arXiv subjects

Kehong Yuan

Publications and source records attributed to Kehong Yuan.

At least 19 recordsLinked to original sources

arXiv2026

Not Every Divergence Should Be Suppressed: Counterfactual Recoverability in On-Policy Distillation

On-policy distillation (OPD) supervises student-visited trajectories, yet divergence-based rules cannot determine whether an erroneous prefix remains correctable. We formulate this decision as counterfactual recoverability and replay each error state through budget-matched teacher-continuation and rollback branches. Based on their relative success, states are categorized as recoverable, irreversible-but-avoidable, or ambiguous, and these labels guide whether training retains, rolls back, or conventionally supervises the corresponding trajectory. On AIME branch diagnostics, the mean continuation-minus-rollback effect is 0.185 for recoverable states and -1.000 for irreversible-but-avoidable states, demonstrating opposite intervention preferences. A branch-derived recoverability proxy achieves an AUC of 1.000, substantially outperforming divergence alone at 0.392. Across frozen evaluations, recoverability-aware control achieves the strongest recorded performance, reaching 0.578 success on held-out AIME2025 compared with 0.517 for the best baseline. It also improves AIME2024-2025 average@32 from 0.2656 to 0.3125 and GPQA-Diamond average@32 from 0.2702 to 0.3070. Component ablations further show that retaining teacher-correctable prefixes provides the largest individual contribution. These findings establish recoverability as an outcome-grounded decision variable for selective supervision in OPD.

cs.LG↗

arXiv2026

CARGO-VL: Counterfactual Arbitration with Risk-Constrained Group Optimization for Vision-Language Models

Vision-language systems combine images with retrieved text, but these sources can disagree or jointly fail to support an answer. Reliable models must identify the trustworthy source and abstain when neither is adequate. Existing post-training objectives score instances independently and therefore do not enforce coherent behavior under counterfactual evidence changes. We introduce CARGO-VL, a group-relative framework that optimizes matched variants covering aligned, image-correct, text-correct, and both-wrong (A/V/T/N) evidence states as one bundle. Its objective couples condition-wise correctness with transition rewards for answer invariance, source equivariance, and answer-to-abstention switching, while a primal-dual controller balances unsafe answers against excessive deferral. We also contribute XMC (eXtended Modal Conflict), a four-condition conflict training resource, and evaluate transfer on CMC-Bench and Modality-Bias. Across multiple seeds, CARGO-VL improves conflict handling, unsupported-answer avoidance, and modality balance over pointwise baselines. Ablations identify complementary benefits from relational transition signals and adaptive risk control, supporting counterfactual consistency as a practical objective for reliable multimodal evidence arbitration.

cs.AI↗

arXiv2026

SAM-R1: Leveraging SAM for Reward Feedback in Multimodal Segmentation via Reinforcement Learning

Leveraging multimodal large models for image segmentation has become a prominent research direction. However, existing approaches typically rely heavily on manually annotated datasets that include explicit reasoning processes, which are costly and time-consuming to produce. Recent advances suggest that reinforcement learning (RL) can endow large models with reasoning capabilities without requiring such reasoning-annotated data. In this paper, we propose SAM-R1, a novel framework that enables multimodal large models to perform fine-grained reasoning in image understanding tasks. Our approach is the first to incorporate fine-grained segmentation settings during the training of multimodal reasoning models. By integrating task-specific, fine-grained rewards with a tailored optimization objective, we further enhance the model's reasoning and segmentation alignment. We also leverage the Segment Anything Model (SAM) as a strong and flexible reward provider to guide the learning process. With only 3k training samples, SAM-R1 achieves strong performance across multiple benchmarks, demonstrating the effectiveness of reinforcement learning in equipping multimodal models with segmentation-oriented reasoning capabilities.

cs.CV↗

arXiv2026

A Cognitive Explainer for Fetal ultrasound images classifier Based on Medical Concepts

Fetal standard scan plane detection during 2-D mid-pregnancy examinations is a highly complex task, which requires extensive medical knowledge and years of training. Although deep neural networks (DNN) can assist inexperienced operators in these tasks, their lack of transparency and interpretability limit their application. Despite some researchers have been committed to visualizing the decision process of DNN, most of them only focus on the pixel-level features and do not take into account the medical prior knowledge. In this work, we propose an interpretable framework based on key medical concepts, which provides explanations from the perspective of clinicians' cognition. Moreover, we utilize a concept-based graph convolutional neural(GCN) network to construct the relationships between key medical concepts. Extensive experimental analysis on a private dataset has shown that the proposed method provides easy-to-understand insights about reasoning results for clinicians.

cs.LG↗

arXiv2026

Unveiling Super Experts in Mixture-of-Experts Large Language Models

In this study, we report, for the first time, the discovery and systematic investigation of a distinct subset of experts that play a pivotal role in the MoE LLMs' forward inference. These experts are prevalent in open-source MoE LLMs, and despite their extremely limited number, pruning them results in a substantial decline in model performance (e.g., prune just three out of 6,144 causes Qwen3-30B-A3B to generate repetitive and uninformative outputs).We refer to these experts as Super Experts (SEs). Our comprehensive analysis provides progressively deeper insights into SEs: (i) SEs are characterized by rare but extreme activation outliers in the output of the down_proj, which give rise to massive activations in the hidden states between decoder layers. Moreover, the distribution of SEs is model-specific, data-agnostic, and remains unaffected by post-training processes. (ii) By pruning SEs, we assess their significance across a variety of tasks, revealing their considerable impact on the model's overall performance, particularly in mathematical reasoning. (iii) We further investigate why compressing SEs exerts such a pronounced impact. We show that, in MoE LLMs, SEs serve as the primary source of the systematic outlier mechanism in Transformers, and that compressing them profoundly disrupts this process, ultimately causing the collapse of attention sinks. These findings advance the understanding of the internal dynamics of MoE LLMs, filling an important gap in the current knowledge. The code is provided in https://github.com/ZunhaiSu/Super-Experts-Profilling.

cs.CL↗

arXiv2025

KVSink: Understanding and Enhancing the Preservation of Attention Sinks in KV Cache Quantization for LLMs

Key-Value (KV) cache quantization has become a widely adopted optimization technique for efficient large language models (LLMs) inference by reducing KV cache memory usage and mitigating memory-bound constraints. Recent studies have emphasized the importance of preserving the original precision of KVs for the first few tokens to ensure the protection of attention sinks. While this approach has proven effective in mitigating performance degradation, its underlying principles remain insufficiently understood. Moreover, it fails to address the recent discovery that attention sinks can emerge beyond the initial token positions. In this work, we elucidate the underlying mechanisms of attention sinks during inference by examining their role in the cross-layer evolution of extreme activation outliers. Additionally, we provide a comprehensive analysis of the interplay between attention sinks and KV cache quantization. Based on our enhanced understanding, we introduce \textit{\textbf{KVSink}}, a plug-and-play method that effectively predicts sink tokens with negligible overhead, enabling more thorough preservation. Extensive experiments demonstrate that KVSink outperforms the existing Preserve-First-N (PFN) strategy, offering more effective preservation of attention sinks during KV cache quantization. Moreover, when applied to the well-established KVQuant method, KVSink further improves perplexity (PPL) and reduces reliance on 16-bit numerical outliers.

cs.CL↗

arXiv2025

RotateKV: Accurate and Robust 2-Bit KV Cache Quantization for LLMs via Outlier-Aware Adaptive Rotations

Key-Value (KV) cache facilitates efficient large language models (LLMs) inference by avoiding recomputation of past KVs. As the batch size and context length increase, the oversized KV caches become a significant memory bottleneck, highlighting the need for efficient compression. Existing KV quantization rely on fine-grained quantization or the retention of a significant portion of high bit-widths caches, both of which compromise compression ratio and often fail to maintain robustness at extremely low average bit-widths. In this work, we explore the potential of rotation technique for 2-bit KV quantization and propose RotateKV, which achieves accurate and robust performance through the following innovations: (i) Outlier-Aware Rotation, which utilizes channel-reordering to adapt the rotations to varying channel-wise outlier distributions without sacrificing the computational efficiency of the fast Walsh-Hadamard transform (FWHT); (ii) Pre-RoPE Grouped-Head Rotation, which mitigates the impact of rotary position embedding (RoPE) on proposed outlier-aware rotation and further smooths outliers across heads; (iii) Attention-Sink-Aware Quantization, which leverages the massive activations to precisely identify and protect attention sinks. RotateKV achieves less than 0.3 perplexity (PPL) degradation with 2-bit quantization on WikiText-2 using LLaMA-2-13B, maintains strong CoT reasoning and long-context capabilities, with less than 1.7\% degradation on GSM8K, outperforming existing methods even at lower average bit-widths. RotateKV also showcases a 3.97x reduction in peak memory usage, supports 5.75x larger batch sizes, and achieves a 2.32x speedup in decoding stage.

cs.LG↗

arXiv2025

AKVQ-VL: Attention-Aware KV Cache Adaptive 2-Bit Quantization for Vision-Language Models

Vision-language models (VLMs) show remarkable performance in multimodal tasks. However, excessively long multimodal inputs lead to oversized Key-Value (KV) caches, resulting in significant memory consumption and I/O bottlenecks. Previous KV quantization methods for Large Language Models (LLMs) may alleviate these issues but overlook the attention saliency differences of multimodal tokens, resulting in suboptimal performance. In this paper, we investigate the attention-aware token saliency patterns in VLM and propose AKVQ-VL. AKVQ-VL leverages the proposed Text-Salient Attention (TSA) and Pivot-Token-Salient Attention (PSA) patterns to adaptively allocate bit budgets. Moreover, achieving extremely low-bit quantization requires effectively addressing outliers in KV tensors. AKVQ-VL utilizes the Walsh-Hadamard transform (WHT) to construct outlier-free KV caches, thereby reducing quantization difficulty. Evaluations of 2-bit quantization on 12 long-context and multimodal tasks demonstrate that AKVQ-VL maintains or even improves accuracy, outperforming LLM-oriented methods. AKVQ-VL can reduce peak memory usage by 2.13x, support up to 3.25x larger batch sizes and 2.46x throughput.

cs.CL↗

arXiv2025

Densely Connected Parameter-Efficient Tuning for Referring Image Segmentation

In the domain of computer vision, Parameter-Efficient Tuning (PET) is increasingly replacing the traditional paradigm of pre-training followed by full fine-tuning. PET is particularly favored for its effectiveness in large foundation models, as it streamlines transfer learning costs and optimizes hardware utilization. However, the current PET methods are mainly designed for single-modal optimization. While some pioneering studies have undertaken preliminary explorations, they still remain at the level of aligned encoders (e.g., CLIP) and lack exploration of misaligned encoders. These methods show sub-optimal performance with misaligned encoders, as they fail to effectively align the multimodal features during fine-tuning. In this paper, we introduce DETRIS, a parameter-efficient tuning framework designed to enhance low-rank visual feature propagation by establishing dense interconnections between each layer and all preceding layers, which enables effective cross-modal feature interaction and adaptation to misaligned encoders. We also suggest using text adapters to improve textual features. Our simple yet efficient approach greatly surpasses state-of-the-art methods with 0.9% to 1.8% backbone parameter updates, evaluated on challenging benchmarks. Our project is available at \url{https://github.com/jiaqihuang01/DETRIS}.

cs.CV↗

arXiv2024

Sharingan: Extract User Action Sequence from Desktop Recordings

Video recordings of user activities, particularly desktop recordings, offer a rich source of data for understanding user behaviors and automating processes. However, despite advancements in Vision-Language Models (VLMs) and their increasing use in video analysis, extracting user actions from desktop recordings remains an underexplored area. This paper addresses this gap by proposing two novel VLM-based methods for user action extraction: the Direct Frame-Based Approach (DF), which inputs sampled frames directly into VLMs, and the Differential Frame-Based Approach (DiffF), which incorporates explicit frame differences detected via computer vision techniques. We evaluate these methods using a basic self-curated dataset and an advanced benchmark adapted from prior work. Our results show that the DF approach achieves an accuracy of 70% to 80% in identifying user actions, with the extracted action sequences being re-playable though Robotic Process Automation. We find that while VLMs show potential, incorporating explicit UI changes can degrade performance, making the DF approach more reliable. This work represents the first application of VLMs for extracting user action sequences from desktop recordings, contributing new methods, benchmarks, and insights for future research.

cs.CV↗

arXiv2024

Diagnosis Assistant for Liver Cancer Utilizing a Large Language Model with Three Types of Knowledge

Liver cancer has a high incidence rate, but primary healthcare settings often lack experienced doctors. Advances in large models and AI technologies offer potential assistance. This work aims to address limitations in liver cancer diagnosis models, such as poor understanding of medical images, insufficient consideration of liver blood vessels, and ensuring accurate medical information. We propose a specialized diagnostic assistant to improve the diagnostic capabilities of less experienced doctors. Our framework combines large and small models, using optimized small models for precise patient image perception. Specifically, a segmentation network iteratively removes ambiguous pixels for liver tumor segmentation, and a multi-scale, multi-level differential network segments liver vessels. Features from these segmentations and medical records form a patient's personalized knowledge base. For diagnosis, Chain of Thought (COT) technology designs prompts mimicking experienced doctors' thought patterns, and Retrieval-Augmented Generation (RAG) technology provides answers based on reliable domain knowledge and trusted cases. Our small model methods improve liver tumor and vessel segmentation performance, resulting in more accurate information extraction. The large model component scores over 1 point higher on a 10-point scale in evaluations by doctors compared to control methods. Our method enhances semantic perception of medical images, improves classification of ambiguous pixels, and optimizes small object perception. It considers blood vessel positions for specific treatments and improves response credibility and interpretability by mimicking experienced doctors' thought processes using reliable resources. This approach has been recognized by doctors and benefits liver cancer auxiliary diagnosis.

physics.med-ph↗

arXiv2024

V3Det Challenge 2024 on Vast Vocabulary and Open Vocabulary Object Detection: Methods and Results

Detecting objects in real-world scenes is a complex task due to various challenges, including the vast range of object categories, and potential encounters with previously unknown or unseen objects. The challenges necessitate the development of public benchmarks and challenges to advance the field of object detection. Inspired by the success of previous COCO and LVIS Challenges, we organize the V3Det Challenge 2024 in conjunction with the 4th Open World Vision Workshop: Visual Perception via Learning in an Open World (VPLOW) at CVPR 2024, Seattle, US. This challenge aims to push the boundaries of object detection research and encourage innovation in this field. The V3Det Challenge 2024 consists of two tracks: 1) Vast Vocabulary Object Detection: This track focuses on detecting objects from a large set of 13204 categories, testing the detection algorithm's ability to recognize and locate diverse objects. 2) Open Vocabulary Object Detection: This track goes a step further, requiring algorithms to detect objects from an open set of categories, including unknown objects. In the following sections, we will provide a comprehensive summary and analysis of the solutions submitted by participants. By analyzing the methods and solutions presented, we aim to inspire future research directions in vast vocabulary and open-vocabulary object detection, driving progress in this field. Challenge homepage: https://v3det.openxlab.org.cn/challenge

cs.CV↗

arXiv2024

Dual-Channel Reliable Breast Ultrasound Image Classification Based on Explainable Attribution and Uncertainty Quantification

This paper focuses on the classification task of breast ultrasound images and researches on the reliability measurement of classification results. We proposed a dual-channel evaluation framework based on the proposed inference reliability and predictive reliability scores. For the inference reliability evaluation, human-aligned and doctor-agreed inference rationales based on the improved feature attribution algorithm SP-RISA are gracefully applied. Uncertainty quantification is used to evaluate the predictive reliability via the Test Time Enhancement. The effectiveness of this reliability evaluation framework has been verified on our breast ultrasound clinical dataset YBUS, and its robustness is verified on the public dataset BUSI. The expected calibration errors on both datasets are significantly lower than traditional evaluation methods, which proves the effectiveness of our proposed reliability measurement.

eess.IV↗

arXiv2022

BSM loss: A superior way in modeling aleatory uncertainty of fine_grained classification

Artificial intelligence(AI)-assisted method had received much attention in the risk field such as disease diagnosis. Different from the classification of disease types, it is a fine-grained task to classify the medical images as benign or malignant. However, most research only focuses on improving the diagnostic accuracy and ignores the evaluation of model reliability, which limits its clinical application. For clinical practice, calibration presents major challenges in the low-data regime extremely for over-parametrized models and inherent noises. In particular, we discovered that modeling data-dependent uncertainty is more conducive to confidence calibrations. Compared with test-time augmentation(TTA), we proposed a modified Bootstrapping loss(BS loss) function with Mixup data augmentation strategy that can better calibrate predictive uncertainty and capture data distribution transformation without additional inference time. Our experiments indicated that BS loss with Mixup(BSM) model can halve the Expected Calibration Error(ECE) compared to standard data augmentation, deep ensemble and MC dropout. The correlation between uncertainty and similarity of in-domain data is up to -0.4428 under the BSM model. Additionally, the BSM model is able to perceive the semantic distance of out-of-domain data, demonstrating high potential in real-world clinical practice.

cs.CV↗

arXiv2022

AI assisted method for efficiently generating breast ultrasound screening reports

Background: Ultrasound is one of the preferred choices for early screening of dense breast cancer. Clinically, doctors have to manually write the screening report which is time-consuming and laborious, and it is easy to miss and miswrite. Aim: We proposed a new pipeline to automatically generate AI breast ultrasound screening reports based on ultrasound images, aiming to assist doctors in improving the efficiency of clinical screening and reducing repetitive report writing. Methods: AI was used to efficiently generate personalized breast ultrasound screening preliminary reports, especially for benign and normal cases which account for the majority. Based on the preliminary AI report, doctors then make simple adjustments or corrections to quickly generate the final report. The approach has been trained and tested using a database of 4809 breast tumor instances. Results: Experimental results indicate that this pipeline improves doctors' work efficiency by up to 90%, which greatly reduces repetitive work. Conclusion: Personalized report generation is more widely recognized by doctors in clinical practice compared with non-intelligent reports based on fixed templates or containing options to fill in the blanks.

eess.IV↗

arXiv2022

MIPR:Automatic Annotation of Medical Images with Pixel Rearrangement

Most of the state-of-the-art semantic segmentation reported in recent years is based on fully supervised deep learning in the medical domain. How?ever, the high-quality annotated datasets require intense labor and domain knowledge, consuming enormous time and cost. Previous works that adopt semi?supervised and unsupervised learning are proposed to address the lack of anno?tated data through assisted training with unlabeled data and achieve good perfor?mance. Still, these methods can not directly get the image annotation as doctors do. In this paper, inspired by self-training of semi-supervised learning, we pro?pose a novel approach to solve the lack of annotated data from another angle, called medical image pixel rearrangement (short in MIPR). The MIPR combines image-editing and pseudo-label technology to obtain labeled data. As the number of iterations increases, the edited image is similar to the original image, and the labeled result is similar to the doctor annotation. Therefore, the MIPR is to get labeled pairs of data directly from amounts of unlabled data with pixel rearrange?ment, which is implemented with a designed conditional Generative Adversarial Networks and a segmentation network. Experiments on the ISIC18 show that the effect of the data annotated by our method for segmentation task is is equal to or even better than that of doctors annotations

eess.IV↗

arXiv2022

Soft-CP: A Credible and Effective Data Augmentation for Semantic Segmentation of Medical Lesions

The medical datasets are usually faced with the problem of scarcity and data imbalance. Moreover, annotating large datasets for semantic segmentation of medical lesions is domain-knowledge and time-consuming. In this paper, we propose a new object-blend method(short in soft-CP) that combines the Copy-Paste augmentation method for semantic segmentation of medical lesions offline, ensuring the correct edge information around the lession to solve the issue above-mentioned. We proved the method's validity with several datasets in different imaging modalities. In our experiments on the KiTS19[2] dataset, Soft-CP outperforms existing medical lesions synthesis approaches. The Soft-CP augementation provides gains of +26.5% DSC in the low data regime(10% of data) and +10.2% DSC in the high data regime(all of data), In offline training data, the ratio of real images to synthetic images is 3:1.

eess.IV↗

arXiv2021