Large language models (LLMs) present a paradox: they can correctly answer a multi-hop factual query in a high-resource language like English, yet fail on the identical query in another language. This raises a fundamental question about the nature of multilingual knowledge: are facts missing, or merely inaccessible? The underlying mechanisms for this knowledge gap have remained largely unexplored. In this work, we resolve this question by introducing a mechanistic interpretability framework that traces the causal pathways of multi-hop knowledge reasoning. Our analysis reveals a core, non-obvious finding: cross-lingual inconsistencies do not stem from a knowledge deficit. Instead, factual knowledge is robustly stored in a set of **shared, language-agnostic semantic neurons**. The failure originates from **misaligned attention pathways**, where a common set of critical attention heads fails to correctly route information along the reasoning chain to the appropriate knowledge neurons in lower-resource languages. This mechanistic diagnosis motivates a targeted alignment strategy: a surgical fine-tuning of only these critical heads. Experiments demonstrate that our method achieves significant improvements in multilingual multi-hop factuality—with positive cross-lingual transfer—while uniquely preserving general model capabilities, offering a scalable and mechanistically-grounded approach to building more reliable multilingual models.

We introduce Mixture-of-Trees (MoT), a novel framework that integrates sparse expert activation with structured tree-based reasoning for efficient LLM inference. MoT employs a learned gating mechanism to selectively activate only the most relevant expert reasoning trees for each problem, where experts use models of varying capacities based on task complexity. The framework features three key innovations: (1) sparse expert activation through unified gating networks, (2) specialized expert trees that leverage domain-specific expertise while optimizing the quality-efficiency trade-off, and (3) collaborative debate mechanisms for conflicting solutions. Additionally, MoT includes a shared baseline tree with early stopping—activated experts perform lightweight validation and terminate early when confidence is high. Experiments across five benchmarks (GSM8K, MATH, AIME 2024, MMLU, HotpotQA) show that MoT achieves 2-7 percentage point accuracy improvements while reducing LLM calls by 37-40% compared to existing multi-path methods.

Artificial intelligence has demonstrated significant potential in clinical decision-making; however, developing models capable of adapting to diverse real-world scenarios and performing complex diagnostic reasoning remains a major challenge. Existing medical multi-modal benchmarks are typically limited to single-image, single-turn tasks, lacking multi-modal medical image integration and failing to capture the longitudinal and multi-modal interactive nature inherent to clinical practice. To address this gap, we introduce MedAtlas, a novel benchmark framework designed to evaluate large language models on realistic medical reasoning tasks. MedAtlas is characterized by four key features: multi-round visual question answering (VQA), Joint reasoning of multiple modalities of medical images, multi-task integration, and high clinical fidelity. It supports four core tasks: open-ended multi-round VQA, closed-ended multi-round VQA, multi-image joint reasoning, and comprehensive disease diagnosis. Each case is derived from real diagnostic workflows and incorporates temporal interactions between textual medical histories and multiple imaging modalities, including CT, MRI, PET, ultrasound, X-ray, etc., requiring models to perform deep integrative reasoning across images and clinical texts. MedAtlas provides expert-annotated gold standards for all tasks. Furthermore, we propose two novel evaluation metrics: Stage Chain Accuracy (SCA) and Error Propagation Suppression Coefficient (EPSC). Benchmark results with existing multi-modal models reveal substantial performance gaps in multi-stage clinical reasoning. MedAtlas establishes a challenging evaluation platform to advance the development of robust and trustworthy medical AI.

Contrastively trained vision-language models like CLIP, have made remarkable progress in learning joint image-text representations, but still face challenges in compositional understanding.They often exhibit a "bag-of-words" behavior-struggling to capture the object relations, attribute-object bindings, and word order dependencies.This limitation arises not only from the reliance on global, single-vector representations for optimization, but also from the insufficient exploitation and modeling of the rich compositional information inherently present in paired image text data.In this work, we propose MACCO (MAsked Compositional Concept MOdeling), a framework that masks compositional concepts in one modality and reconstructs them conditioned on the full contextual information from the other, enabling the model to capture and align cross-modal compositional structures more effectively.To facilitate this process, we introduce two auxiliary objectives that jointly align and regularize masked features both inter-modally and intramodally.Extensive experiments on five compositional benchmarks, along with in-depth analyses, demonstrate that our approach not only significantly enhances compositionality in VLMs but also improves their ability to capture syntactic structure and linguistic information.Additionally, the improved compositionality also benefits text-to-image generation and multimodal large language model.

Wei Li, Zhen Huang, Houqiang Li, Le Lu, Yang Lu, Xinmei Tian, Xu Shen, Jieping Ye. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Chenghao Sun, Zhen Huang, Yonggang Zhang, Le Lu, Houqiang Li, Xinmei Tian, Xu Shen, Jieping Ye. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Thermal management in 3D ICs is increasingly challenging due to higher power densities. Traditional PDESolving based methods, while accurate, are too slow for iterative design. Machine learning approaches like FNO provide faster alternatives but suffer from high-frequency information loss and high-fidelity data dependency. We introduce Self-Attention UNet Fourier Neural Operator (SAU-FNO), a novel framework combining self-attention and U-Net with FNO to capture longrange dependencies and model local high-frequency features effectively. Transfer learning is employed to fine-tune low-fidelity data, minimizing the need for extensive high-fidelity datasets and speeding up training. Experiments demonstrate that SAUFNO achieves state-of-the-art thermal prediction accuracy and provides an $842 \times$ speedup over traditional FEM methods, making it an efficient tool for advanced 3D IC thermal simulations.

Device models require large numbers of parameters to characterize complex physical effects. Although the latest advancements in machine learning and automated tools have drastically improved efficiency over the classic methods, they still demand a considerable amount of human intervention in the loop to gain accuracy. This drastically limits further automation. Inspired by the success of Multimodal Large Language Models (MLLMs) in addressing tasks across diverse fields, we propose ModelGen, the first in-depth study to leverage MLLMs with RAG (Retrieval-Augmented Generation) to significantly reduce human effort in parameter extraction for compact model. Our contributions include (1) Automated Agentic Workflow Construction that learns to build and refine extraction workflows through iterative optimization, (2) MLLM Judge, a visual scoring mechanism that evaluates fitting quality using actual device characteristic plots rather than simple numerical metrics, and (3) Model-specific RAG for providing relevant domain knowledge during the extraction process. Experimental results demonstrate that ModelGen achieves a 26.8%–33.1% improvement in pass@1,3,5 compared to base LLM methods. The system completes complex model extractions for BSIMs and ASM-HEMT in hours (up to 168× faster) rather than days or weeks, making parameter extraction more accessible to non-experts while maintaining professional engineer-level accuracy.

Contrastively trained vision-language models such as CLIP have achieved remarkable progress in vision and language representation learning.Despite the promising progress, their proficiency in compositional reasoning over attributes and relations (e.g., distinguishing between "the car is underneath the person" and "the person is underneath the car") remains notably inadequate.We investigate the cause for this deficient behavior is the composition attribution issue, where the attribution scores (e.g., attention scores or GradCAM scores) for relations (e.g., underneath) or attributes (e.g., red) in the text are substantially lower than those for object terms.In this work, we show such issue is mitigated via a novel framework called CAE (Composition Attribution Enhancement).This generic framework incorporates various interpretable attribution methods to encourage the model to pay greater attention to composition words denoting relationships and attributes within the text.Detailed analysis shows that our approach enables the models to adjust and rectify the attribution of the texts.Extensive experiments across seven benchmarks reveal that our framework significantly enhances the ability to discern intricate details and construct more sophisticated interpretations of combined visual and linguistic elements.

Large Language Models (LLMs) tend to prioritize adherence to user prompts over providing veracious responses, leading to the sycophancy issue. When challenged by users, LLMs tend to admit mistakes and provide inaccurate responses even if they initially provided the correct answer. Recent works propose to employ supervised fine-tuning (SFT) to mitigate the sycophancy issue, while it typically leads to the degeneration of LLMs' general capability. To address the challenge, we propose a novel supervised pinpoint tuning (SPT), where the region-of-interest modules are tuned for a given objective. Specifically, SPT first reveals and verifies a small percentage (<5%) of the basic modules, which significantly affect a particular behavior of LLMs. i.e., sycophancy. Subsequently, SPT merely fine-tunes these identified modules while freezing the rest. To verify the effectiveness of the proposed SPT, we conduct comprehensive experiments, demonstrating that SPT significantly mitigates the sycophancy issue of LLMs (even better than SFT). Moreover, SPT introduces limited or even no side effects on the general capability of LLMs. Our results shed light on how to precisely, effectively, and efficiently explain and improve the targeted ability of LLMs.

In real teaching scenarios, an excellent teacher always teaches what he (or she) is good at but the student is not. This gives the student the best assistance in making up for his (or her) weaknesses and becoming a good one overall. Enlightened by this, we introduce the "Teaching what you Should Teach" strategy into a knowledge distillation framework, and propose a data-based distillation method named "TST" that searches for desirable augmented samples to assist in distilling more efficiently and rationally. To be specific, we design a neural network-based data augmentation module with priori bias to find out what meets the teacher's strengths but the student's weaknesses, by learning magnitudes and probabilities to generate suitable data samples. By training the data augmentation module and the generalized distillation paradigm alternately, a student model is learned with excellent generalization ability. To verify the effectiveness of our method, we conducted extensive comparative experiments on object recognition, detection, and segmentation tasks. The results on the CIFAR-100, ImageNet-1k, MS-COCO, and Cityscapes datasets demonstrate that our method achieves state-of-the-art performance on almost all teacher-student pairs. Furthermore, we conduct visualization studies to explore what magnitudes and probabilities are needed for the distillation process.

Cloth-Changing person re-identification (CC-ReID) aims at matching the same person across different locations over a long-duration, e.g., over days, and therefore inevitably has cases of changing clothing. In this paper, we focus on handling well the CC-ReID problem under a more challenging setting, i.e., just from a single image, which enables an efficient and latency-free person identity matching for surveillance. Specifically, we introduce Gait recognition as an auxiliary task to drive the Image ReID model to learn cloth-agnostic representations by leveraging personal unique and cloth-independent gait information, we name this framework as GI-ReID. GI-ReID adopts a two-stream architecture that consists of an image ReID-Stream and an auxiliary gait recognition stream (Gait-Stream). The Gait-Stream, that is discarded in the inference for high efficiency, acts as a regulator to encourage the ReID-Stream to capture cloth-invariant biometric motion features during the training. To get temporal continuous motion cues from a single image, we design a Gait Sequence Prediction (GSP) module for Gait-Stream to enrich gait information. Finally, a semantics consistency constraint over two streams is enforced for effective knowledge regularization. Extensive experiments on multiple image-based Cloth-Changing ReID benchmarks, e.g., LTCC, PRCC, Real28, and VC-Clothes, demonstrate that GI-ReID performs favorably against the state-of-the-art methods.

Deep convolutional neural networks (CNNs) have become dominant in the task of single image rain removal. Most of current CNN methods, however, suffer from the problem of overfitting on one single synthetic dataset as they neglect the intrinsic prior of the physical properties of rain streaks. To address this issue, we propose a simple but effective prior - Fourier prior to improve the generalization ability of an image rain removal model. The Fourier prior is a kind of property of rainy images. It is based on a key observation of us - replacing the Fourier amplitude of rainy images with that of clean images greatly suppresses the synthetic and real-world rain streaks. This means the amplitude contains most of the rain streak information and the phase keeps the similar structures of the background. So it is natural for single image rain removal to process the amplitude and phase information of the rainy images separately. In this paper, we develop a two-stage model where the first stage restores the amplitude of rainy images to clean rain streaks, and the second stage restores the phase information to refine fine-grained background structures. Extensive experiments on synthetic rainy data demonstrate the power of Fourier prior. Moreover, when trained on synthetic data, a robust generalization ability to real-world images can also be obtained. The code will be publicly available at https://github.com/willinglucky/ExploringFourier-Prior-for-Single-Image-Rain-Removal.

Knowledge Distillation (KD) is a popular technique to transfer knowledge from a teacher model or ensemble to a student model. Its success is generally attributed to the privileged information on similarities/consistency between the class distributions or intermediate feature representations of the teacher model and the student model. However, directly pushing the student model to mimic the probabilities/features of the teacher model to a large extent limits the student model in learning undiscovered knowledge/features. In this paper, we propose a novel inheritance and exploration knowledge distillation framework (IE-KD), in which a student model is split into two parts - inheritance and exploration. The inheritance part is learned with a similarity loss to transfer the existing learned knowledge from the teacher model to the student model, while the exploration part is encouraged to learn representations different from the inherited ones with a dis-similarity loss. Our IE-KD framework is generic and can be easily combined with existing distillation or mutual learning methods for training deep neural networks. Extensive experiments demonstrate that these two parts can jointly push the student model to learn more diversified and effective representations, and our IE-KD can be a general technique to improve the student network to achieve SOTA performance. Furthermore, by applying our IE-KD to the training of two networks, the performance of both can be improved w.r.t. deep mutual learning.

The goal of gait recognition is to learn the unique spatiotemporal pattern about the human body shape from its temporal changing characteristics. As different body parts behave differently during walking, it is intuitive to model the spatio-temporal patterns of each part separately. However, existing part-based methods equally divide the feature maps of each frame into fixed horizontal stripes to get local parts. It is obvious that these stripe partition-based methods cannot accurately locate the body parts. First, different body parts can appear at the same stripe (e.g., arms and the torso), and one part can appear at different stripes in different frames (e.g., hands). Second, different body parts possess different scales, and even the same part in different frames can appear at different locations and scales. Third, different parts also exhibit distinct movement patterns (e.g., at which frame the movement starts, the position change frequency, how long it lasts). To overcome these issues, we propose novel 3D local operations as a generic family of building blocks for 3D gait recognition backbones. The proposed 3D local operations support the extraction of local 3D volumes of body parts in a sequence with adaptive spatial and temporal scales, locations and lengths. In this way, the spatio-temporal patterns of the body parts are well learned from the 3D local neighborhood in partspecific scales, locations, frequencies and lengths. Experiments demonstrate that our 3D local convolutional neural networks achieve state-of-the-art performance on popular gait datasets. Code is available at: https://github.com/yellowtownhz/3DLocalCNN.

Skeleton-based human action recognition has attracted much attention with the prevalence of accessible depth sensors. Recently, graph convolutional networks (GCNs) have been widely used for this task due to their powerful capability to model graph data. The topology of the adjacency graph is a key factor for modeling the correlations of the input skeletons. Thus, previous methods mainly focus on the design/learning of the graph topology. But once the topology is learned, only a single-scale feature and one transformation exist in each layer of the networks. Many insights, such as multi-scale information and multiple sets of transformations, that have been proven to be very effective in convolutional neural networks (CNNs), have not been investigated in GCNs. The reason is that, due to the gap between graph-structured skeleton data and conventional image/video data, it is very challenging to embed these insights into GCNs. To overcome this gap, we reinvent the split-transform-merge strategy in GCNs for skeleton sequence processing. Specifically, we design a simple and highly modularized graph convolutional network architecture for skeleton-based action recognition. Our network is constructed by repeating a building block that aggregates multi-granularity information from both the spatial and temporal paths. Extensive experiments demonstrate that our network outperforms state-of-the-art methods by a significant margin with only 1/5 of the parameters and 1/10 of the FLOPs.