Yutong Gao

Yutong Gao, Qinglin Meng, Yuan Zhou et al.

While Large Language Models (LLMs) have achieved strong performance across many NLP tasks, their opaque internal mechanisms hinder trustworthiness and safe deployment. Existing surveys in explainable AI largely focus on post-hoc explanation methods that interpret trained models through external approximations. In contrast, intrinsic interpretability, which builds transparency directly into model architectures and computations, has recently emerged as a promising alternative. This paper presents a systematic review of the recent advances in intrinsic interpretability for LLMs, categorizing existing approaches into five design paradigms: functional transparency, concept alignment, representational decomposability, explicit modularization, and latent sparsity induction. We further discuss open challenges and outline future research directions in this emerging field. The paper list is available at: https://github.com/PKU-PILLAR-Group/Survey-Intrinsic-Interpretability-of-LLMs.

Maoyuan Shao, Yutong Gao, Xinyang Huang et al.

Vision-language models like CLIP have achieved remarkable progress in cross-modal representation learning, yet suffer from systematic misclassifications among visually and semantically similar categories. We observe that such confusion patterns are not random but persistently occur between specific category pairs, revealing the model's intrinsic bias and limited fine-grained discriminative ability. To address this, we propose CAPT, a Confusion-Aware Prompt Tuning framework that enables models to learn from their own misalignment. Specifically, we construct a Confusion Bank to explicitly model stable confusion relationships across categories and misclassified samples. On this basis, we introduce a Semantic Confusion Miner (SEM) to capture global inter-class confusion through semantic difference and commonality prompts, and a Sample Confusion Miner (SAM) to retrieve representative misclassified instances from the bank and capture sample-level cues through a Diff-Manner Adapter that integrates global and local contexts. To further unify confusion information across different granularities, a Multi-Granularity Difference Expert (MGDE) module is designed to jointly leverage semantic- and sample-level experts for more robust confusion-aware reasoning. Extensive experiments on 11 benchmark datasets demonstrate that our method significantly reduces confusion-induced errors while enhancing the discriminability and generalization of both base and novel classes, successfully resolving 50.72 percent of confusable sample pairs. Code will be released at https://github.com/greatest-gourmet/CAPT.

Fei Shen, Yutong Gao, Jian Yu et al.

Recent diffusion models have advanced image editing by improving fidelity and controllability across creative and personalized applications. However, multi-object scenes remain challenging, as reliable control over object categories, counts, and spatial layout is difficult to achieve. For that, we first study quantity and layout consistent image editing, abbreviated as QL-Edit, which targets control of object quantity and spatial layout in multi-object scenes. Then, we present IMAGHarmony, a straightforward framework featuring a plug-and-play harmony aware (HA) module that fuses perception semantics while modeling object counts and locations, resulting in accurate edits and strong structural consistency. We further observe that diffusion models are sensitive to the choice of initial noise and tend to prefer certain noise patterns. Based on this finding, we present a preference-guided noise selection (PNS) strategy that selects semantically aligned initial noise through vision and language matching, thereby further improving generation stability and layout consistency in multiple object editing. To support evaluation, we develop HarmonyBench, a comprehensive benchmark that covers a diverse range of quantity and layout control scenarios. Extensive experiments demonstrate that IMAGHarmony outperforms prior methods in both structural alignment and semantic accuracy, utilizing only 200 training images and 10.6M of trainable parameters. Code, models, and data are available at https://github.com/muzishen/IMAGHarmony.

Siyuan Huang, Yutong Gao, Juyang Bai et al.

Creating high-quality figures and visualizations for scientific papers is a time-consuming task that requires both deep domain knowledge and professional design skills. Despite over 2.5 million scientific papers published annually, the figure generation process remains largely manual. We introduce $\textbf{SciFig}$, an end-to-end AI agent system that generates publication-ready pipeline figures directly from research paper texts. SciFig uses a hierarchical layout generation strategy, which parses research descriptions to identify component relationships, groups related elements into functional modules, and generates inter-module connections to establish visual organization. Furthermore, an iterative chain-of-thought (CoT) feedback mechanism progressively improves layouts through multiple rounds of visual analysis and reasoning. We introduce a rubric-based evaluation framework that analyzes 2,219 real scientific figures to extract evaluation rubrics and automatically generates comprehensive evaluation criteria. SciFig demonstrates remarkable performance: achieving 70.1$\%$ overall quality on dataset-level evaluation and 66.2$\%$ on paper-specific evaluation, and consistently high scores across metrics such as visual clarity, structural organization, and scientific accuracy. SciFig figure generation pipeline and our evaluation benchmark will be open-sourced.

Xinye Cao, Hongcan Guo, Jiawen Qian et al.

Understanding hour-long videos with multi-modal large language models (MM-LLMs) enriches the landscape of human-centered AI applications. However, for end-to-end video understanding with LLMs, uniformly sampling video frames results in LLMs being overwhelmed by a vast amount of irrelevant information as video length increases. Existing hierarchical key frame extraction methods improve the accuracy of video understanding but still face two critical challenges. 1) How can the interference of extensive redundant information in long videos be mitigated? 2) How can a model dynamically adapt to complex hierarchical structures while accurately identifying key frames? To address these issues, we propose VideoMiner, which iteratively segments, captions, and clusters long videos, forming a hierarchical tree structure. The proposed VideoMiner progresses from long videos to events to frames while preserving temporal coherence, effectively addressing the first challenge. To precisely locate key frames, we introduce T-GRPO, a tree-based group relative policy optimization in reinforcement learning method that guides the exploration of the VideoMiner. The proposed T-GRPO is specifically designed for tree structures, integrating spatiotemporal information at the event level while being guided by the question, thus solving the second challenge. We achieve superior performance in all long-video understanding tasks and uncover several interesting insights. Our proposed T-GRPO surprisingly incentivizes the model to spontaneously generate a reasoning chain. Additionally, the designed tree growth auxin dynamically adjusts the expansion depth, obtaining accuracy and efficiency gains. The code is publicly available at https://github.com/caoxinye/VideoMiner.

Yutong Gao, Maoyuan Shao, Xinyang Huang et al.

CLIP's success has demonstrated that prompt tuning can achieve robust cross-modal semantic alignment for tasks ranging from open-domain recognition to fine-grained classification. However, redundant or weakly relevant feature components introduce noise and incur unnecessary computational costs. In this work, we propose Spotlighter, a lightweight token-selection framework that simultaneously enhances accuracy and efficiency in prompt tuning. Spotlighter evaluates each visual token's activation from both sample-wise and semantic-wise perspectives and retains only the top-scoring tokens for downstream prediction. A class-specific semantic memory bank of learned prototypes refines this selection, ensuring semantic representativeness and compensating for discarded features. To further prioritize informative signals, we introduce a two-level ranking mechanism that dynamically weights token--prototype interactions. Across 11 few-shot benchmarks, Spotlighter outperforms CLIP by up to 11.19\% in harmonic mean accuracy and achieves up to 0.8K additional FPS, with only 21 extra parameters. These results establish Spotlighter as an effective and scalable baseline for prompt tuning. Code for our method will be available at https://github.com/greatest-gourmet/Spotlighter.

Yutong Gao, Charles A. Ellis, Vince D. Calhoun et al.

The high dimensionality and complexity of neuroimaging data necessitate large datasets to develop robust and high-performing deep learning models. However, the neuroimaging field is notably hampered by the scarcity of such datasets. In this work, we proposed a data augmentation and validation framework that utilizes dynamic forecasting with Long Short-Term Memory (LSTM) networks to enrich datasets. We extended multivariate time series data by predicting the time courses of independent component networks (ICNs) in both one-step and recursive configurations. The effectiveness of these augmented datasets was then compared with the original data using various deep learning models designed for chronological age prediction tasks. The results suggest that our approach improves model performance, providing a robust solution to overcome the challenges presented by the limited size of neuroimaging datasets.

Yutong Gao, Vince D. Calhoun, Robyn L. Miller

Understanding the relationship between cognition and intrinsic brain activity through purely data-driven approaches remains a significant challenge in neuroscience. Resting-state functional magnetic resonance imaging (rs-fMRI) offers a non-invasive method to monitor regional neural activity, providing a rich and complex spatiotemporal data structure. Deep learning has shown promise in capturing these intricate representations. However, the limited availability of large datasets, especially for disease-specific groups such as Alzheimer's Disease (AD), constrains the generalizability of deep learning models. In this study, we focus on multivariate time series forecasting of independent component networks derived from rs-fMRI as a form of data augmentation, using both a conventional LSTM-based model and the novel Transformer-based BrainLM model. We assess their utility in AD classification, demonstrating how generative forecasting enhances classification performance. Post-hoc interpretation of BrainLM reveals class-specific brain network sensitivities associated with AD.

Yutong Gao, Liqian Liang, Congyan Lang et al.

In this work, we focus on Interactive Human Parsing (IHP), which aims to segment a human image into multiple human body parts with guidance from users' interactions. This new task inherits the class-aware property of human parsing, which cannot be well solved by traditional interactive image segmentation approaches that are generally class-agnostic. To tackle this new task, we first exploit user clicks to identify different human parts in the given image. These clicks are subsequently transformed into semantic-aware localization maps, which are concatenated with the RGB image to form the input of the segmentation network and generate the initial parsing result. To enable the network to better perceive user's purpose during the correction process, we investigate several principal ways for the refinement, and reveal that random-sampling-based click augmentation is the best way for promoting the correction effectiveness. Furthermore, we also propose a semantic-perceiving loss (SP-loss) to augment the training, which can effectively exploit the semantic relationships of clicks for better optimization. To the best knowledge, this work is the first attempt to tackle the human parsing task under the interactive setting. Our IHP solution achieves 85\% mIoU on the benchmark LIP, 80\% mIoU on PASCAL-Person-Part and CIHP, 75\% mIoU on Helen with only 1.95, 3.02, 2.84 and 1.09 clicks per class respectively. These results demonstrate that we can simply acquire high-quality human parsing masks with only a few human effort. We hope this work can motivate more researchers to develop data-efficient solutions to IHP in the future.

Yutong Gao, Yi Pan

Machine learning techniques are immensely deployed in both industry and academy. Recent studies indicate that machine learning models used for classification tasks are vulnerable to adversarial examples, which limits the usage of applications in the fields with high precision requirements. We propose a new approach called Feature Map Denoising to detect the adversarial inputs and show the performance of detection on the mixed dataset consisting of adversarial examples generated by different attack algorithms, which can be used to associate with any pre-trained DNNs at a low cost. Wiener filter is also introduced as the denoise algorithm to the defense model, which can further improve performance. Experimental results indicate that good accuracy of detecting the adversarial examples can be achieved through our Feature Map Denoising algorithm.