arXiv subjects

Yizhe Zhang

Publications and source records attributed to Yizhe Zhang.

At least 55 records · Page 3Linked to original sources

arXiv2025

Continuously Augmented Discrete Diffusion model for Categorical Generative Modeling

Standard discrete diffusion models treat all unobserved states identically by mapping them to an absorbing [MASK] token. This creates an 'information void' where semantic information that could be inferred from unmasked tokens is lost between denoising steps. We introduce Continuously Augmented Discrete Diffusion (CADD), a framework that augments the discrete state space with a paired diffusion in a continuous latent space. This yields graded, gradually corrupted states in which masked tokens are represented by noisy yet informative latent vectors rather than collapsed 'information voids'. At each reverse step, CADD may leverage the continuous latent as a semantic hint to guide discrete denoising. The design is clean and compatible with existing discrete diffusion training. At sampling time, the strength and choice of estimator for the continuous latent vector enables a controlled trade-off between mode-coverage (generating diverse outputs) and mode-seeking (generating contextually precise outputs) behaviors. Empirically, we demonstrate CADD improves generative quality over mask-based diffusion across text generation, image synthesis, and code modeling, with consistent gains on both qualitative and quantitative metrics against strong discrete baselines.

stat.ML↗

arXiv2025

Can General-Purpose Omnimodels Compete with Specialists? A Case Study in Medical Image Segmentation

The emergence of powerful, general-purpose omnimodels capable of processing diverse data modalities has raised a critical question: can these ``jack-of-all-trades'' systems perform on par with highly specialized models in knowledge-intensive domains? This work investigates this question within the high-stakes field of medical image segmentation. We conduct a comparative study analyzing the zero-shot performance of a state-of-the-art omnimodel (Gemini, the ``Nano Banana'' model) against domain-specific deep learning models on three distinct tasks: polyp (endoscopy), retinal vessel (fundus), and breast tumor segmentation (ultrasound). Our study focuses on performance at the extremes by curating subsets of the ``easiest'' and ``hardest'' cases based on the specialist models' accuracy. Our findings reveal a nuanced and task-dependent landscape. For polyp and breast tumor segmentation, specialist models excel on easy samples, but the omnimodel demonstrates greater robustness on hard samples where specialists fail catastrophically. Conversely, for the fine-grained task of retinal vessel segmentation, the specialist model maintains superior performance across both easy and hard cases. Intriguingly, qualitative analysis suggests omnimodels may possess higher sensitivity, identifying subtle anatomical features missed by human annotators. Our results indicate that while current omnimodels are not yet a universal replacement for specialists, their unique strengths suggest a potential complementary role with specialist models, particularly in enhancing robustness on challenging edge cases.

eess.IV↗

arXiv2025

Aerial-ground Cross-modal Localization: Dataset, Ground-truth, and Benchmark

Accurate visual localization in dense urban environments poses a fundamental task in photogrammetry, geospatial information science, and robotics. While imagery is a low-cost and widely accessible sensing modality, its effectiveness on visual odometry is often limited by textureless surfaces, severe viewpoint changes, and long-term drift. The growing public availability of airborne laser scanning (ALS) data opens new avenues for scalable and precise visual localization by leveraging ALS as a prior map. However, the potential of ALS-based localization remains underexplored due to three key limitations: (1) the lack of platform-diverse datasets, (2) the absence of reliable ground-truth generation methods applicable to large-scale urban environments, and (3) limited validation of existing Image-to-Point Cloud (I2P) algorithms under aerial-ground cross-platform settings. To overcome these challenges, we introduce a new large-scale dataset that integrates ground-level imagery from mobile mapping systems with ALS point clouds collected in Wuhan, Hong Kong, and San Francisco.

cs.RO↗

arXiv2025

Hierarchical Spatio-temporal Segmentation Network for Ejection Fraction Estimation in Echocardiography Videos

Automated segmentation of the left ventricular endocardium in echocardiography videos is a key research area in cardiology. It aims to provide accurate assessment of cardiac structure and function through Ejection Fraction (EF) estimation. Although existing studies have achieved good segmentation performance, their results do not perform well in EF estimation. In this paper, we propose a Hierarchical Spatio-temporal Segmentation Network (\ourmodel) for echocardiography video, aiming to improve EF estimation accuracy by synergizing local detail modeling with global dynamic perception. The network employs a hierarchical design, with low-level stages using convolutional networks to process single-frame images and preserve details, while high-level stages utilize the Mamba architecture to capture spatio-temporal relationships. The hierarchical design balances single-frame and multi-frame processing, avoiding issues such as local error accumulation when relying solely on single frames or neglecting details when using only multi-frame data. To overcome local spatio-temporal limitations, we propose the Spatio-temporal Cross Scan (STCS) module, which integrates long-range context through skip scanning across frames and positions. This approach helps mitigate EF calculation biases caused by ultrasound image noise and other factors.

cs.CV↗

arXiv2025

A Closer Look at Edema Area Segmentation in SD-OCT Images Using Adversarial Framework

The development of artificial intelligence models for macular edema (ME) analy-sis always relies on expert-annotated pixel-level image datasets which are expen-sive to collect prospectively. While anomaly-detection-based weakly-supervised methods have shown promise in edema area (EA) segmentation task, their per-formance still lags behind fully-supervised approaches. In this paper, we leverage the strong correlation between EA and retinal layers in spectral-domain optical coherence tomography (SD-OCT) images, along with the update characteristics of weakly-supervised learning, to enhance an off-the-shelf adversarial framework for EA segmentation with a novel layer-structure-guided post-processing step and a test-time-adaptation (TTA) strategy. By incorporating additional retinal lay-er information, our framework reframes the dense EA prediction task as one of confirming intersection points between the EA contour and retinal layers, result-ing in predictions that better align with the shape prior of EA. Besides, the TTA framework further helps address discrepancies in the manifestations and presen-tations of EA between training and test sets. Extensive experiments on two pub-licly available datasets demonstrate that these two proposed ingredients can im-prove the accuracy and robustness of EA segmentation, bridging the gap between weakly-supervised and fully-supervised models.

eess.IV↗

arXiv2025

Uncertainty-aware Cross-training for Semi-supervised Medical Image Segmentation

Semi-supervised learning has gained considerable popularity in medical image segmentation tasks due to its capability to reduce reliance on expert-examined annotations. Several mean-teacher (MT) based semi-supervised methods utilize consistency regularization to effectively leverage valuable information from unlabeled data. However, these methods often heavily rely on the student model and overlook the potential impact of cognitive biases within the model. Furthermore, some methods employ co-training using pseudo-labels derived from different inputs, yet generating high-confidence pseudo-labels from perturbed inputs during training remains a significant challenge. In this paper, we propose an Uncertainty-aware Cross-training framework for semi-supervised medical image Segmentation (UC-Seg). Our UC-Seg framework incorporates two distinct subnets to effectively explore and leverage the correlation between them, thereby mitigating cognitive biases within the model. Specifically, we present a Cross-subnet Consistency Preservation (CCP) strategy to enhance feature representation capability and ensure feature consistency across the two subnets. This strategy enables each subnet to correct its own biases and learn shared semantics from both labeled and unlabeled data. Additionally, we propose an Uncertainty-aware Pseudo-label Generation (UPG) component that leverages segmentation results and corresponding uncertainty maps from both subnets to generate high-confidence pseudo-labels. We extensively evaluate the proposed UC-Seg on various medical image segmentation tasks involving different modality images, such as MRI, CT, ultrasound, colonoscopy, and so on. The results demonstrate that our method achieves superior segmentation accuracy and generalization performance compared to other state-of-the-art semi-supervised methods. Our code will be released at https://github.com/taozh2017/UCSeg.

cs.CV↗

arXiv2025

Beyond Manual Annotation: A Human-AI Collaborative Framework for Medical Image Segmentation Using Only "Better or Worse" Expert Feedback

Manual annotation of medical images is a labor-intensive and time-consuming process, posing a significant bottleneck in the development and deployment of robust medical imaging AI systems. This paper introduces a novel hands-free Human-AI collaborative framework for medical image segmentation that substantially reduces the annotation burden by eliminating the need for explicit manual pixel-level labeling. The core innovation lies in a preference learning paradigm, where human experts provide minimal, intuitive feedback -- simply indicating whether an AI-generated segmentation is better or worse than a previous version. The framework comprises four key components: (1) an adaptable foundation model (FM) for feature extraction, (2) label propagation based on feature similarity, (3) a clicking agent that learns from human better-or-worse feedback to decide where to click and with which label, and (4) a multi-round segmentation learning procedure that trains a state-of-the-art segmentation network using pseudo-labels generated by the clicking agent and FM-based label propagation. Experiments on three public datasets demonstrate that the proposed approach achieves competitive segmentation performance using only binary preference feedback, without requiring experts to directly manually annotate the images.

eess.IV↗

arXiv2025

EndoFinder: Online Lesion Retrieval for Explainable Colorectal Polyp Diagnosis Leveraging Latent Scene Representations

Colorectal cancer (CRC) remains a leading cause of cancer-related mortality, underscoring the importance of timely polyp detection and diagnosis. While deep learning models have improved optical-assisted diagnostics, they often demand extensive labeled datasets and yield "black-box" outputs with limited interpretability. In this paper, we propose EndoFinder, an online polyp retrieval framework that leverages multi-view scene representations for explainable and scalable CRC diagnosis. First, we develop a Polyp-aware Image Encoder by combining contrastive learning and a reconstruction task, guided by polyp segmentation masks. This self-supervised approach captures robust features without relying on large-scale annotated data. Next, we treat each polyp as a three-dimensional "scene" and introduce a Scene Representation Transformer, which fuses multiple views of the polyp into a single latent representation. By discretizing this representation through a hashing layer, EndoFinder enables real-time retrieval from a compiled database of historical polyp cases, where diagnostic information serves as interpretable references for new queries. We evaluate EndoFinder on both public and newly collected polyp datasets for re-identification and pathology classification. Results show that EndoFinder outperforms existing methods in accuracy while providing transparent, retrieval-based insights for clinical decision-making. By contributing a novel dataset and a scalable, explainable framework, our work addresses key challenges in polyp diagnosis and offers a promising direction for more efficient AI-driven colonoscopy workflows. The source code is available at https://github.com/ku262/EndoFinder-Scene.

cs.IR↗

arXiv2025

Text-driven Multiplanar Visual Interaction for Semi-supervised Medical Image Segmentation

Semi-supervised medical image segmentation is a crucial technique for alleviating the high cost of data annotation. When labeled data is limited, textual information can provide additional context to enhance visual semantic understanding. However, research exploring the use of textual data to enhance visual semantic embeddings in 3D medical imaging tasks remains scarce. In this paper, we propose a novel text-driven multiplanar visual interaction framework for semi-supervised medical image segmentation (termed Text-SemiSeg), which consists of three main modules: Text-enhanced Multiplanar Representation (TMR), Category-aware Semantic Alignment (CSA), and Dynamic Cognitive Augmentation (DCA). Specifically, TMR facilitates text-visual interaction through planar mapping, thereby enhancing the category awareness of visual features. CSA performs cross-modal semantic alignment between the text features with introduced learnable variables and the intermediate layer of visual features. DCA reduces the distribution discrepancy between labeled and unlabeled data through their interaction, thus improving the model's robustness. Finally, experiments on three public datasets demonstrate that our model effectively enhances visual features with textual information and outperforms other methods. Our code is available at https://github.com/taozh2017/Text-SemiSeg.

cs.CV↗

arXiv2025

Spatial Lifting for Dense Prediction

We present Spatial Lifting (SL), a novel methodology for dense prediction tasks. SL operates by lifting standard inputs, such as 2D images, into a higher-dimensional space and subsequently processing them using networks designed for that higher dimension, such as a 3D U-Net. Counterintuitively, this dimensionality lifting allows us to achieve good performance on benchmark tasks compared to conventional approaches, while reducing inference costs and significantly lowering the number of model parameters. The SL framework produces intrinsically structured outputs along the lifted dimension. This emergent structure facilitates dense supervision during training and enables robust, near-zero-additional-cost prediction quality assessment at test time. We validate our approach across 19 benchmark datasets (13 for semantic segmentation and 6 for depth estimation), demonstrating competitive dense prediction performance while reducing the model parameter count by over 98% (in the U-Net case) and lowering inference costs. Spatial Lifting introduces a new vision modeling paradigm that offers a promising path toward more efficient, accurate, and reliable deep networks for dense prediction tasks in vision.

cs.CV↗

arXiv2025

SAGE: Steering Dialog Generation with Future-Aware State-Action Augmentation

Recent advances in large language models have demonstrated impressive capabilities in task-oriented applications, yet building emotionally intelligent chatbots that can engage in natural, strategic conversations remains a challenge. We present a novel approach called SAGE that uses latent variables to control long-horizon behavior in dialogue generation. At the core of our method is the State-Action Chain (SAC), which augments standard language model fine-tuning by introducing latent variables that encapsulate emotional states and conversational strategies between dialogue turns. During inference, these variables are generated before each response, enabling coarse-grained control over dialogue progression while maintaining natural interaction patterns. We also introduce a self-improvement pipeline that leverages dialogue tree search, LLM-based reward modeling, and targeted fine-tuning to optimize conversational trajectories. Our experimental results show that models trained with this approach demonstrate improved performance in emotional intelligence metrics while maintaining strong capabilities on LLM benchmarks. The discrete nature of our latent variables facilitates search-based strategies and provides a foundation for future applications of reinforcement learning to dialogue systems, where learning can occur at the state level rather than the token level. https://github.com/apple/ml-sage-dialog-gen

cs.CL↗

arXiv2025

Flexible Language Modeling in Continuous Space with Transformer-based Autoregressive Flows

Autoregressive models have driven remarkable progress in language modeling. Their foundational reliance on discrete tokens, unidirectional context, and single-pass decoding, while central to their success, also inspires the exploration of a design space that could offer new axes of modeling flexibility. In this work, we explore an alternative paradigm, shifting language modeling from a discrete token space to a continuous latent space. We propose a novel framework TarFlowLM, that employs transformer-based autoregressive normalizing flows to model these continuous representations. This approach unlocks substantial flexibility, enabling the construction of models that can capture global bi-directional context through stacked, alternating-direction autoregressive transformations, support block-wise generation with flexible token patch sizes, and facilitate a hierarchical multi-pass generation process. We further propose new mixture-based coupling transformations designed to capture complex dependencies within the latent space shaped by discrete data, and demonstrate theoretical connections to conventional discrete autoregressive models. Extensive experiments on language modeling benchmarks demonstrate strong likelihood performance and highlight the flexible modeling capabilities inherent in our framework.

cs.LG↗

arXiv2025

What Makes the Preferred Thinking Direction for LLMs in Multiple-choice Questions?

Language models usually use left-to-right (L2R) autoregressive factorization. However, L2R factorization may not always be the best inductive bias. Therefore, we investigate whether alternative factorizations of the text distribution could be beneficial in some tasks. We investigate right-to-left (R2L) training as a compelling alternative, focusing on multiple-choice questions (MCQs) as a test bed for knowledge extraction and reasoning. Through extensive experiments across various model sizes (2B-8B parameters) and training datasets, we find that R2L models can significantly outperform L2R models on several MCQ benchmarks, including logical reasoning, commonsense understanding, and truthfulness assessment tasks. Our analysis reveals that this performance difference may be fundamentally linked to multiple factors including calibration, computability, and directional conditional entropy. We analyze the impact of these factors through controlled simulation studies using arithmetic tasks, where the impacting factors can be better disentangled. Our work demonstrates that exploring alternative factorizations of the text distribution can lead to improvements in LLM capabilities and provides theoretical insights into optimal factorization towards approximating human language distribution, and when each reasoning order might be more advantageous. Our code and checkpoints are released at https://github.com/apple/ml-reversal-blessing.

cs.CL↗

arXiv2025

DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation

Diffusion large language models (dLLMs) are compelling alternatives to autoregressive (AR) models because their denoising models operate over the entire sequence. The global planning and iterative refinement features of dLLMs are particularly useful for code generation. However, current training and inference mechanisms for dLLMs in coding are still under-explored. To demystify the decoding behavior of dLLMs and unlock their potential for coding, we systematically investigate their denoising processes and reinforcement learning (RL) methods. We train a 7B dLLM, \textbf{DiffuCoder}, on 130B tokens of code. Using this model as a testbed, we analyze its decoding behavior, revealing how it differs from that of AR models: (1) dLLMs can decide how causal their generation should be without relying on semi-AR decoding, and (2) increasing the sampling temperature diversifies not only token choices but also their generation order. This diversity creates a rich search space for RL rollouts. For RL training, to reduce the variance of token log-likelihood estimates and maintain training efficiency, we propose \textbf{coupled-GRPO}, a novel sampling scheme that constructs complementary mask noise for completions used in training. In our experiments, coupled-GRPO significantly improves DiffuCoder's performance on code generation benchmarks (+4.4\% on EvalPlus) and reduces reliance on AR bias during decoding. Our work provides deeper insight into the machinery of dLLM generation and offers an effective, diffusion-native RL training framework. https://github.com/apple/ml-diffucoder.

cs.CL↗

arXiv2025

Eliciting In-context Retrieval and Reasoning for Long-context Large Language Models

Recent advancements in long-context language models (LCLMs) promise to transform Retrieval-Augmented Generation (RAG) by simplifying pipelines. With their expanded context windows, LCLMs can process entire knowledge bases and perform retrieval and reasoning directly -- a capability we define as In-Context Retrieval and Reasoning (ICR^2). However, existing benchmarks like LOFT often overestimate LCLM performance by providing overly simplified contexts. To address this, we introduce ICR^2, a benchmark that evaluates LCLMs in more realistic scenarios by including confounding passages retrieved with strong retrievers. We then propose three methods to enhance LCLM performance: (1) retrieve-then-generate fine-tuning, (2) retrieval-attention-probing, which uses attention heads to filter and de-noise long contexts during decoding, and (3) joint retrieval head training alongside the generation head. Our evaluation of five well-known LCLMs on LOFT and ICR^2 demonstrates significant gains with our best approach applied to Mistral-7B: +17 and +15 points by Exact Match on LOFT, and +13 and +2 points on ICR^2, compared to vanilla RAG and supervised fine-tuning, respectively. It even outperforms GPT-4-Turbo on most tasks despite being a much smaller model.

cs.CL↗

arXiv2025

Training Software Engineering Agents and Verifiers with SWE-Gym

We present SWE-Gym, the first environment for training real-world software engineering (SWE) agents. SWE-Gym contains 2,438 real-world Python task instances, each comprising a codebase with an executable runtime environment, unit tests, and a task specified in natural language. We use SWE-Gym to train language model based SWE agents, achieving up to 19% absolute gains in resolve rate on the popular SWE-Bench Verified and Lite test sets. We also experiment with inference-time scaling through verifiers trained on agent trajectories sampled from SWE-Gym. When combined with our fine-tuned SWE agents, we achieve 32.0% and 26.0% on SWE-Bench Verified and Lite, respectively, reflecting a new state-of-the-art for open-weight SWE agents. To facilitate further research, we publicly release SWE-Gym, models, and agent trajectories.

cs.SE↗

arXiv2025

Scaling Diffusion Language Models via Adaptation from Autoregressive Models

Diffusion Language Models (DLMs) have emerged as a promising new paradigm for text generative modeling, potentially addressing limitations of autoregressive (AR) models. However, current DLMs have been studied at a smaller scale compared to their AR counterparts and lack fair comparison on language modeling benchmarks. Additionally, training diffusion models from scratch at scale remains challenging. Given the prevalence of open-source AR language models, we propose adapting these models to build text diffusion models. We demonstrate connections between AR and diffusion modeling objectives and introduce a simple continual pre-training approach for training diffusion models. Through systematic evaluation on language modeling, reasoning, and commonsense benchmarks, we show that we can convert AR models ranging from 127M to 7B parameters (GPT2 and LLaMA) into diffusion models DiffuGPT and DiffuLLaMA, using less than 200B tokens for training. Our experimental results reveal that these models outperform earlier DLMs and are competitive with their AR counterparts. We release a suite of DLMs (127M-355M-7B) capable of generating fluent text, performing in-context learning, filling in the middle without prompt re-ordering, and following instructions https://github.com/HKUNLP/DiffuLLaMA.

cs.CL↗

arXiv2025

Learn to Reason Efficiently with Adaptive Length-based Reward Shaping

Large Reasoning Models (LRMs) have shown remarkable capabilities in solving complex problems through reinforcement learning (RL), particularly by generating long reasoning traces. However, these extended outputs often exhibit substantial redundancy, which limits the efficiency of LRMs. In this paper, we investigate RL-based approaches to promote reasoning efficiency. Specifically, we first present a unified framework that formulates various efficient reasoning methods through the lens of length-based reward shaping. Building on this perspective, we propose a novel Length-bAsed StEp Reward shaping method (LASER), which employs a step function as the reward, controlled by a target length. LASER surpasses previous methods, achieving a superior Pareto-optimal balance between performance and efficiency. Next, we further extend LASER based on two key intuitions: (1) The reasoning behavior of the model evolves during training, necessitating reward specifications that are also adaptive and dynamic; (2) Rather than uniformly encouraging shorter or longer chains of thought (CoT), we posit that length-based reward shaping should be difficulty-aware i.e., it should penalize lengthy CoTs more for easy queries. This approach is expected to facilitate a combination of fast and slow thinking, leading to a better overall tradeoff. The resulting method is termed LASER-D (Dynamic and Difficulty-aware). Experiments on DeepSeek-R1-Distill-Qwen-1.5B, DeepSeek-R1-Distill-Qwen-7B, and DeepSeek-R1-Distill-Qwen-32B show that our approach significantly enhances both reasoning performance and response length efficiency. For instance, LASER-D and its variant achieve a +6.1 improvement on AIME2024 while reducing token usage by 63%. Further analysis reveals our RL-based compression produces more concise reasoning patterns with less redundant "self-reflections". Resources are at https://github.com/hkust-nlp/Laser.

cs.CL↗