Ziyi Liu

Dong-Ho Lee, Akshen Kadakia, Brihi Joshi et al. · meta-ai

NLP models are susceptible to learning spurious biases (i.e., bugs) that work on some datasets but do not properly reflect the underlying task. Explanation-based model debugging aims to resolve spurious biases by showing human users explanations of model behavior, asking users to give feedback on the behavior, then using the feedback to update the model. While existing model debugging methods have shown promise, their prototype-level implementations provide limited practical utility. Thus, we propose XMD: the first open-source, end-to-end framework for explanation-based model debugging. Given task- or instance-level explanations, users can flexibly provide various forms of feedback via an intuitive, web-based UI. After receiving user feedback, XMD automatically updates the model in real time, by regularizing the model so that its explanations align with the user feedback. The new model can then be easily deployed into real-world applications via Hugging Face. Using XMD, we can improve the model's OOD performance on text classification tasks by up to 18%.

Jun Rui Huang, Wang Bill Zhu, Ziyi Liu et al.

Large language models (LLMs) are increasingly used as conversational partners for companionship, emotional disclosure, and interpersonal advice, but the social dynamics of these interactions can create harms that are not captured by capability-oriented or traditional safety evaluations. We introduce the Social AI Design Code, a framework for evaluating whether LLMs align with user welfare in social interactions, including whether they encourage harmful intimacy, dependence, or prolonged engagement. To evaluate these risks in natural and diverse user-LLM interactions, we operationalize the code with EUDAIMONIA, a benchmark of 969 user inputs and 3,147 design-requirement violation checks built from WildChat through weak-to-strong filtration, multi-model relabeling, and controlled rewriting. Evaluating 22 recent LLMs, we find that even the strongest models, Claude-Opus-4.7 and GPT-5.5, violate 30.7% and 27.2% of checks, respectively. Extended thinking does not reduce violation rates, suggesting that these failures are persistent social-alignment problems rather than deficits solvable through test-time reasoning alone.

Brihi Joshi, Aaron Chan, Ziyi Liu et al. · meta-ai

By explaining how humans would solve a given task, human rationales can provide strong learning signal for neural language models (LMs). Explanation regularization (ER) aims to improve LM generalization by pushing the LM's machine rationales (Which input tokens did the LM focus on?) to align with human rationales (Which input tokens would humans focus on?). Though prior works primarily study ER via in-distribution (ID) evaluation, out-of-distribution (OOD) generalization is often more critical in real-world scenarios, yet ER's effect on OOD generalization has been underexplored. In this paper, we introduce ER-Test, a framework for evaluating ER models' OOD generalization along three dimensions: unseen dataset tests, contrast set tests, and functional tests. Using ER-Test, we extensively analyze how ER models' OOD generalization varies with different ER design choices. Across two tasks and six datasets, ER-Test shows that ER has little impact on ID performance but can yield large OOD performance gains. Also, we find that ER can improve OOD performance even with limited rationale supervision. ER-Test's results help demonstrate ER's utility and establish best practices for using ER effectively.

Fengtao Zhou, Yingxue Xu, Zhengyu Zhang et al.

Comprehensive molecular profiling is essential for modern precision oncology but remains hindered by prohibitive costs, specimen exhaustion, and protracted turnaround times. While pathology foundation models (PFMs) have demonstrated potential for inferring molecular phenotypes from routine hematoxylin and eosin (H&E) whole-slide images (WSIs), current architectures primarily rely on vision-centric self-supervised learning or vision-language alignment, lacking the spatially resolved molecular supervision required to connect subtle morphological features with underlying genomic alterations. Spatial transcriptomics (ST) emerges as a transformative technology that enables transcriptomic quantification within intact tissue sections, thereby preserving the precise spatial link between histology and molecular profiles. In this study, we present a Spatial Transcriptomics-guided Alignment framework for Molecular Profiling (STAMP), which endows PFMs with intrinsic molecular awareness. To support this paradigm, we curated HumanST-1k, a human ST dataset spanning diverse anatomical organs and sequencing platforms. This atlas yields 1.8 million pairs of H&E patches and corresponding transcriptomic profiles, providing a corpus that links histological structures with their molecular states. To mitigate the technical noise inherent to raw transcriptomics, STAMP applies a pathway-informed alignment strategy that aggregates transcriptomic data into biologically functional pathways, which are subsequently integrated into PFMs via parameter-efficient fine-tuning. This alignment enriches the representation space of PFMs and unlocks their capacity to resolve sub-visual molecular signatures. The clinical utility of these augmented representations was validated through a multi-tier evaluation framework.

Jesse Zhang, Jiahui Zhang, Karl Pertsch et al.

We propose BOSS, an approach that automatically learns to solve new long-horizon, complex, and meaningful tasks by growing a learned skill library with minimal supervision. Prior work in reinforcement learning require expert supervision, in the form of demonstrations or rich reward functions, to learn long-horizon tasks. Instead, our approach BOSS (BOotStrapping your own Skills) learns to accomplish new tasks by performing "skill bootstrapping," where an agent with a set of primitive skills interacts with the environment to practice new skills without receiving reward feedback for tasks outside of the initial skill set. This bootstrapping phase is guided by large language models (LLMs) that inform the agent of meaningful skills to chain together. Through this process, BOSS builds a wide range of complex and useful behaviors from a basic set of primitive skills. We demonstrate through experiments in realistic household environments that agents trained with our LLM-guided bootstrapping procedure outperform those trained with naive bootstrapping as well as prior unsupervised skill acquisition methods on zero-shot execution of unseen, long-horizon tasks in new environments. Website at clvrai.com/boss.

Ziyi Liu, Soumya Sanyal, Isabelle Lee et al. · amazon-science, uw

In a plethora of recent work, large language models (LLMs) demonstrated impressive reasoning ability, but many proposed downstream reasoning tasks only focus on final answers. Two fundamental questions persist: 1) how consistent is the reasoning, and 2) can models detect unreliable reasoning? In this paper, we investigate self-contradictory (Self-Contra) reasoning, where the model reasoning does not support its answers. To answer 1), we define and assess the Self-Contra rate across three datasets and delve into finer-grained categories of Self-Contra reasoning. We find that LLMs often contradict themselves in reasoning tasks involving contextual information understanding or commonsense. The model may generate correct answers by taking shortcuts in reasoning or overlooking contextual evidence, leading to compromised reasoning. For 2), we task the state-of-the-art model GPT-4 with identifying Self-Contra reasoning and finer-grained fallacies. We find that finer-grained categories enhanced detection can improve GPT-4's ability to detect Self-Contra. However, it is only able to detect Self-Contra with a 52.2% F1 score, much lower compared to 66.7% for humans. Our results indicate that current LLMs lack the robustness necessary for reliable reasoning and we emphasize the urgent need for establishing best practices in comprehensive reasoning evaluations beyond pure performance-based metrics.

Firas Gabetni, Alexandre Rocchi Henry, Nacim Belkhir et al.

Detecting covariate shift is critical for building reliable vision systems. While most prior work focuses on improving robustness to shift, explicitly detecting covariate shift remains underexplored. Existing approaches typically rely on fully supervised training, requiring labeled examples from both original and shifted distributions, which is often impractical. In this paper, we show that covariate shift detection can be effectively addressed with weaker supervision using Positive Unlabeled (PU) learning. However, under covariate shift, in distribution and shifted data overlap significantly, making classical PU methods unstable and sensitive to noise. To overcome this challenge, we introduce Spectral PU Neighborhood Annotation (SPUNA), a geometry aware framework that progressively discovers shifted data by leveraging the local manifold structure of visual features. Extensive experiments show that SPUNA achieves state of the art performance in PU settings and remarkably matches the performances of fully supervised methods. Moreover, our approach transfers robustly across different types of shifts, demonstrating strong generalization capabilities.

Bowen Jiang, Yuan Yuan, Maohao Shen et al. · uw

Personalization is one of the next milestones in advancing AI capability and alignment. We introduce PersonaMem-v2, the state-of-the-art dataset for LLM personalization that simulates 1,000 realistic user-chatbot interactions on 300+ scenarios, 20,000+ user preferences, and 128k-token context windows, where most user preferences are implicitly revealed to reflect real-world interactions. Using this data, we investigate how reinforcement fine-tuning enables a model to improve its long-context reasoning capabilities for user understanding and personalization. We also develop a framework for training an agentic memory system, which maintains a single, human-readable memory that grows with each user over time. In our experiments, frontier LLMs still struggle with implicit personalization, achieving only 37-48% accuracy. While they support long context windows, reasoning remains the bottleneck for implicit personalization tasks. Using reinforcement fine-tuning, we successfully train Qwen3-4B to outperforms GPT-5, reaching 53% accuracy in implicit personalization. Moreover, our agentic memory framework achieves state-of-the-art 55% accuracy while using 16x fewer input tokens, relying on a 2k-token memory instead of full 32k conversation histories. These results underscore the impact of our dataset and demonstrate agentic memory as a scalable path toward real-world personalized intelligence.

Jiawei Li, Ziyi Liu, Weijie Shi et al.

3D object grounding localizes referred objects in a 3D scene from natural language. Unified instance-centric 3D-LLMs aim to solve grounding together with dialog, QA, and captioning, yet many rely on a single pointer-style grounding decision that compresses a relational instruction into one selection. This is brittle for fine-grained queries where multiple same-class candidates must be ruled out by context objects and spatial relations. We propose Structured Spatial Reasoning 3D-LLM (SSR3D-LLM), a structured grounding interface for unified 3D-LLMs. Given fixed Mask3D object proposals, the LLM writes a sequence of latent spatial reasoning steps and memory tokens from the query, and a geometry-aware scorer reads these latent steps in order to refine candidate rankings step by step with step-length masking. The latent steps are learned from standard benchmark target supervision with auxiliary referential-cue supervision during training, while inference uses only the input query and Mask3D proposals. Across ReferIt3D, ScanRefer, and Multi3DRef, SSR3D-LLM achieves the strongest results among unified 3D-LLM baselines, with substantial gains over the single-pointer QPG baseline on fine-grained grounding and consistent improvements over prior unified 3D-LLMs, while preserving the default language-task route.

Xue Yu, Ziyi Liu, Wu Wang et al.

This study investigates clustered federated learning (FL), one of the formulations of FL with non-i.i.d. data, where the devices are partitioned into clusters and each cluster optimally fits its data with a localized model. We propose a clustered FL framework that incorporates a nonconvex penalty to pairwise differences of parameters. Without a priori knowledge of the set of devices in each cluster and the number of clusters, this framework can autonomously estimate cluster structures. To implement the proposed framework, we introduce a novel clustered FL method called Fusion Penalized Federated Clustering (FPFC). Building upon the standard alternating direction method of multipliers (ADMM), FPFC can perform partial updates at each communication round and allows parallel computation with variable workload. These strategies significantly reduce the communication cost while ensuring privacy, making it practical for FL. We also propose a new warmup strategy for hyperparameter tuning in FL settings and explore the asynchronous variant of FPFC (asyncFPFC). Theoretical analysis provides convergence guarantees for FPFC with general losses and establishes the statistical convergence rate under a linear model with squared loss. Extensive experiments have demonstrated the superiority of FPFC compared to current methods, including robustness and generalization capability.

Jingjing Jiang, Ziyi Liu, Nanning Zheng

Benefiting from large-scale pretrained vision language models (VLMs), the performance of visual question answering (VQA) has approached human oracles. However, finetuning such models on limited data often suffers from overfitting and poor generalization issues, leading to a lack of model robustness. In this paper, we aim to improve input robustness from an information bottleneck perspective when adapting pretrained VLMs to the downstream VQA task. Input robustness refers to the ability of models to defend against visual and linguistic input variations, as well as shortcut learning involved in inputs. Generally, the representations obtained by pretrained VLMs inevitably contain irrelevant and redundant information for a specific downstream task, resulting in statistically spurious correlations and insensitivity to input variations. To encourage representations to converge to a minimal sufficient statistic in multimodal learning, we propose Correlation Information Bottleneck (CIB), which seeks a tradeoff between compression and redundancy in representations by minimizing the mutual information (MI) between inputs and representations while maximizing the MI between outputs and representations. Moreover, we derive a tight theoretical upper bound for the mutual information between multimodal inputs and representations, incorporating different internal correlations that guide models to learn more robust representations and facilitate modality alignment. Extensive experiments consistently demonstrate the effectiveness and superiority of the proposed CIB in terms of input robustness and accuracy.

Ziyi Liu, Rakshitha Godahewa, Kasun Bandara et al.

Machine learning (ML) based time series forecasting models often require and assume certain degrees of stationarity in the data when producing forecasts. However, in many real-world situations, the data distributions are not stationary and they can change over time while reducing the accuracy of the forecasting models, which in the ML literature is known as concept drift. Handling concept drift in forecasting is essential for many ML methods in use nowadays, however, the prior work only proposes methods to handle concept drift in the classification domain. To fill this gap, we explore concept drift handling methods in particular for Global Forecasting Models (GFM) which recently have gained popularity in the forecasting domain. We propose two new concept drift handling methods, namely: Error Contribution Weighting (ECW) and Gradient Descent Weighting (GDW), based on a continuous adaptive weighting concept. These methods use two forecasting models which are separately trained with the most recent series and all series, and finally, the weighted average of the forecasts provided by the two models are considered as the final forecasts. Using LightGBM as the underlying base learner, in our evaluation on three simulated datasets, the proposed models achieve significantly higher accuracy than a set of statistical benchmarks and LightGBM baselines across four evaluation metrics.

Yusen Zhang, Zhongli Li, Qingyu Zhou et al.

To automatically correct handwritten assignments, the traditional approach is to use an OCR model to recognize characters and compare them to answers. The OCR model easily gets confused on recognizing handwritten Chinese characters, and the textual information of the answers is missing during the model inference. However, teachers always have these answers in mind to review and correct assignments. In this paper, we focus on the Chinese cloze tests correction and propose a multimodal approach (named AiM). The encoded representations of answers interact with the visual information of students' handwriting. Instead of predicting 'right' or 'wrong', we perform the sequence labeling on the answer text to infer which answer character differs from the handwritten content in a fine-grained way. We take samples of OCR datasets as the positive samples for this task, and develop a negative sample augmentation method to scale up the training data. Experimental results show that AiM outperforms OCR-based methods by a large margin. Extensive studies demonstrate the effectiveness of our multimodal approach.

Zhengrui Guo, Zhengyu Zhang, Jiabo Ma et al.

Pathological assessment guides lung cancer diagnosis, treatment selection, and prognostic evaluation, yet current CPath approaches rely on task-specific models for isolated objectives. Although pan-cancer foundation models offer versatility, they lack subspecialty-level depth and have not been evaluated across clinical workflows or prospectively validated in real-world settings. We introduce PulmoFoundation, a multi-center, prospectively validated, randomized controlled trial (RCT)-evaluated foundation model for comprehensive lung pathology assessment across pre-operative, intra-operative, and post-operative care. Built upon Virchow2 via subspecialty-specific pretraining using ~40,000 diagnostic H&E-stained whole-slide images (WSIs), PulmoFoundation was systematically evaluated on ~26,000 WSIs across 32 clinically relevant tasks. In addition to accurately predicting molecular markers and patient survival, our model achieves clinical-grade performance in core diagnostic tasks across biopsy, frozen section, and surgical resection slides. In a registered prospective study of 1,357 patients across 11 diagnostic tasks, our model achieved an average AUC of 92.3%. Using pre-specified triage thresholds, PulmoFoundation could reduce additional second-review burden for 68.8% of biopsies and 83.0% of frozen sections, and defer 44.5% of IHC stain orders, with PPVs of 1.0, 0.991, and 0.966. Beyond prospective validation, we conducted a crossover RCT with eight pathologists, in which AI assistance improved diagnostic accuracy across 4,928 case-reader pairs (91.7% w/ AI vs. 83.8% w/o AI). AI assistance also reduced median diagnostic time by 19.6%, increased diagnostic confidence by 8.7%, and improved inter-rater agreement from moderate (kappa = 0.56) to substantial (kappa = 0.76). Together, these evaluations support PulmoFoundation as a clinically validated decision-support system for lung pathology.

Runtao Liu, Ziyi Liu, Jiaqi Tang et al.

Recent advances in multimodal LLMs and systems that use tools for long-video QA point to the promise of reasoning over hour-long episodes. However, many methods still compress content into lossy summaries or rely on limited toolsets, weakening temporal grounding and missing fine-grained cues. We propose a multi-agent framework in which a master LLM coordinates a grounding agent to localize question-relevant segments and a vision agent to extract targeted textual observations. The master agent plans with a step limit, and is trained with reinforcement learning to encourage concise, correct, and efficient multi-agent cooperation. This design helps the master agent focus on relevant clips via grounding, complements subtitles with visual detail, and yields interpretable trajectories. On our proposed LongTVQA and LongTVQA+ which are episode-level datasets aggregated from TVQA/TVQA+, our multi-agent system significantly outperforms strong non-agent baselines. Experiments also show reinforcement learning further strengthens reasoning and planning for the trained agent. Code and data will be shared at https://longvideoagent.github.io/.

Ziyi Liu, Idan Attias, Daniel M. Roy

In this work, we investigate the problem of adapting to the presence or absence of causal structure in multi-armed bandit problems. In addition to the usual reward signal, we assume the learner has access to additional variables, observed in each round after acting. When these variables $d$-separate the action from the reward, existing work in causal bandits demonstrates that one can achieve strictly better (minimax) rates of regret (Lu et al., 2020). Our goal is to adapt to this favorable "conditionally benign" structure, if it is present in the environment, while simultaneously recovering worst-case minimax regret, if it is not. Notably, the learner has no prior knowledge of whether the favorable structure holds. In this paper, we establish the Pareto optimal frontier of adaptive rates. We prove upper and matching lower bounds on the possible trade-offs in the performance of learning in conditionally benign and arbitrary environments, resolving an open question raised by Bilodeau et al. (2022). Furthermore, we are the first to obtain instance-dependent bounds for causal bandits, by reducing the problem to the linear bandit setting. Finally, we examine the common assumption that the marginal distributions of the post-action contexts are known and show that a nontrivial estimate is necessary for better-than-worst-case minimax rates.

Ziyi Liu, Dengpan Ye, Long Tang et al.

With the development of artificial intelligence, neural networks play a key role in network intrusion detection systems (NIDS). Despite the tremendous advantages, neural networks are susceptible to adversarial attacks. To improve the reliability of NIDS, many research has been conducted and plenty of solutions have been proposed. However, the existing solutions rarely consider the adversarial attacks against recurrent neural networks (RNN) with time steps, which would greatly affect the application of NIDS in real world. Therefore, we first propose a novel RNN adversarial attack model based on feature reconstruction called \textbf{T}emporal adversarial \textbf{E}xamples \textbf{A}ttack \textbf{M}odel \textbf{(TEAM)}, which applied to time series data and reveals the potential connection between adversarial and time steps in RNN. That is, the past adversarial examples within the same time steps can trigger further attacks on current or future original examples. Moreover, TEAM leverages Time Dilation (TD) to effectively mitigates the effect of temporal among adversarial examples within the same time steps. Experimental results show that in most attack categories, TEAM improves the misjudgment rate of NIDS on both black and white boxes, making the misjudgment rate reach more than 96.68%. Meanwhile, the maximum increase in the misjudgment rate of the NIDS for subsequent original samples exceeds 95.57%.

Jingyuan Huang, Jen-tse Huang, Ziyi Liu et al. · pku, tencent-ai

Visual-Language Models (VLMs) have shown remarkable performance across various tasks, particularly in recognizing geographic information from images. However, VLMs still show regional biases in this task. To systematically evaluate these issues, we introduce a benchmark consisting of 1,200 images paired with detailed geographic metadata. Evaluating four VLMs, we find that while these models demonstrate the ability to recognize geographic information from images, achieving up to 53.8% accuracy in city prediction, they exhibit significant biases. Specifically, performance is substantially higher for economically developed and densely populated regions compared to less developed (-12.5%) and sparsely populated (-17.0%) areas. Moreover, regional biases of frequently over-predicting certain locations remain. For instance, they consistently predict Sydney for images taken in Australia, shown by the low entropy scores for these countries. The strong performance of VLMs also raises privacy concerns, particularly for users who share images online without the intent of being identified. Our code and dataset are publicly available at https://github.com/uscnlp-lime/FairLocator.

Yunming Zhang, Dengpan Ye, Caiyun Xie et al.

The malicious applications of deep forgery, represented by face swapping, have introduced security threats such as misinformation dissemination and identity fraud. While some research has proposed the use of robust watermarking methods to trace the copyright of facial images for post-event traceability, these methods cannot effectively prevent the generation of forgeries at the source and curb their dissemination. To address this problem, we propose a novel comprehensive active defense mechanism that combines traceability and adversariality, called Dual Defense. Dual Defense invisibly embeds a single robust watermark within the target face to actively respond to sudden cases of malicious face swapping. It disrupts the output of the face swapping model while maintaining the integrity of watermark information throughout the entire dissemination process. This allows for watermark extraction at any stage of image tracking for traceability. Specifically, we introduce a watermark embedding network based on original-domain feature impersonation attack. This network learns robust adversarial features of target facial images and embeds watermarks, seeking a well-balanced trade-off between watermark invisibility, adversariality, and traceability through perceptual adversarial encoding strategies. Extensive experiments demonstrate that Dual Defense achieves optimal overall defense success rates and exhibits promising universality in anti-face swapping tasks and dataset generalization ability. It maintains impressive adversariality and traceability in both original and robust settings, surpassing current forgery defense methods that possess only one of these capabilities, including CMUA-Watermark, Anti-Forgery, FakeTagger, or PGD methods.

Kai Qian, Weijie Shi, Jiaqi Wang et al.

Multimodal fake news detection (MFND) aims to verify news credibility by jointly exploiting textual and visual evidence. However, real-world news dissemination frequently suffers from missing modality due to deleted images, corrupted screenshots, and similar issues. Thus, robust detection in this scenario requires preserving strong verification ability for each modality, which is challenging in MFND due to insufficient learning of the low-contribution modality and scarce unimodal annotations. To address this issue, we propose Head-wise Modality Specialization within Multimodal Large Language Models (MLLMs) for robust MFND under missing modality. Specifically, we first systematically study attention heads in MLLMs and their relationship with performance under missing modality, showing that modality-critical heads serve as key carriers of unimodal verification ability through their modality specialization. Based on this observation, to better preserve verification ability for the low-contribution modality, we introduce a head-wise specialization mechanism that explicitly allocates these heads to different modalities and preserves their specialization through lower-bound attention constraints. Furthermore, to better exploit scarce unimodal annotations, we propose a Unimodal Knowledge Retention strategy that prevents these heads from drifting away from the unimodal knowledge learned from limited supervision. Experiments show that our method improves robustness under missing modality while preserving performance with full multimodal input.

Ziyi Liu, Qingyue Long, Zhiwen Xue et al.

User mobility trajectory and mobile traffic data are essential for a wide spectrum of applications including urban planning, network optimization, and emergency management. However, large-scale and fine-grained mobility data remains difficult to obtain due to privacy concerns and collection costs, making it essential to simulate realistic mobility and traffic patterns. User trajectories and mobile traffic are fundamentally coupled, reflecting both physical mobility and cyber behavior in urban environments. Despite this strong interdependence, existing studies often model them separately, limiting the ability to capture cross-modal dynamics. Therefore, a unified framework is crucial. In this paper, we propose MSTDiff, a Multi-Scale Diffusion Transformer for joint simulation of mobile traffic and user trajectories. First, MSTDiff applies discrete wavelet transforms for multi-resolution traffic decomposition. Second, it uses a hybrid denoising network to process continuous traffic volumes and discrete location sequences. A transition mechanism based on urban knowledge graph embedding similarity is designed to guide semantically informed trajectory generation. Finally, a multi-scale Transformer with cross-attention captures dependencies between trajectories and traffic. Experiments show that MSTDiff surpasses state-of-the-art baselines in traffic and trajectory generation tasks, reducing Jensen-Shannon divergence (JSD) across key statistical metrics by up to 17.38% for traffic generation, and by an average of 39.53% for trajectory generation. The source code is available at: https://github.com/tsinghua-fib-lab/MSTDiff .

Haitao Li, Ziyu Li, Yiheng Mao et al.

The advent of multimodal large language models (MLLMs) has sparked interest in their application to electrocardiogram (ECG) analysis. However, existing ECG-focused MLLMs primarily focus on report generation tasks, often limited to single 12-lead, short-duration (10s) ECG inputs, thereby underutilizing the potential of MLLMs. To this end, we aim to develop a MLLM for ECG analysis that supports a broader range of tasks and more flexible ECG inputs. However, existing ECG-QA datasets are often monotonous. To address this gap, we first constructed the anyECG dataset, which encompasses a wide variety of tasks, including report generation, abnormal waveform localization, and open-ended question answering. In addition to standard hospital ECGs, we introduced long-duration reduced-lead ECGs for home environments and multiple ECG comparison scenarios commonly encountered in clinical practice. Furthermore, we propose the anyECG-chat model, which supports dynamic-length ECG inputs and multiple ECG inputs. We trained the model using a three-stage curriculum training recipe with the anyECG dataset. A comprehensive evaluation was conducted, demonstrating that anyECG-chat is capable of supporting various practical application scenarios, including not only common report generation tasks but also abnormal waveform localization for long-duration reduced-lead ECGs in home environments and comprehensive comparative analysis of multiple ECGs. Our code and data are available at: https://github.com/CuCl-2/anyECG-chat.

Ziyi Liu, Claudio Affolter, Sidi Wu et al.

Historical maps provide valuable information and knowledge about the past. However, as they often feature non-standard projections, hand-drawn styles, and artistic elements, it is challenging for non-experts to identify and interpret them. While existing image captioning methods have achieved remarkable success on natural images, their performance on maps is suboptimal as maps are underrepresented in their pre-training process. Despite the recent advance of GPT-4 in text recognition and map captioning, it still has a limited understanding of maps, as its performance wanes when texts (e.g., titles and legends) in maps are missing or inaccurate. Besides, it is inefficient or even impractical to fine-tune the model with users' own datasets. To address these problems, we propose a novel and lightweight map-captioning counterpart. Specifically, we fine-tune the state-of-the-art vision-language model CLIP to generate captions relevant to historical maps and enrich the captions with GPT-3.5 to tell a brief story regarding where, what, when and why of a given map. We propose a novel decision tree architecture to only generate captions relevant to the specified map type. Our system shows invariance to text alterations in maps. The system can be easily adapted and extended to other map types and scaled to a larger map captioning system. The code is open-sourced at https://github.com/claudaff/automatic-map-storytelling.

Jingjing Jiang, Ziyi Liu, Yifan Liu et al.

Encouraging progress has been made towards Visual Question Answering (VQA) in recent years, but it is still challenging to enable VQA models to adaptively generalize to out-of-distribution (OOD) samples. Intuitively, recompositions of existing visual concepts (\ie, attributes and objects) can generate unseen compositions in the training set, which will promote VQA models to generalize to OOD samples. In this paper, we formulate OOD generalization in VQA as a compositional generalization problem and propose a graph generative modeling-based training scheme (X-GGM) to implicitly model the problem. X-GGM leverages graph generative modeling to iteratively generate a relation matrix and node representations for the predefined graph that utilizes attribute-object pairs as nodes. Furthermore, to alleviate the unstable training issue in graph generative modeling, we propose a gradient distribution consistency loss to constrain the data distribution with adversarial perturbations and the generated distribution. The baseline VQA model (LXMERT) trained with the X-GGM scheme achieves state-of-the-art OOD performance on two standard VQA OOD benchmarks, \ie, VQA-CP v2 and GQA-OOD. Extensive ablation studies demonstrate the effectiveness of X-GGM components. Code is available at \url{https://github.com/jingjing12110/x-ggm}.

Jie Yang, Ziyi Liu, Ulian Shahnovich et al.

In this paper, we present a novel tone mapping algorithm that can be used for displaying wide dynamic range (WDR) images on low dynamic range (LDR) devices. The proposed algorithm is mainly motivated by the logarithmic response and local adaptation features of the human visual system (HVS). HVS perceives luminance differently when under different adaptation levels, and therefore our algorithm uses functions built upon different scales to tone map pixels to different values. Functions of large scales are used to maintain image brightness consistency and functions of small scales are used to preserve local detail and contrast. An efficient method using local variance has been proposed to fuse the values of different scales and to remove artifacts. The algorithm utilizes integral images and integral histograms to reduce computation complexity and processing time. Experimental results show that the proposed algorithm can generate high brightness, good contrast, and appealing images that surpass the performance of many state-of-the-art tone mapping algorithms. This project is available at https://github.com/jieyang1987/ToneMapping-Based-on-Multi-scale-Histogram-Synthesis.

Jie Yang, Ziyi Liu, Mengchen Lin et al.

Wide dynamic range (WDR) images contain more scene details and contrast when compared to common images. However, it requires tone mapping to process the pixel values in order to display properly. The details of WDR images can diminish during the tone mapping process. In this work, we address the problem by combining a novel reformulated Laplacian pyramid and deep learning. The reformulated Laplacian pyramid always decompose a WDR image into two frequency bands where the low-frequency band is global feature-oriented, and the high-frequency band is local feature-oriented. The reformulation preserves the local features in its original resolution and condenses the global features into a low-resolution image. The generated frequency bands are reconstructed and fine-tuned to output the final tone mapped image that can display on the screen with minimum detail and contrast loss. The experimental results demonstrate that the proposed method outperforms state-of-the-art WDR image tone mapping methods. The code is made publicly available at https://github.com/linmc86/Deep-Reformulated-Laplacian-Tone-Mapping.

Zhoujian Sun, Cheng Luo, Ziyi Liu et al.

The development of large language models (LLMs) has brought unprecedented possibilities for artificial intelligence (AI) based medical diagnosis. However, the application perspective of LLMs in real diagnostic scenarios is still unclear because they are not adept at collecting patient data proactively. This study presents a LLM-based diagnostic system that enhances planning capabilities by emulating doctors. Our system involves two external planners to handle planning tasks. The first planner employs a reinforcement learning approach to formulate disease screening questions and conduct initial diagnoses. The second planner uses LLMs to parse medical guidelines and conduct differential diagnoses. By utilizing real patient electronic medical record data, we constructed simulated dialogues between virtual patients and doctors and evaluated the diagnostic abilities of our system. We demonstrated that our system obtained impressive performance in both disease screening and differential diagnoses tasks. This research represents a step towards more seamlessly integrating AI into clinical settings, potentially enhancing the accuracy and accessibility of medical diagnostics.

Yangcen Liu, Ziyi Liu, Yuanhao Zhai et al.

Weakly-supervised temporal action localization (WTAL) aims to recognize and localize action instances with only video-level labels. Despite the significant progress, existing methods suffer from severe performance degradation when transferring to different distributions and thus may hardly adapt to real-world scenarios . To address this problem, we propose the Generalizable Temporal Action Localization task (GTAL), which focuses on improving the generalizability of action localization methods. We observed that the performance decline can be primarily attributed to the lack of generalizability to different action scales. To address this problem, we propose STAT (Self-supervised Temporal Adaptive Teacher), which leverages a teacher-student structure for iterative refinement. Our STAT features a refinement module and an alignment module. The former iteratively refines the model's output by leveraging contextual information and helps adapt to the target scale. The latter improves the refinement process by promoting a consensus between student and teacher models. We conduct extensive experiments on three datasets, THUMOS14, ActivityNet1.2, and HACS, and the results show that our method significantly improves the Baseline methods under the cross-distribution evaluation setting, even approaching the same-distribution evaluation performance.

Yunming Zhang, Dengpan Ye, Caiyun Xie et al.

The wide deployment of Face Recognition (FR) systems poses privacy risks. One countermeasure is adversarial attack, deceiving unauthorized malicious FR, but it also disrupts regular identity verification of trusted authorizers, exacerbating the potential threat of identity impersonation. To address this, we propose the first double identity protection scheme based on traceable adversarial watermarking, termed DIP-Watermark. DIP-Watermark employs a one-time watermark embedding to deceive unauthorized FR models and allows authorizers to perform identity verification by extracting the watermark. Specifically, we propose an information-guided adversarial attack against FR models. The encoder embeds an identity-specific watermark into the deep feature space of the carrier, guiding recognizable features of the image to deviate from the source identity. We further adopt a collaborative meta-optimization strategy compatible with sub-tasks, which regularizes the joint optimization direction of the encoder and decoder. This strategy enhances the representation of universal carrier features, mitigating multi-objective optimization conflicts in watermarking. Experiments confirm that DIP-Watermark achieves significant attack success rates and traceability accuracy on state-of-the-art FR models, exhibiting remarkable robustness that outperforms the existing privacy protection methods using adversarial attacks and deep watermarking, or simple combinations of the two. Our work potentially opens up new insights into proactive protection for FR privacy.

Ziyi Liu, Yangcen Liu

Weakly-supervised Temporal Action Localization (WTAL) has achieved notable success but still suffers from a lack of temporal annotations, leading to a performance and framework gap compared with fully-supervised methods. While recent approaches employ pseudo labels for training, three key challenges: generating high-quality pseudo labels, making full use of different priors, and optimizing training methods with noisy labels remain unresolved. Due to these perspectives, we propose PseudoFormer, a novel two-branch framework that bridges the gap between weakly and fully-supervised Temporal Action Localization (TAL). We first introduce RickerFusion, which maps all predicted action proposals to a global shared space to generate pseudo labels with better quality. Subsequently, we leverage both snippet-level and proposal-level labels with different priors from the weak branch to train the regression-based model in the full branch. Finally, the uncertainty mask and iterative refinement mechanism are applied for training with noisy pseudo labels. PseudoFormer achieves state-of-the-art WTAL results on the two commonly used benchmarks, THUMOS14 and ActivityNet1.3. Besides, extensive ablation studies demonstrate the contribution of each component of our method.

Jiabo MA, Wenqiang Li, Jinbang Li et al.

Accurate histopathological diagnosis often requires multiple differently stained tissue sections, a process that is time-consuming, labor-intensive, and environmentally taxing due to the use of multiple chemical stains. Recently, virtual staining has emerged as a promising alternative that is faster, tissue-conserving, and environmentally friendly. However, existing virtual staining methods face significant challenges in clinical applications, primarily due to their reliance on well-aligned paired data. Obtaining such data is inherently difficult because chemical staining processes can distort tissue structures, and a single tissue section cannot undergo multiple staining procedures without damage or loss of information. As a result, most available virtual staining datasets are either unpaired or roughly paired, making it difficult for existing methods to achieve accurate pixel-level supervision. To address this challenge, we propose a robust virtual staining framework featuring cascaded registration mechanisms to resolve spatial mismatches between generated outputs and their corresponding ground truth. Experimental results demonstrate that our method significantly outperforms state-of-the-art models across five datasets, achieving an average improvement of 3.2% on internal datasets and 10.1% on external datasets. Moreover, in datasets with substantial misalignment, our approach achieves a remarkable 23.8% improvement in peak signal-to-noise ratio compared to baseline models. The exceptional robustness of the proposed method across diverse datasets simplifies the data acquisition process for virtual staining and offers new insights for advancing its development.

Zhe Xu, Cheng Jin, Yihui Wang et al.

Multimodal pathological image understanding has garnered widespread interest due to its potential to improve diagnostic accuracy and enable personalized treatment through integrated visual and textual data. However, existing methods exhibit limited reasoning capabilities, which hamper their ability to handle complex diagnostic scenarios. Additionally, the enormous size of pathological images leads to severe computational burdens, further restricting their practical deployment. To address these limitations, we introduce a novel bilateral reinforcement learning framework comprising two synergistic branches. One reinforcement branch enhances the reasoning capability by enabling the model to learn task-specific decision processes, i.e., pathology rationales, directly from labels without explicit reasoning supervision. While the other branch dynamically allocates a tailored number of tokens to different images based on both their visual content and task context, thereby optimizing computational efficiency. We apply our method to various pathological tasks such as visual question answering, cancer subtyping, and lesion detection. Extensive experiments show an average +41.7 absolute performance improvement with 70.3% lower inference costs over the base models, achieving both reasoning accuracy and computational efficiency.

Haitao Li, Che Liu, Zhengyao Ding et al.

Electrocardiograms (ECGs) are essential for diagnosing cardiovascular diseases. However, existing ECG-Report contrastive learning methods focus on whole-ECG and report alignment, missing the link between local ECG features and individual report tags. In this paper, we propose FG-CLEP (Fine-Grained Contrastive Language ECG Pre-training), which achieves fine-grained alignment between specific ECG segments and each tag in the report via tag-specific ECG representations. Furthermore, we found that nearly 55\% of ECG reports in the MIMIC-ECG training dataset lack detailed waveform features, which hinders fine-grained alignment. To address this, we introduce a coarse-to-fine training process that leverages large language models (LLMs) to recover these missing waveform features and validate the LLM outputs using a coarse model. Additionally, fine-grained alignment at the tag level, rather than at the report level, exacerbates the false negative problem, as different reports may share common tags. To mitigate this, we introduce a semantic similarity matrix to guide the model in identifying and correcting false negatives. Experiments on six datasets demonstrate that FG-CLEP significantly improves fine-grained alignment, outperforming state-of-the-art methods in both zero-shot prediction and linear probing. Meanwhile, the fine-grained reports we generate also enhance the performance of other methods.

Ziyi Liu, Phuc Luong, Mario Boley et al.

Gaussian process regression is a popular model in the small data regime due to its sound uncertainty quantification and the exploitation of the smoothness of the regression function that is encountered in a wide range of practical problems. However, Gaussian processes perform sub-optimally when the degree of smoothness is non-homogeneous across the input domain. Random forest regression partially addresses this issue by providing local basis functions of variable support set sizes that are chosen in a data-driven way. However, they do so at the expense of forgoing any degree of smoothness, which often results in poor performance in the small data regime. Here, we aim to combine the advantages of both models by applying a kernel-based smoothing mechanism to a learned random forest or any other piecewise constant prediction function. As we demonstrate empirically, the resulting model consistently improves the predictive performance of the underlying random forests and, in almost all test cases, also improves the log loss of the usual uncertainty quantification based on inter-tree variance. The latter advantage can be attributed to the ability of the smoothing model to take into account the uncertainty over the exact tree-splitting locations.

Zhoujian Sun, Ziyi Liu, Cheng Luo et al.

Recent advances in large language models (LLMs) have shown promising results in medical diagnosis, with some studies indicating superior performance compared to human physicians in specific scenarios. However, the diagnostic capabilities of LLMs are often overestimated, as their performance significantly deteriorates in interactive diagnostic settings that require active information gathering. This study investigates the underlying mechanisms behind the performance degradation phenomenon and proposes a solution. We identified that the primary deficiency of LLMs lies in the initial diagnosis phase, particularly in information-gathering efficiency and initial diagnosis formation, rather than in the subsequent differential diagnosis phase. To address this limitation, we developed a plug-and-play method enhanced (PPME) LLM agent, leveraging over 3.5 million electronic medical records from Chinese and American healthcare facilities. Our approach integrates specialized models for initial disease diagnosis and inquiry into the history of the present illness, trained through supervised and reinforcement learning techniques. The experimental results indicate that the PPME LLM achieved over 30% improvement compared to baselines. The final diagnostic accuracy of the PPME LLM in interactive diagnostic scenarios approached levels comparable to those achieved using complete clinical data. These findings suggest a promising potential for developing autonomous diagnostic systems, although further validation studies are needed.

Ziyi Liu, Bahar Sarrafzadeh, Pei Zhou et al.

While Large Language Models (LLMs) are increasingly used in agentic frameworks to assist individual users, there is a growing need for agents that can proactively manage complex, multi-party collaboration. Systematic evaluation methods for such proactive agents remain scarce, limiting progress in developing AI that can effectively support multiple people together. Negotiation offers a demanding testbed for this challenge, requiring socio-cognitive intelligence to navigate conflicting interests between multiple participants and multiple topics and build consensus. Here, we present ProMediate, the first framework for evaluating proactive AI mediator agents in complex, multi-topic, multi-party negotiations. ProMediate consists of two core components: (i) a simulation testbed based on realistic negotiation cases and theory-driven difficulty levels (ProMediate-Easy, ProMediate-Medium, and ProMediate-Hard), with a plug-and-play proactive AI mediator grounded in socio-cognitive mediation theories, capable of flexibly deciding when and how to intervene; and (ii) a socio-cognitive evaluation framework with a new suite of metrics to measure consensus changes, intervention latency, mediator effectiveness, and intelligence. Together, these components establish a systematic framework for assessing the socio-cognitive intelligence of proactive AI agents in multi-party settings. Our results show that a socially intelligent mediator agent outperforms a generic baseline, via faster, better-targeted interventions. In the ProMediate-Hard setting, our social mediator increases consensus change by 3.6 percentage points compared to the generic baseline (10.65\% vs 7.01\%) while being 77\% faster in response (15.98s vs. 3.71s). In conclusion, ProMediate provides a rigorous, theory-grounded testbed to advance the development of proactive, socially intelligent agents.

Fuxiang Huang, Jiayi Zhu, Yunfang Yu et al.

Breast cancer is the most commonly diagnosed cancer and the leading cause of cancer-related mortality in women globally. Mammography is essential for the early detection and diagnosis of breast lesions. Despite recent progress in foundation models (FMs) for mammogram analysis, their clinical translation remains constrained by several fundamental limitations, including insufficient diversity in training data, limited model generalizability, and a lack of comprehensive evaluation across clinically relevant tasks. Here, we introduce VersaMammo, a versatile foundation model for mammograms, designed to overcome these limitations. We curated the largest multi-institutional mammogram dataset to date, comprising 706,239 images from 21 sources. To improve generalization, we propose a two-stage pre-training strategy to develop VersaMammo, a mammogram foundation model. First, a teacher model is trained via self-supervised learning to extract transferable features from unlabeled mammograms. Then, supervised learning combined with knowledge distillation transfers both features and clinical knowledge into VersaMammo. To ensure a comprehensive evaluation, we established a benchmark comprising 92 specific tasks, including 68 internal tasks and 24 external validation tasks, spanning 5 major clinical task categories: lesion detection, segmentation, classification, image retrieval, and visual question answering. VersaMammo achieves state-of-the-art performance, ranking first in 50 out of 68 specific internal tasks and 20 out of 24 external validation tasks, with average ranks of 1.5 and 1.2, respectively. These results demonstrate its superior generalization and clinical utility, offering a substantial advancement toward reliable and scalable breast cancer screening and diagnosis.

Ziyi Liu

Large Language Models (LLMs) struggle with information forgetting and inefficiency in long-horizon, multi-turn dialogues. To address this, we propose a training-free prompt engineering method, the State-Update Multi-turn Dialogue Strategy. It utilizes "State Reconstruction" and "History Remind" mechanisms to effectively manage dialogue history. Our strategy shows strong performance across multiple multi-hop QA datasets. For instance, on the HotpotQA dataset, it improves the core information filtering score by 32.6%, leading to a 14.1% increase in the downstream QA score, while also reducing inference time by 73.1% and token consumption by 59.4%. Ablation studies confirm the pivotal roles of both components. Our work offers an effective solution for optimizing LLMs in long-range interactions, providing new insights for developing more robust Agents.

Ziyi Liu, Zhe Xu, Jiabo Ma et al.

Foundation models have revolutionized computational pathology by achieving remarkable success in high-level diagnostic tasks, yet the critical challenge of low-level image enhancement remains largely unaddressed. Real-world pathology images frequently suffer from degradations such as noise, blur, and low resolution due to slide preparation artifacts, staining variability, and imaging constraints, while the reliance on physical staining introduces significant costs, delays, and inconsistency. Although existing methods target individual problems like denoising or super-resolution, their task-specific designs lack the versatility to handle the diverse low-level vision challenges encountered in practice. To bridge this gap, we propose the first unified Low-level Pathology Foundation Model (LPFM), capable of enhancing image quality in restoration tasks, including super-resolution, deblurring, and denoising, as well as facilitating image translation tasks like virtual staining (H&E and special stains), all through a single adaptable architecture. Our approach introduces a contrastive pre-trained encoder that learns transferable, stain-invariant feature representations from 190 million unlabeled pathology images, enabling robust identification of degradation patterns. A unified conditional diffusion process dynamically adapts to specific tasks via textual prompts, ensuring precise control over output quality. Trained on a curated dataset of 87,810 whole slied images (WSIs) across 34 tissue types and 5 staining protocols, LPFM demonstrates statistically significant improvements (p<0.01) over state-of-the-art methods in most tasks (56/66), achieving Peak Signal-to-Noise Ratio (PSNR) gains of 10-15% for image restoration and Structural Similarity Index Measure (SSIM) improvements of 12-18% for virtual staining.

Yifan Wang, Ziyi Liu, Xiaolong Sun et al.

Temporal sentence grounding (TSG) is a highly challenging task aiming to localize the temporal segment within an untrimmed video corresponding to a given natural language description. Benefiting from the design of learnable queries, the DETR-based models have achieved substantial advancements in the TSG task. However, the absence of explicit supervision often causes the learned queries to overlap in roles, leading to redundant predictions. Therefore, we propose to improve TSG by making each query fulfill its designated role, leveraging the length priors of the video-description pairs. In this paper, we introduce the Length-Aware Transformer (LATR) for TSG, which assigns different queries to handle predictions based on varying temporal lengths. Specifically, we divide all queries into three groups, responsible for segments with short, middle, and long temporal durations, respectively. During training, an additional length classification task is introduced. Predictions from queries with mismatched lengths are suppressed, guiding each query to specialize in its designated function. Extensive experiments demonstrate the effectiveness of our LATR, achieving state-of-the-art performance on three public benchmarks. Furthermore, the ablation studies validate the contribution of each component of our method and the critical role of incorporating length priors into the TSG task.

Penghan Zhu, Shurui Mei, Shushan Chen et al.

This study proposes a deep learning-based framework for automated segmentation of brain regions and classification of amyloid positivity using positron emission tomography (PET) images alone, without the need for structural MRI or CT. A 3D U-Net architecture with four layers of depth was trained and validated on a dataset of 200 F18-florbetapir amyloid-PET scans, with an 130/20/50 train/validation/test split. Segmentation performance was evaluated using Dice similarity coefficients across 30 brain regions, with scores ranging from 0.45 to 0.88, demonstrating high anatomical accuracy, particularly in subcortical structures. Quantitative fidelity of PET uptake within clinically relevant regions. Precuneus, prefrontal cortex, gyrus rectus, and lateral temporal cortex was assessed using normalized root mean square error, achieving values as low as 0.0011. Furthermore, the model achieved a classification accuracy of 0.98 for amyloid positivity based on regional uptake quantification, with an area under the ROC curve (AUC) of 0.99. These results highlight the model's potential for integration into PET only diagnostic pipelines, particularly in settings where structural imaging is not available. This approach reduces dependence on coregistration and manual delineation, enabling scalable, reliable, and reproducible analysis in clinical and research applications. Future work will focus on clinical validation and extension to diverse PET tracers including C11 PiB and other F18 labeled compounds.

Zhe Xu, Ziyi Liu, Junlin Hou et al.

Multimodal large language models (MLLMs) have emerged as powerful tools for computational pathology, offering unprecedented opportunities to integrate pathological images with language context for comprehensive diagnostic analysis. These models hold particular promise for automating complex tasks that traditionally require expert interpretation of pathologists. However, current MLLM approaches in pathology demonstrate significantly constrained reasoning capabilities, primarily due to their reliance on expensive chain-of-thought annotations. Additionally, existing methods remain limited to simplex application of visual question answering (VQA) at the region-of-interest (ROI) level, failing to address the full spectrum of diagnostic needs such as ROI classification, detection, segmentation, whole-slide-image (WSI) classification and VQA in clinical practice. In this study, we present SmartPath-R1, a versatile MLLM capable of simultaneously addressing both ROI-level and WSI-level tasks while demonstrating robust pathological reasoning capability. Our framework combines scale-dependent supervised fine-tuning and task-aware reinforcement fine-tuning, which circumvents the requirement for chain-of-thought supervision by leveraging the intrinsic knowledge within MLLM. Furthermore, SmartPath-R1 integrates multiscale and multitask analysis through a mixture-of-experts mechanism, enabling dynamic processing for diverse tasks. We curate a large-scale dataset comprising 2.3M ROI samples and 188K WSI samples for training and evaluation. Extensive experiments across 72 tasks validate the effectiveness and superiority of the proposed approach. This work represents a significant step toward developing versatile, reasoning-enhanced AI systems for precision pathology.

Ziyi Liu, Priyanka Dey, Jen-tse Huang et al.

Cultural Intelligence (CQ) refers to the ability to understand unfamiliar cultural contexts, a crucial skill for large language models (LLMs) to effectively engage with globally diverse users. Existing studies often focus on explicitly stated cultural norms, but fail to capture the subtle, implicit values that are common in daily conversation. To address this gap, we introduce CQBench, a benchmark specifically designed to assess LLMs' capability to infer implicit cultural values from natural conversational contexts. CQBench consists of multi character conversation based stories using values from the World Value Survey and the GlobalOpinions, with topics including ethical, religious, social, etc. Our automatic dataset construction pipeline integrates rigorous validation procedures (incorporation, consistency, and implicitness checks), achieving a 94.5% human model agreement in the final validation. To leverage CQBench data, we design three tasks of increasing complexity: attitude detection, value selection, and value extraction. These tasks evaluate whether models can detect attitude and recognize values embedded within natural dialogues rather than relying on explicit cultural knowledge. We find that while frontier models like o1 reach human level performance in value selection (0.809 F1), they still fall short in nuanced attitude detection (0.622 F1). Notably, finetuning a smaller LLaMA-3.2-3B on only 500 culturally rich examples improves performance by over 10%, even outperforming o3-mini in some cases. Using CQ-Bench, we provide insights into the current challenges in LLMs' CQ research and suggest practical pathways for enhancing LLMs' cross-cultural reasoning abilities.

Long Tang, Dengpan Ye, Sirun Chen et al.

The fine-tuning technique for text-to-image diffusion models facilitates image customization but risks privacy breaches and opinion manipulation. Current research focuses on prompt- or image-level adversarial attacks for anti-customization, yet it overlooks the correlation between these two levels and the relationship between internal modules and inputs. This hinders anti-customization performance in practical threat scenarios. We propose Dual Anti-Diffusion (DADiff), a two-stage adversarial attack targeting diffusion customization, which, for the first time, integrates the adversarial prompt-level attack into the generation process of image-level adversarial examples. In stage 1, we generate prompt-level adversarial vectors to guide the subsequent image-level attack. In stage 2, besides conducting the end-to-end attack on the UNet model, we disrupt its self- and cross-attention modules, aiming to break the correlations between image pixels and align the cross-attention results computed using instance prompts and adversarial prompt vectors within the images. Furthermore, we introduce a local random timestep gradient ensemble strategy, which updates adversarial perturbations by integrating random gradients from multiple segmented timesets. Experimental results on various mainstream facial datasets demonstrate 10%-30% improvements in cross-prompt, keyword mismatch, cross-model, and cross-mechanism anti-customization with DADiff compared to existing methods.

Haitao Li, Ziyu Li, Yiheng Mao et al.

Multimodal large language models (MLLMs) are increasingly being applied in the medical field, particularly in medical imaging. However, developing MLLMs for ECG signals, which are crucial in clinical settings, has been a significant challenge beyond medical imaging. Previous studies have attempted to address this by converting ECGs into several text tags using an external classifier in a training-free manner. However, this approach significantly compresses the information in ECGs and underutilizes the reasoning capabilities of LLMs. In this work, we directly feed the embeddings of ECGs into the LLM through a projection layer, retaining more information about ECGs and better leveraging the reasoning abilities of LLMs. Our method can also effectively handle a common situation in clinical practice where it is necessary to compare two ECGs taken at different times. Recent studies found that MLLMs may rely solely on text input to provide answers, ignoring inputs from other modalities. We analyzed this phenomenon from a causal perspective in the context of ECG MLLMs and discovered that the confounder, severity of illness, introduces a spurious correlation between the question and answer, leading the model to rely on this spurious correlation and ignore the ECG input. Such models do not comprehend the ECG input and perform poorly in adversarial tests where different expressions of the same question are used in the training and testing sets. We designed a de-biased pre-training method to eliminate the confounder's effect according to the theory of backdoor adjustment. Our model performed well on the ECG-QA task under adversarial testing and demonstrated zero-shot capabilities. An interesting random ECG test further validated that our model effectively understands and utilizes the input ECG signal.

Zhengyao Ding, Ziyu Li, Yujian Hu et al.

Cardiovascular diseases (CVDs) are the leading cause of global mortality, necessitating accessible and accurate diagnostic tools. While cardiac magnetic resonance imaging (CMR) provides gold-standard insights into cardiac structure and function, its clinical utility is limited by high cost and complexity. In contrast, electrocardiography (ECG) is inexpensive and widely available but lacks the granularity of CMR. We propose CardioNets, a deep learning framework that translates 12-lead ECG signals into CMR-level functional parameters and synthetic images, enabling scalable cardiac assessment. CardioNets integrates cross-modal contrastive learning and generative pretraining, aligning ECG with CMR-derived cardiac phenotypes and synthesizing high-resolution CMR images via a masked autoregressive model. Trained on 159,819 samples from five cohorts, including the UK Biobank (n=42,483) and MIMIC-IV-ECG (n=164,550), and externally validated on independent clinical datasets (n=3,767), CardioNets achieved strong performance across disease screening and phenotype estimation tasks. In the UK Biobank, it improved cardiac phenotype regression R2 by 24.8% and cardiomyopathy AUC by up to 39.3% over baseline models. In MIMIC, it increased AUC for pulmonary hypertension detection by 5.6%. Generated CMR images showed 36.6% higher SSIM and 8.7% higher PSNR than prior approaches. In a reader study, ECG-only CardioNets achieved 13.9% higher accuracy than human physicians using both ECG and real CMR. These results suggest that CardioNets offers a promising, low-cost alternative to CMR for large-scale CVD screening, particularly in resource-limited settings. Future efforts will focus on clinical deployment and regulatory validation of ECG-based synthetic imaging.

Isabelle Lee, Joshua Lum, Ziyi Liu et al.

While interpretability research has shed light on some internal algorithms utilized by transformer-based LLMs, reasoning in natural language, with its deep contextuality and ambiguity, defies easy categorization. As a result, formulating clear and motivating questions for circuit analysis that rely on well-defined in-domain and out-of-domain examples required for causal interventions is challenging. Although significant work has investigated circuits for specific tasks, such as indirect object identification (IOI), deciphering natural language reasoning through circuits remains difficult due to its inherent complexity. In this work, we take initial steps to characterize causal reasoning in LLMs by analyzing clear-cut cause-and-effect sentences like "I opened an umbrella because it started raining," where causal interventions may be possible through carefully crafted scenarios using GPT-2 small. Our findings indicate that causal syntax is localized within the first 2-3 layers, while certain heads in later layers exhibit heightened sensitivity to nonsensical variations of causal sentences. This suggests that models may infer reasoning by (1) detecting syntactic cues and (2) isolating distinct heads in the final layers that focus on semantic relationships.

Ziyi Liu, Abhishek Anand, Pei Zhou et al.

Large language models (LLMs) have demonstrated the potential to mimic human social intelligence. However, most studies focus on simplistic and static self-report or performance-based tests, which limits the depth and validity of the analysis. In this paper, we developed a novel framework, InterIntent, to assess LLMs' social intelligence by mapping their ability to understand and manage intentions in a game setting. We focus on four dimensions of social intelligence: situational awareness, self-regulation, self-awareness, and theory of mind. Each dimension is linked to a specific game task: intention selection, intention following, intention summarization, and intention guessing. Our findings indicate that while LLMs exhibit high proficiency in selecting intentions, achieving an accuracy of 88%, their ability to infer the intentions of others is significantly weaker, trailing human performance by 20%. Additionally, game performance correlates with intention understanding, highlighting the importance of the four components towards success in this game. These findings underline the crucial role of intention understanding in evaluating LLMs' social intelligence and highlight the potential of using social deduction games as a complex testbed to enhance LLM evaluation. InterIntent contributes a structured approach to bridging the evaluation gap in social intelligence within multiplayer games.

Yuanhao Zhai, Ziyi Liu, Zhenyu Wu et al.

Deep learning models have a risk of utilizing spurious clues to make predictions, such as recognizing actions based on the background scene. This issue can severely degrade the open-set action recognition performance when the testing samples have different scene distributions from the training samples. To mitigate this problem, we propose a novel method, called Scene-debiasing Open-set Action Recognition (SOAR), which features an adversarial scene reconstruction module and an adaptive adversarial scene classification module. The former prevents the decoder from reconstructing the video background given video features, and thus helps reduce the background information in feature learning. The latter aims to confuse scene type classification given video features, with a specific emphasis on the action foreground, and helps to learn scene-invariant information. In addition, we design an experiment to quantify the scene bias. The results indicate that the current open-set action recognizers are biased toward the scene, and our proposed SOAR method better mitigates such bias. Furthermore, our extensive experiments demonstrate that our method outperforms state-of-the-art methods, and the ablation studies confirm the effectiveness of our proposed modules.

Brihi Joshi, Ziyi Liu, Sahana Ramnath et al.

Among the remarkable emergent capabilities of large language models (LMs) is free-text rationalization; beyond a certain scale, large LMs are capable of generating seemingly useful rationalizations, which in turn, can dramatically enhance their performances on leaderboards. This phenomenon raises a question: can machine generated rationales also be useful for humans, especially when lay humans try to answer questions based on those machine rationales? We observe that human utility of existing rationales is far from satisfactory, and expensive to estimate with human studies. Existing metrics like task performance of the LM generating the rationales, or similarity between generated and gold rationales are not good indicators of their human utility. While we observe that certain properties of rationales like conciseness and novelty are correlated with their human utility, estimating them without human involvement is challenging. We show that, by estimating a rationale's helpfulness in answering similar unseen instances, we can measure its human utility to a better extent. We also translate this finding into an automated score, GEN-U, that we propose, which can help improve LMs' ability to generate rationales with better human utility, while maintaining most of its task performance. Lastly, we release all code and collected data with this project.