Zirui Zhu

Yang Luo, Shengju Qian, Xiaohang Tang et al.

Autoregressive video generators are attractive for streaming, long-horizon, and interactive applications, but distilling strong black-box teachers into causal students remains difficult. The student must learn under its own rollout distribution, whereas practical teachers may expose only prompt-conditioned completed videos and may differ in architecture, capacity, temporal design, and sampling schedule. This interface makes supervised fine-tuning off-policy, score-based distillation inapplicable, and direct adversarial imitation too sparse for denoising-time credit assignment. We propose Adversarial Flow Distillation (AFD), an on-policy framework for heterogeneous black-box video distillation. AFD queries the teacher and rolls out the current student on the same prompts, trains a prompt-paired Bradley-Terry discriminator to estimate clean-sample teacher-student discrepancy, and converts the resulting on-policy advantage into forward-process flow-matching updates on the student's own noised states. Thus, AFD provides dense velocity-field supervision while requiring no teacher scores, latents, denoising trajectories, step alignment, or reverse-chain reinforcement learning. Experiments across two causal AR student families show that AFD consistently improves motion- and physics-sensitive generation while preserving general video quality, and ablations validate the importance of adaptive on-policy feedback and forward-process credit assignment. The method requires only clean teacher videos and student rollouts, providing a practical route for distilling proprietary or heterogeneous video generators into efficient autoregressive students.

Yong Liu, Di Fu, Yang Luo et al.

We introduce Post-Optimization Model Edit (POME), a new algorithm that enhances the performance of fine-tuned large language models using only their pretrained and fine-tuned checkpoints, without requiring extra data or further optimization. The core idea is to apply a muon-style projection to $ΔW$, the difference between the fine-tuned and pretrained weights. This projection uses truncated singular value decomposition (SVD) to equalize the influence of dominant update directions and prune small singular values, which often represent noise. As a simple post-processing step, POME is completely decoupled from the training pipeline. It requires zero modifications and imposes no overhead, making it universally compatible with any optimizer or distributed framework. POME delivers consistent gains, boosting average performance by +2.5\% on GSM8K and +1.0\% on code generation. Its broad applicability -- from 7B foundation models to 72B RLHF-instructed models -- establishes it as a practical, zero-cost enhancement for any fine-tuning pipeline. Code is available at https://github.com/NUS-HPC-AI-Lab/POME.

Zirui Zhu, Chen Gao, Xu Chen et al.

Incorporating social relations into the recommendation system, i.e. social recommendation, has been widely studied in academic and industrial communities. While many promising results have been achieved, existing methods mostly assume that the social relations can be homogeneously applied to all the items, which is not practical for users' actually diverse preferences. In this paper, we argue that the effect of the social relations should be inhomogeneous, that is, two socially-related users may only share the same preference on some specific items, while for the other products, their preferences can be inconsistent or even contradictory. Inspired by this idea, we build a novel social recommendation model, where the traditional pair-wise "user-user'' relation is extended to the triple relation of "user-item-user''. To well handle such high-order relations, we base our framework on the hypergraph. More specifically, each hyperedge connects a user-user-item triplet, representing that the two users share similar preferences on the item. We develop a Social HyperGraph Convolutional Network (short for SHGCN) to learn from the complex triplet social relations. With the hypergraph convolutional networks, the social relations can be modeled in a more fine-grained manner, which more accurately depicts real users' preferences, and benefits the recommendation performance. Extensive experiments on two real-world datasets demonstrate our model's effectiveness. Studies on data sparsity and hyper-parameter studies further validate our model's rationality. Our codes and dataset are available at https://github.com/ziruizhu/SHGCN.

Zirui Zhu, Hailun Xu, Yang Luo et al.

Multimodal large language models (MLLMs) represent images and video frames as visual tokens. Scaling from single images to hour-long videos, however, inflates the token budget far beyond practical limits. Popular pipelines therefore either uniformly subsample or apply keyframe selection with retrieval-style scoring using smaller vision-language models. However, these keyframe selection methods still rely on pre-filtering before selection to reduce the inference cost and can miss the most informative moments. We propose FOCUS, Frame-Optimistic Confidence Upper-bound Selection, a training-free, model-agnostic keyframe selection module that selects query-relevant frames under a strict token budget. FOCUS formulates keyframe selection as a combinatorial pure-exploration (CPE) problem in multi-armed bandits: it treats short temporal clips as arms, and uses empirical means and Bernstein confidence radius to identify informative regions while preserving exploration of uncertain areas. The resulting two-stage exploration-exploitation procedure reduces from a sequential policy with theoretical guarantees, first identifying high-value temporal regions, then selecting top-scoring frames within each region On two long-video question-answering benchmarks, FOCUS delivers substantial accuracy improvements while processing less than 2% of video frames. For videos longer than 20 minutes, it achieves an 11.9% gain in accuracy on LongVideoBench, demonstrating its effectiveness as a keyframe selection method and providing a simple and general solution for scalable long-video understanding with MLLMs.

Yong Liu, Zirui Zhu, Chaoyu Gong et al.

While fine-tuning large language models (LLMs) for specific tasks often yields impressive results, it comes at the cost of memory inefficiency due to back-propagation in gradient-based training. Memory-efficient Zeroth-order (MeZO) optimizers, recently proposed to address this issue, only require forward passes during training, making them more memory-friendly. However, the quality of gradient estimates in zeroth order optimization often depends on the data dimensionality, potentially explaining why MeZO still exhibits significant performance drops compared to standard fine-tuning across various tasks. Inspired by the success of Parameter-Efficient Fine-Tuning (PEFT), this paper introduces Sparse MeZO, a novel memory-efficient zeroth-order optimization approach that applies ZO only to a carefully chosen subset of parameters. We propose a simple yet effective parameter selection scheme that yields significant performance gains with Sparse-MeZO. Additionally, we develop a memory-optimized implementation for sparse masking, ensuring the algorithm requires only inference-level memory consumption, allowing Sparse-MeZO to fine-tune LLaMA-30b on a single A100 GPU. Experimental results illustrate that Sparse-MeZO consistently improves both performance and convergence speed over MeZO without any overhead. For example, it achieves a 9\% absolute accuracy improvement and 3.5x speedup over MeZO on the RTE task.

Zirui Zhu, Ziyi Ye

Federated Learning (FL) is a privacy-protected machine learning paradigm that allows model to be trained directly at the edge without uploading data. One of the biggest challenges faced by FL in practical applications is the heterogeneity of edge node data, which will slow down the convergence speed and degrade the performance of the model. For the above problems, a representative solution is to add additional constraints in the local training, such as FedProx, FedCurv and FedCL. However, the above algorithms still have room for improvement. We propose to use the aggregation of all models obtained in the past as new constraint target to further improve the performance of such algorithms. Experiments in various settings demonstrate that our method significantly improves the convergence speed and performance of the model.