Ziwen Xu

Ziwen Xu, Chenyan Wu, Hengyu Sun et al.

Methods for controlling large language models (LLMs), including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model's valid-generation manifold. Finally, we introduce a new steering approach SPLIT guided by this analysis that improves preference while better preserving utility. Code is available at https://github.com/zjunlp/EasyEdit/blob/main/examples/SPLIT.md.

Ziwen Xu, Kewei Xu, Haoming Xu et al.

Large Language Models (LLMs) are increasingly deployed in socially sensitive domains, yet their unpredictable behaviors, ranging from misaligned intent to inconsistent personality, pose significant risks. We introduce SteerEval, a hierarchical benchmark for evaluating LLM controllability across three domains: language features, sentiment, and personality. Each domain is structured into three specification levels: L1 (what to express), L2 (how to express), and L3 (how to instantiate), connecting high-level behavioral intent to concrete textual output. Using SteerEval, we systematically evaluate contemporary steering methods, revealing that control often degrades at finer-grained levels. Our benchmark offers a principled and interpretable framework for safe and controllable LLM behavior, serving as a foundation for future research.

Yunshui Li, Yiyuan Ma, Shen Yan et al.

Model merging has emerged as a promising technique for enhancing large language models, though its application in large-scale pre-training remains relatively unexplored. In this paper, we present a comprehensive investigation of model merging techniques during the pre-training process. Through extensive experiments with both dense and Mixture-of-Experts (MoE) architectures ranging from millions to over 100 billion parameters, we demonstrate that merging checkpoints trained with constant learning rates not only achieves significant performance improvements but also enables accurate prediction of annealing behavior. These improvements lead to both more efficient model development and significantly lower training costs. Our detailed ablation studies on merging strategies and hyperparameters provide new insights into the underlying mechanisms while uncovering novel applications. Through comprehensive experimental analysis, we offer the open-source community practical pre-training guidelines for effective model merging.

Chenxi Wang, Jizhan Fang, Xiang Chen et al.

Recent advancements in Large Multimodal Models (LMMs) have shown promise in Autonomous Driving Systems (ADS). However, their direct application to ADS is hindered by challenges such as misunderstanding of traffic knowledge, complex road conditions, and diverse states of vehicle. To address these challenges, we propose the use of Knowledge Editing, which enables targeted modifications to a model's behavior without the need for full retraining. Meanwhile, we introduce ADS-Edit, a multimodal knowledge editing dataset specifically designed for ADS, which includes various real-world scenarios, multiple data types, and comprehensive evaluation metrics. We conduct comprehensive experiments and derive several interesting conclusions. We hope that our work will contribute to the further advancement of knowledge editing applications in the field of autonomous driving. Code and data are available in https://github.com/zjunlp/EasyEdit/blob/main/examples/ADSEdit.md.

Ziwen Xu, beiji Zou, Qing Liu

Retinal image quality assessment is an essential prerequisite for diagnosis of retinal diseases. Its goal is to identify retinal images in which anatomic structures and lesions attracting ophthalmologists' attention most are exhibited clearly and definitely while reject poor quality fundus images. Motivated by this, we mimic the way that ophthalmologists assess the quality of retinal images and propose a method termed SalStructuIQA. First, two salient structures for automated retinal quality assessment. One is the large-size salient structures including optic disc region and exudates in large-size. The other is the tiny-size salient structures which mainly include vessels. Then we incorporate the proposed two salient structure priors with deep convolutional neural network (CNN) to shift the focus of CNN to salient structures. Accordingly, we develop two CNN architectures: Dual-branch SalStructIQA and Single-branch SalStructIQA. Dual-branch SalStructIQA contains two CNN branches and one is guided by large-size salient structures while the other is guided by tiny-size salient structures. Single-branch SalStructIQA contains one CNN branch, which is guided by the concatenation of salient structures in both large-size and tiny-size. Experimental results on Eye-Quality dataset show that our proposed Dual-branch SalStructIQA outperforms the state-of-the-art methods for retinal image quality assessment and Single-branch SalStructIQA is much light-weight comparing with state-of-the-art deep retinal image quality assessment methods and still achieves competitive performances.