to avoid generating the observed incorrect responses rather than capitalizing on the hard-to-sample correct ones on hard prompts, resulting in low training efficiency. To improve training efficiency, we propose Direct Advantage Amplification (DAA), which amplifies the advantages of hard-to-sample correct responses on hard prompts, as obtained by Dynamic Sampling. This ensures that, when Dynamic Sampling is used, these hard-to-sample responses can be effectively capitalized on, implying higher training efficiency. By integrating DAA into DAPO, we obtain Difficulty-aware Advantage Amplification Policy Optimization (DA3PO), which is implemented with fewer than 30 lines of code from DAPO. Experiments show that DA3PO significantly outperforms GRPO and other classical GRPO variants.

cs.AI↗

arXiv2026

Exploring Dualistic Meta-Learning to Enhance Domain Generalization in Open Set Scenarios

Domain generalization learns from multiple source domains to generalize to unseen target domains. However, it often neglects the realistic case of label mismatch between source and target. Open set domain generalization is then proposed to recognize unseen classes in unseen domains. A simple approach trains one-vs-all classifiers to separate each class and detect outliers as unknown. Yet, the imbalance between few positive samples and many negative samples skews the decision boundary towards the positive ones, leading the model to over-reject out-of-distribution data, even from known classes in unseen domains. In this paper, we propose a novel meta-learning stategy called dualistic MEta-learning with joint DomaIn-Class matching (MEDIC), which considers implicit gradient matching towards inter-domain and inter-class task splits simultaneously to find optimal boundaries balanced for both domains and classes. Experimental results show that MEDIC not only outperforms prior methods in open set scenarios, but also maintains competitive close set generalization ability.

cs.LG↗

arXiv2026

JoyAI-LLM Flash: Advancing Mid-Scale LLMs with Token Efficiency

We introduce JoyAI-LLM Flash, an efficient Mixture-of-Experts (MoE) language model designed to redefine the trade-off between strong performance and token efficiency in the sub-50B parameter regime. JoyAI-LLM Flash is pretrained on a massive corpus of 20 trillion tokens and further optimized through a rigorous post-training pipeline, including supervised fine-tuning (SFT), Direct Preference Optimization (DPO), and large-scale reinforcement learning (RL) across diverse environments. To improve token efficiency, JoyAI-LLM Flash strategically balances \emph{thinking} and \emph{non-thinking} cognitive modes and introduces FiberPO, a novel RL algorithm inspired by fibration theory that decomposes trust-region maintenance into global and local components, providing unified multi-scale stability control for LLM policy optimization. To enhance architectural sparsity, the model comprises 48B total parameters while activating only 2.7B parameters per forward pass, achieving a substantially higher sparsity ratio than contemporary industry leading models of comparable scale. To further improve inference throughput, we adopt a joint training-inference co-design that incorporates dense Multi-Token Prediction (MTP) and Quantization-Aware Training (QAT). We release the checkpoints for both JoyAI-LLM-48B-A3B Base and its post-trained variants on Hugging Face to support the open-source community.

cs.CL↗

arXiv2025

Balanced Direction from Multifarious Choices: Arithmetic Meta-Learning for Domain Generalization

Domain generalization is proposed to address distribution shift, arising from statistical disparities between training source and unseen target domains. The widely used first-order meta-learning algorithms demonstrate strong performance for domain generalization by leveraging the gradient matching theory, which aims to establish balanced parameters across source domains to reduce overfitting to any particular domain. However, our analysis reveals that there are actually numerous directions to achieve gradient matching, with current methods representing just one possible path. These methods actually overlook another critical factor that the balanced parameters should be close to the centroid of optimal parameters of each source domain. To address this, we propose a simple yet effective arithmetic meta-learning with arithmetic-weighted gradients. This approach, while adhering to the principles of gradient matching, promotes a more precise balance by estimating the centroid between domain-specific optimal parameters. Experimental results validate the effectiveness of our strategy.

cs.LG↗

arXiv2023

Generalizable Decision Boundaries: Dualistic Meta-Learning for Open Set Domain Generalization

Domain generalization (DG) is proposed to deal with the issue of domain shift, which occurs when statistical differences exist between source and target domains. However, most current methods do not account for a common realistic scenario where the source and target domains have different classes. To overcome this deficiency, open set domain generalization (OSDG) then emerges as a more practical setting to recognize unseen classes in unseen domains. An intuitive approach is to use multiple one-vs-all classifiers to define decision boundaries for each class and reject the outliers as unknown. However, the significant class imbalance between positive and negative samples often causes the boundaries biased towards positive ones, resulting in misclassification for known samples in the unseen target domain. In this paper, we propose a novel meta-learning-based framework called dualistic MEta-learning with joint DomaIn-Class matching (MEDIC), which considers gradient matching towards inter-domain and inter-class splits simultaneously to find a generalizable boundary balanced for all tasks. Experimental results demonstrate that MEDIC not only outperforms previous methods in open set scenarios, but also maintains competitive close set generalization ability at the same time. Our code is available at https://github.com/zzwdx/MEDIC.

cs.CV↗

arXiv2020

Automatic Foreground Extraction from Imperfect Backgrounds using Multi-Agent Consensus Equilibrium

Extracting accurate foreground objects from a scene is an essential step for many video applications. Traditional background subtraction algorithms can generate coarse estimates, but generating high quality masks requires professional softwares with significant human interventions, e.g., providing trimaps or labeling key frames. We propose an automatic foreground extraction method in applications where a static but imperfect background is available. Examples include filming and surveillance where the background can be captured before the objects enter the scene or after they leave the scene. Our proposed method is very robust and produces significantly better estimates than state-of-the-art background subtraction, video segmentation and alpha matting methods. The key innovation of our method is a novel information fusion technique. The fusion framework allows us to integrate the individual strengths of alpha matting, background subtraction and image denoising to produce an overall better estimate. Such integration is particularly important when handling complex scenes with imperfect background. We show how the framework is developed, and how the individual components are built. Extensive experiments and ablation studies are conducted to evaluate the proposed method.

cs.CV↗

arXiv2016

Plug-and-Play ADMM for Image Restoration: Fixed Point Convergence and Applications

Alternating direction method of multiplier (ADMM) is a widely used algorithm for solving constrained optimization problems in image restoration. Among many useful features, one critical feature of the ADMM algorithm is its modular structure which allows one to plug in any off-the-shelf image denoising algorithm for a subproblem in the ADMM algorithm. Because of the plug-in nature, this type of ADMM algorithms is coined the name "Plug-and-Play ADMM". Plug-and-Play ADMM has demonstrated promising empirical results in a number of recent papers. However, it is unclear under what conditions and by using what denoising algorithms would it guarantee convergence. Also, since Plug-and-Play ADMM uses a specific way to split the variables, it is unclear if fast implementation can be made for common Gaussian and Poissonian image restoration problems. In this paper, we propose a Plug-and-Play ADMM algorithm with provable fixed point convergence. We show that for any denoising algorithm satisfying an asymptotic criteria, called bounded denoisers, Plug-and-Play ADMM converges to a fixed point under a continuation scheme. We also present fast implementations for two image restoration problems on super-resolution and single-photon imaging. We compare Plug-and-Play ADMM with state-of-the-art algorithms in each problem type, and demonstrate promising experimental results of the algorithm.

cs.CV↗