Zhiyu Xu

Zhiyu Xu, Jia Liu, Yixin Wang et al.

The proliferation of Large Language Models (LLMs) necessitates valid evaluation methods to provide guidance for both downstream applications and actionable future improvements. The Item Response Theory (IRT) model with Computerized Adaptive Testing has recently emerged as a promising framework for evaluating LLMs via their response accuracy. Beyond simple response accuracy, LLMs' chain of thought (CoT) lengths serve as a vital indicator of their reasoning ability. To leverage the CoT length information to assist the evaluation of LLMs, we propose the Latency-Response Theory (LaRT) model, which jointly models both the response accuracy and CoT length by introducing a key correlation parameter between the latent ability and the latent speed. We derive an efficient stochastic approximation Expectation-Maximization algorithm for parameter estimation. We establish rigorous identifiability results for the latent ability and latent speed parameters to ensure the statistical validity of their estimation. Through both theoretical asymptotic analyses and simulation studies, we demonstrate LaRT's advantages over IRT in terms of superior estimation accuracy and shorter confidence intervals for latent trait estimation. To evaluate LaRT in real data, we collect responses from diverse LLMs on popular benchmark datasets. We find that LaRT yields different LLM rankings than IRT and outperforms IRT across multiple key evaluation metrics including predictive power, item efficiency, ranking validity, and LLM evaluation efficiency. Code and data are available at https://github.com/Toby-X/Latency-Response-Theory-Model.

Zhiyu Xu, Lean Wang, Yuanxin Liu et al.

Vision-Language Models (VLMs) have demonstrated remarkable proficiency in general multi-modal understanding; yet they struggle to efficiently acquire continually evolving domain-specific skills. Conventional approaches to enhancing VLM capabilities, such as Supervised Fine-Tuning (SFT), require extensive dataset curation and substantial computational resources. Model merging has emerged as an efficient alternative that enables the transfer of domain-specific expertise from Large Language Models (LLMs) to VLMs without incurring additional training data requirements or significant computational overhead. Unlike conventional merging of homogeneous LLMs, which mainly aggregates existing capabilities, cross-modal skill injection aims to induce emergent cross-modal capabilities by integrating a domain-expert LLM into a VLM. However, existing research lacks a systematic analysis of the applicability and methodology of cross-modal skill injection. In this study, we investigate cross-modal skill injection across three main aspects: scenarios, methods, and hyperparameters. For scenarios, we find that cross-modal skill injection generally performs well in instruction-following and cross-lingual settings, yet struggles with mathematical reasoning. For methods, we find that classic approaches such as TA and DARE consistently achieve superior performance over alternative merging methods. We also provide a systematic and quantitative analysis of the hyperparameter tuning that these classic methods critically depend on.

Zhiyu Xu, Qingliang Chen

Glass-like objects can be seen everywhere in our daily life which are very hard for existing methods to segment them. The properties of transparencies pose great challenges of detecting them from the chaotic background and the vague separation boundaries further impede the acquisition of their exact contours. Moving machines which ignore glasses have great risks of crashing into transparent barriers or difficulties in analysing objects reflected in the mirror, thus it is of substantial significance to accurately locate glass-like objects and completely figure out their contours. In this paper, inspired by the scale integration strategy and the refinement method, we proposed a brand-new network, named as MGNet, which consists of a Fine-Rescaling and Merging module (FRM) to improve the ability to extract spatially relationship and a Primary Prediction Guiding module (PPG) to better mine the leftover semantics from the fused features. Moreover, we supervise the model with a novel loss function with the uncertainty-aware loss to produce high-confidence segmentation maps. Unlike the existing glass segmentation models that must be trained on different settings with respect to varied datasets, our model are trained under consistent settings and has achieved superior performance on three popular public datasets. Code is available at

Zhiyu Xu, Weilong Yan, Yufei Shi et al.

Recent advancements in multimodal large language models (MLLMs) and video agent systems have significantly improved general video understanding. However, when applied to scientific video understanding and educating, a domain that demands external professional knowledge integration and rigorous step-wise reasoning, existing approaches often struggle. To bridge this gap, we propose SciEducator, the first iterative self-evolving multi-agent system for scientific video comprehension and education. Rooted in the classical Deming Cycle from management science, our design reformulates its Plan-Do-Study-Act philosophy into a self-evolving reasoning and feedback mechanism, which facilitates the interpretation of intricate scientific activities in videos. Moreover, SciEducator can produce multimodal educational content tailored to specific scientific processes, including textual instructions, visual guides, audio narrations, and interactive references. To support evaluation, we construct SciVBench, a benchmark consisting of 500 expert-verified and literature-grounded science QA pairs across five categories, covering physical, chemical, and everyday phenomena. Extensive experiments demonstrate that SciEducator substantially outperforms leading closed-source MLLMs (e.g., Gemini, GPT-4o) and state-of-the-art video agents on the benchmark, establishing a new paradigm for the community.

Zhiyu Xu, Qingliang Chen

Driven by large data trained segmentation models, such as SAM , research in one-shot segmentation has experienced significant advancements. Recent contributions like PerSAM and MATCHER , presented at ICLR 2024, utilize a similar approach by leveraging SAM with one or a few reference images to generate high quality segmentation masks for target images. Specifically, they utilize raw encoded features to compute cosine similarity between patches within reference and target images along the channel dimension, effectively generating prompt points or boxes for the target images a technique referred to as the matching strategy. However, relying solely on raw features might introduce biases and lack robustness for such a complex task. To address this concern, we delve into the issues of feature interaction and uneven distribution inherent in raw feature based matching. In this paper, we propose a simple and training-free method to enhance the validity and robustness of the matching strategy at no additional computational cost (NubbleDrop). The core concept involves randomly dropping feature channels (setting them to zero) during the matching process, thereby preventing models from being influenced by channels containing deceptive information. This technique mimics discarding pathological nubbles, and it can be seamlessly applied to other similarity computing scenarios. We conduct a comprehensive set of experiments, considering a wide range of factors, to demonstrate the effectiveness and validity of our proposed method. Our results showcase the significant improvements achieved through this simmple and straightforward approach.