Zekun Li

Zekun Li, Baolin Peng, Pengcheng He et al. · microsoft-research

We introduce Directional Stimulus Prompting, a novel framework for guiding black-box large language models (LLMs) toward specific desired outputs. Instead of directly adjusting LLMs, our method employs a small tunable policy model (e.g., T5) to generate an auxiliary directional stimulus prompt for each input instance. These directional stimulus prompts act as nuanced, instance-specific hints and clues to guide LLMs in generating desired outcomes, such as including specific keywords in the generated summary. Our approach sidesteps the challenges of direct LLM tuning by optimizing the policy model to explore directional stimulus prompts that align LLMs with desired behaviors. The policy model can be optimized through 1) supervised fine-tuning using labeled data and 2) reinforcement learning from offline or online rewards based on the LLM's output. We assess our method across summarization, dialogue response generation, and chain-of-thought reasoning tasks. Our experiments demonstrate that the framework consistently improves LLMs' (e.g., ChatGPT, Codex, InstructGPT) performance on these supervised tasks using minimal labeled data. Notably, using just 80 dialogues on the MultiWOZ dataset, our approach enhances ChatGPT's performance by an impressive 41.4%, matching or surpassing some fully supervised start-of-the-art models. Additionally, the instance-specific chain-of-thought prompt generated by our approach improves InstructGPT's reasoning accuracy compared to human-crafted or automatically generated prompts. The code and data are publicly available at \url{https://github.com/Leezekun/Directional-Stimulus-Prompting}.

Zekun Li, Baolin Peng, Pengcheng He et al. · microsoft-research

Large Language Models (LLMs) have demonstrated exceptional proficiency in instruction-following, becoming increasingly crucial across various applications. However, this capability brings with it the risk of prompt injection attacks, where attackers inject instructions into LLMs' input to elicit undesirable actions or content. Understanding the robustness of LLMs against such attacks is vital for their safe implementation. In this work, we establish a benchmark to evaluate the robustness of instruction-following LLMs against prompt injection attacks. Our objective is to determine the extent to which LLMs can be influenced by injected instructions and their ability to differentiate between these injected and original target instructions. Through extensive experiments with leading instruction-following LLMs, we uncover significant vulnerabilities in their robustness to such attacks. Our results indicate that some models are overly tuned to follow any embedded instructions in the prompt, overly focusing on the latter parts of the prompt without fully grasping the entire context. By contrast, models with a better grasp of the context and instruction-following capabilities will potentially be more susceptible to compromise by injected instructions. This underscores the need to shift the focus from merely enhancing LLMs' instruction-following capabilities to improving their overall comprehension of prompts and discernment of instructions that are appropriate to follow. We hope our in-depth analysis offers insights into the underlying causes of these vulnerabilities, aiding in the development of future solutions. Code and data are available at https://github.com/Leezekun/instruction-following-robustness-eval

Zekun Li, Wenhu Chen, Shiyang Li et al.

Building dialogue systems requires a large corpus of annotated dialogues. Such datasets are usually created via crowdsourcing, which is expensive and time-consuming. In this paper, we propose \textsc{Dialogic}, a novel dialogue simulation method based on large language model in-context learning to automate dataset creation. Seeded with a few annotated dialogues, \textsc{Dialogic} automatically selects in-context examples for demonstration and prompts GPT-3 to generate new dialogues and annotations in a controllable way. Our method can rapidly expand a small set of dialogue data with minimum or zero \textit{human involvement} and \textit{parameter update} and is thus much more cost-efficient and time-saving than crowdsourcing. Experimental results on the MultiWOZ dataset demonstrate that training a model on the simulated dialogues leads to even better performance than using the same amount of human-generated dialogues under the challenging low-resource settings, with as few as 85 dialogues as a seed. When enough data is available, our method can still serve as an effective data augmentation method. Human evaluation results also show that our simulated dialogues have near-human fluency and annotation accuracy. The code and data are available at \textbf{\url{https://github.com/Leezekun/dialogic}}.

Zekun Li, Lei Qi, Yinghuan Shi et al.

Semi-supervised learning (SSL) aims to leverage massive unlabeled data when labels are expensive to obtain. Unfortunately, in many real-world applications, the collected unlabeled data will inevitably contain unseen-class outliers not belonging to any of the labeled classes. To deal with the challenging open-set SSL task, the mainstream methods tend to first detect outliers and then filter them out. However, we observe a surprising fact that such approach could result in more severe performance degradation when labels are extremely scarce, as the unreliable outlier detector may wrongly exclude a considerable portion of valuable inliers. To tackle with this issue, we introduce a novel open-set SSL framework, IOMatch, which can jointly utilize inliers and outliers, even when it is difficult to distinguish exactly between them. Specifically, we propose to employ a multi-binary classifier in combination with the standard closed-set classifier for producing unified open-set classification targets, which regard all outliers as a single new class. By adopting these targets as open-set pseudo-labels, we optimize an open-set classifier with all unlabeled samples including both inliers and outliers. Extensive experiments have shown that IOMatch significantly outperforms the baseline methods across different benchmark datasets and different settings despite its remarkable simplicity. Our code and models are available at https://github.com/nukezil/IOMatch.

Zekun Li, Xianjun Yang, Kyuri Choi et al.

Scientific figure interpretation is a crucial capability for AI-driven scientific assistants built on advanced Large Vision Language Models. However, current datasets and benchmarks primarily focus on simple charts or other relatively straightforward figures from limited science domains. To address this gap, we present a comprehensive dataset compiled from peer-reviewed Nature Communications articles covering 72 scientific fields, encompassing complex visualizations such as schematic diagrams, microscopic images, and experimental data which require graduate-level expertise to interpret. We evaluated 19 proprietary and open-source models on two benchmark tasks, figure captioning and multiple-choice, and conducted human expert annotation. Our analysis revealed significant task challenges and performance gaps among models. Beyond serving as a benchmark, this dataset serves as a valuable resource for large-scale training. Fine-tuning Qwen2-VL-7B with our task-specific data achieved better performance than GPT-4o and even human experts in multiple-choice evaluations. Furthermore, continuous pre-training on our interleaved article and figure data substantially enhanced the model's downstream task performance in materials science. We have released our dataset to support further research.

Jing Qian, Hong Wang, Zekun Li et al.

Recent work has shown that large pretrained Language Models (LMs) can not only perform remarkably well on a range of Natural Language Processing (NLP) tasks but also start improving on reasoning tasks such as arithmetic induction, symbolic manipulation, and commonsense reasoning with increasing size of models. However, it is still unclear what the underlying capabilities of these LMs are. Surprisingly, we find that these models have limitations on certain basic symbolic manipulation tasks such as copy, reverse, and addition. When the total number of symbols or repeating symbols increases, the model performance drops quickly. We investigate the potential causes behind this phenomenon and examine a set of possible methods, including explicit positional markers, fine-grained computation steps, and LMs with callable programs. Experimental results show that none of these techniques can solve the simplest addition induction problem completely. In the end, we introduce LMs with tutor, which demonstrates every single step of teaching. LMs with tutor is able to deliver 100% accuracy in situations of OOD and repeating symbols, shedding new insights on the boundary of large LMs in induction.

Zhiyuan Hu, Yue Feng, Yang Deng et al.

Recently, the development of large language models (LLMs) has been significantly enhanced the question answering and dialogue generation, and makes them become increasingly popular in current practical scenarios. While unlike the general dialogue system which emphasizes the semantic performance, the task-oriented dialogue (ToD) systems aim to achieve the dialogue goal efficiently and successfully in multiple turns. Unfortunately, existing LLM-induced ToD systems lack the direct reward toward the final goal and do not take account of the dialogue proactivity that can strengthen the dialogue efficiency. To fill these gaps, we introduce the ProToD (Proactively Goal-Driven LLM-Induced ToD) approach, which anticipates the future dialogue actions and incorporates the goal-oriented reward signal to enhance ToD systems. Additionally, we present a novel evaluation method that assesses ToD systems based on goal-driven dialogue simulations. This method allows us to gauge user satisfaction, system efficiency and successful rate while overcoming the limitations of current Information and Success metrics. Empirical experiments conducted on the MultiWoZ 2.1 dataset demonstrate that our model can achieve superior performance using only 10% of the data compared to previous end-to-end fully supervised models. This improvement is accompanied by enhanced user satisfaction and efficiency.

Xing Cui, Zekun Li, Peipei Li et al.

This paper explores interactive facial image editing via dialogue and introduces the ChatEdit benchmark dataset for evaluating image editing and conversation abilities in this context. ChatEdit is constructed from the CelebA-HQ dataset, incorporating annotated multi-turn dialogues corresponding to user edit requests on the images. The dataset is challenging, as it requires the system to dynamically track user requests, edit images, and generate appropriate responses. Accordingly, we propose three benchmark tasks: (i) user edit request tracking, (ii) image editing, and (iii) response generation. We present a novel baseline framework that integrates a dialogue module for both tracking user requests and generating responses and an image editing module for image editing. Unlike previous approaches, our framework directly tracks user edit requests from the entire dialogue history up to the current turn and modifies the original image rather than adjusting the previous turn's output, thereby reducing error accumulation and preventing attribute forgetfulness. Extensive experiments on the ChatEdit dataset underline our framework's superior performance against prior models, while also highlighting potential room for further research. We will release the code and data publicly to facilitate advancements in complex interactive facial image editing.

Zekun Li, Zhiyu Zoey Chen, Mike Ross et al. · microsoft-research

Large language models (LLMs) are increasingly prevalent in conversational systems due to their advanced understanding and generative capabilities in general contexts. However, their effectiveness in task-oriented dialogues (TOD), which requires not only response generation but also effective dialogue state tracking (DST) within specific tasks and domains, remains less satisfying. In this work, we propose a novel approach FnCTOD for solving DST with LLMs through function calling. This method improves zero-shot DST, allowing adaptation to diverse domains without extensive data collection or model tuning. Our experimental results demonstrate that our approach achieves exceptional performance with both modestly sized open-source and also proprietary LLMs: with in-context prompting it enables various 7B or 13B parameter models to surpass the previous state-of-the-art (SOTA) achieved by ChatGPT, and improves ChatGPT's performance beating the SOTA by 5.6% average joint goal accuracy (JGA). Individual model results for GPT-3.5 and GPT-4 are boosted by 4.8% and 14%, respectively. We also show that by fine-tuning on a small collection of diverse task-oriented dialogues, we can equip modestly sized models, specifically a 13B parameter LLaMA2-Chat model, with function-calling capabilities and DST performance comparable to ChatGPT while maintaining their chat capabilities. We have made the code publicly available at https://github.com/facebookresearch/FnCTOD

Qinghe Ma, Jian Zhang, Zekun Li et al.

Large pretrained visual foundation models exhibit impressive general capabilities. However, the extensive prior knowledge inherent in these models can sometimes be a double-edged sword when adapting them to downstream tasks in specific domains. In the context of semi-supervised medical image segmentation with domain shift, foundation models like MedSAM tend to make overconfident predictions, some of which are incorrect. The error accumulation hinders the effective utilization of unlabeled data and limits further improvements. In this paper, we introduce a Synergistic training framework for Foundation and Conventional models (SynFoC) to address the issue. We observe that a conventional model trained from scratch has the ability to correct the high-confidence mispredictions of the foundation model, while the foundation model can supervise it with high-quality pseudo-labels in the early training stages. Furthermore, to enhance the collaborative training effectiveness of both models and promote reliable convergence towards optimization, the consensus-divergence consistency regularization is proposed. We demonstrate the superiority of our method across four public multi-domain datasets. In particular, our method improves the Dice score by 10.31\% on the Prostate dataset. Our code is available at https://github.com/MQinghe/SynFoC .

Hengjia Li, Haonan Qiu, Shiwei Zhang et al.

The current text-to-video (T2V) generation has made significant progress in synthesizing realistic general videos, but it is still under-explored in identity-specific human video generation with customized ID images. The key challenge lies in maintaining high ID fidelity consistently while preserving the original motion dynamic and semantic following after the identity injection. Current video identity customization methods mainly rely on reconstructing given identity images on text-to-image models, which have a divergent distribution with the T2V model. This process introduces a tuning-inference gap, leading to dynamic and semantic degradation. To tackle this problem, we propose a novel framework, dubbed $\textbf{PersonalVideo}$, that applies a mixture of reward supervision on synthesized videos instead of the simple reconstruction objective on images. Specifically, we first incorporate identity consistency reward to effectively inject the reference's identity without the tuning-inference gap. Then we propose a novel semantic consistency reward to align the semantic distribution of the generated videos with the original T2V model, which preserves its dynamic and semantic following capability during the identity injection. With the non-reconstructive reward training, we further employ simulated prompt augmentation to reduce overfitting by supervising generated results in more semantic scenarios, gaining good robustness even with only a single reference image. Extensive experiments demonstrate our method's superiority in delivering high identity faithfulness while preserving the inherent video generation qualities of the original T2V model, outshining prior methods.

Yueying Zou, Peipei Li, Zekun Li et al.

The proliferation of AI-generated media poses significant challenges to information authenticity and social trust, making reliable detection methods highly demanded. Methods for detecting AI-generated media have evolved rapidly, paralleling the advancement of Multimodal Large Language Models (MLLMs). Current detection approaches can be categorized into two main groups: Non-MLLM-based and MLLM-based methods. The former employs high-precision, domain-specific detectors powered by deep learning techniques, while the latter utilizes general-purpose detectors based on MLLMs that integrate authenticity verification, explainability, and localization capabilities. Despite significant progress in this field, there remains a gap in literature regarding a comprehensive survey that examines the transition from domain-specific to general-purpose detection methods. This paper addresses this gap by providing a systematic review of both approaches, analyzing them from single-modal and multi-modal perspectives. We present a detailed comparative analysis of these categories, examining their methodological similarities and differences. Through this analysis, we explore potential hybrid approaches and identify key challenges in forgery detection, providing direction for future research. Additionally, as MLLMs become increasingly prevalent in detection tasks, ethical and security considerations have emerged as critical global concerns. We examine the regulatory landscape surrounding Generative AI (GenAI) across various jurisdictions, offering valuable insights for researchers and practitioners in this field.

Xuannan Liu, Zekun Li, Zheqi He et al.

The increasing deployment of Large Vision-Language Models (LVLMs) raises safety concerns under potential malicious inputs. However, existing multimodal safety evaluations primarily focus on model vulnerabilities exposed by static image inputs, ignoring the temporal dynamics of video that may induce distinct safety risks. To bridge this gap, we introduce Video-SafetyBench, the first comprehensive benchmark designed to evaluate the safety of LVLMs under video-text attacks. It comprises 2,264 video-text pairs spanning 48 fine-grained unsafe categories, each pairing a synthesized video with either a harmful query, which contains explicit malice, or a benign query, which appears harmless but triggers harmful behavior when interpreted alongside the video. To generate semantically accurate videos for safety evaluation, we design a controllable pipeline that decomposes video semantics into subject images (what is shown) and motion text (how it moves), which jointly guide the synthesis of query-relevant videos. To effectively evaluate uncertain or borderline harmful outputs, we propose RJScore, a novel LLM-based metric that incorporates the confidence of judge models and human-aligned decision threshold calibration. Extensive experiments show that benign-query video composition achieves average attack success rates of 67.2%, revealing consistent vulnerabilities to video-induced attacks. We believe Video-SafetyBench will catalyze future research into video-based safety evaluation and defense strategies.

You Wang, Zekun Li, Lei Qi et al.

Despite the promising performance achieved by current semi-supervised models in segmenting individual medical targets, many of these models suffer a notable decrease in performance when tasked with the simultaneous segmentation of multiple targets. A vital factor could be attributed to the imbalanced scales among different targets: during simultaneously segmenting multiple targets, large targets dominate the loss, leading to small targets being misclassified as larger ones. To this end, we propose a novel method, which consists of a Collaborative Generalist and several Specialists, termed CGS. It is centered around the idea of employing a specialist for each target class, thus avoiding the dominance of larger targets. The generalist performs conventional multi-target segmentation, while each specialist is dedicated to distinguishing a specific target class from the remaining target classes and the background. Based on a theoretical insight, we demonstrate that CGS can achieve a more balanced training. Moreover, we develop cross-consistency losses to foster collaborative learning between the generalist and the specialists. Lastly, regarding their intrinsic relation that the target class of any specialized head should belong to the remaining classes of the other heads, we introduce an inter-head error detection module to further enhance the quality of pseudo-labels. Experimental results on three popular benchmarks showcase its superior performance compared to state-of-the-art methods.

Xuannan Liu, Zekun Li, Peipei Li et al.

Current multimodal misinformation detection (MMD) methods often assume a single source and type of forgery for each sample, which is insufficient for real-world scenarios where multiple forgery sources coexist. The lack of a benchmark for mixed-source misinformation has hindered progress in this field. To address this, we introduce MMFakeBench, the first comprehensive benchmark for mixed-source MMD. MMFakeBench includes 3 critical sources: textual veracity distortion, visual veracity distortion, and cross-modal consistency distortion, along with 12 sub-categories of misinformation forgery types. We further conduct an extensive evaluation of 6 prevalent detection methods and 15 Large Vision-Language Models (LVLMs) on MMFakeBench under a zero-shot setting. The results indicate that current methods struggle under this challenging and realistic mixed-source MMD setting. Additionally, we propose MMD-Agent, a novel approach to integrate the reasoning, action, and tool-use capabilities of LVLM agents, significantly enhancing accuracy and generalization. We believe this study will catalyze future research into more realistic mixed-source multimodal misinformation and provide a fair evaluation of misinformation detection methods.

Zekun Li, Wei Zhao, Feng Shi et al.

How to fast and accurately assess the severity level of COVID-19 is an essential problem, when millions of people are suffering from the pandemic around the world. Currently, the chest CT is regarded as a popular and informative imaging tool for COVID-19 diagnosis. However, we observe that there are two issues -- weak annotation and insufficient data that may obstruct automatic COVID-19 severity assessment with CT images. To address these challenges, we propose a novel three-component method, i.e., 1) a deep multiple instance learning component with instance-level attention to jointly classify the bag and also weigh the instances, 2) a bag-level data augmentation component to generate virtual bags by reorganizing high confidential instances, and 3) a self-supervised pretext component to aid the learning process. We have systematically evaluated our method on the CT images of 229 COVID-19 cases, including 50 severe and 179 non-severe cases. Our method could obtain an average accuracy of 95.8%, with 93.6% sensitivity and 96.4% specificity, which outperformed previous works.