arXiv subjects

Jun Zhou

Publications and source records attributed to Jun Zhou.

At least 163 records · Page 9Linked to original sources

arXiv2025

Optimizing Transmission FLASH Radiotherapy for Large-Field Post-Mastectomy Breast Treatment

We investigated the effects of scanning speed, beam configuration, and dose-rate modeling on the FLASH effect in post-mastectomy proton transmission-beam (TB) planning and evaluated whether optimizing the spot-scanning path can enhance FLASH. Five left-sided post-mastectomy patients (32 Gy in 5 fractions) were replanned with single-energy (249 MeV) tangential TBs plus a clinical en face background beam. FLASH was evaluated with two models: Krieger's FLASH effectiveness model (FEM) and Folkerts' average dose-rate (ADR) framework. Plans used conventional pencil-beam scanning, split-field delivery, and GA-optimized spot sequences, with vertical scan speeds varied from 10 to 20 mm/ms. FLASH in normal tissues was defined as the percentage of voxels meeting the threshold (>= 4 Gy at >= 40 Gy/s); once a voxel met the criterion, a dose-adjustment factor of 0.67 was applied. The FLASH effect was highly sensitive to scanning pattern and model choice. Increasing vertical scan speed from 10 to 20 mm/ms increased FLASH in the CTV by 22% (ADR) and 12% (FEM); in skin it rose from 41.4% to 58.8% (ADR) and from 8.4% to 13.1% (FEM). Split-field delivery increased the temporal separation between vertical spot columns and yielded superior FLASH, including up to a 9.2 Gy reduction in CTV Dmean with ADR. GA-based optimization shortened scan time and achieved FLASH comparable to split-field delivery, with a CTV Dmean reduction of 7.87 Gy (ADR-GA) and skin Dmean reductions of 2-3 Gy. These findings indicate that FLASH outcomes depend strongly on scanning trajectory, scan speed, and model selection. In addition, path-minimizing spot-delivery optimization (e.g., GA) can further improve dose-rate distributions in healthy voxels.

physics.med-ph↗

arXiv2025

Large Language Diffusion Models

The capabilities of large language models (LLMs) are widely regarded as relying on autoregressive models (ARMs). We challenge this notion by introducing LLaDA, a diffusion model trained from scratch under the pre-training and supervised fine-tuning (SFT) paradigm. LLaDA employs a forward data masking process and a reverse generation process, parameterized by a Transformer to predict masked tokens. It provides a principled generative approach for probabilistic inference by optimizing a likelihood lower bound. Across extensive benchmarks on general tasks, math, code, and so on, LLaDA demonstrates strong scalability and performs comparably to our self-constructed ARM baselines. Remarkably, LLaDA 8B is competitive with strong LLMs like LLaMA3 8B in in-context learning and, after SFT, exhibits impressive instruction-following abilities in case studies such as multi-turn dialogue. Moreover, LLaDA addresses the reversal curse, surpassing GPT-4o in a reversal poem completion task. Our findings show the promise of diffusion models for language modeling at scale and challenge the common assumption that core LLM capabilities discussed above inherently depend on ARMs. Project page and codes: https://ml-gsai.github.io/LLaDA-demo/.

cs.CL↗

arXiv2025

LLaDA 1.5: Variance-Reduced Preference Optimization for Large Language Diffusion Models

While Masked Diffusion Models (MDMs), such as LLaDA, present a promising paradigm for language modeling, there has been relatively little effort in aligning these models with human preferences via reinforcement learning. The challenge primarily arises from the high variance in Evidence Lower Bound (ELBO)-based likelihood estimates required for preference optimization. To address this issue, we propose Variance-Reduced Preference Optimization (VRPO), a framework that formally analyzes the variance of ELBO estimators and derives bounds on both the bias and variance of preference optimization gradients. Building on this theoretical foundation, we introduce unbiased variance reduction strategies, including optimal Monte Carlo budget allocation and antithetic sampling, that significantly improve the performance of MDM alignment. We demonstrate the effectiveness of VRPO by applying it to LLaDA, and the resulting model, LLaDA 1.5, outperforms its SFT-only predecessor consistently and significantly across mathematical (GSM8K +4.7), code (HumanEval +3.0, MBPP +1.8), and alignment benchmarks (IFEval +4.0, Arena-Hard +4.3). Furthermore, LLaDA 1.5 demonstrates a highly competitive mathematical performance compared to strong language MDMs and ARMs. Project page: https://ml-gsai.github.io/LLaDA-1.5-Demo/.

cs.LG↗

arXiv2025

From Entity Reliability to Clean Feedback: An Entity-Aware Denoising Framework Beyond Interaction-Level Signals

Implicit feedback is central to modern recommender systems but is inherently noisy, often impairing model training and degrading user experience. At scale, such noise can mislead learning processes, reducing both recommendation accuracy and platform value. Existing denoising strategies typically overlook the entity-specific nature of noise while introducing high computational costs and complex hyperparameter tuning. To address these challenges, we propose \textbf{EARD} (\textbf{E}ntity-\textbf{A}ware \textbf{R}eliability-\textbf{D}riven Denoising), a lightweight framework that shifts the focus from interaction-level signals to entity-level reliability. Motivated by the empirical observation that training loss correlates with noise, EARD quantifies user and item reliability via their average training losses as a proxy for reputation, and integrates these entity-level factors with interaction-level confidence. The framework is \textbf{model-agnostic}, \textbf{computationally efficient}, and requires \textbf{only two intuitive hyperparameters}. Extensive experiments across multiple datasets and backbone models demonstrate that EARD yields substantial improvements over state-of-the-art baselines (e.g., up to 27.01\% gain in NDCG@50), while incurring negligible additional computational cost. Comprehensive ablation studies and mechanism analyses further confirm EARD's robustness to hyperparameter choices and its practical scalability. These results highlight the importance of entity-aware reliability modeling for denoising implicit feedback and pave the way for more robust recommendation research.

cs.IR↗

arXiv2025

Ming-UniVision: Joint Image Understanding and Generation with a Unified Continuous Tokenizer

Visual tokenization remains a core challenge in unifying visual understanding and generation within the autoregressive paradigm. Existing methods typically employ tokenizers in discrete latent spaces to align with the tokens from large language models, where the quantization errors can limit semantic expressiveness and degrade the capability of vision-language understanding. To address this, we introduce MingTok, a new family of visual tokenizers with a continuous latent space, for unified autoregressive generation and understanding. While understanding tasks favor discriminative high-dimensional features, generation tasks prefer compact low-level codes. Thus, to reconcile these competing demands, MingTok adopts a three-stage sequential architecture involving low-level encoding, semantic expansion, and visual reconstruction. Built on top of it, Ming-UniVision eliminates the need for task-specific visual representations, and unifies diverse vision-language tasks under a single autoregrsssive prediction paradigm. By formulating both understanding and generation as next-token prediction in a shared continuous space, it seamlessly supports multi-round, in-context tasks such as iterative understanding, generation and editing. Empirically, we find that using a unified continuous visual representation reconciles the competing requirements on the tokenizers by the understanding and generation tasks, thereby leading to state-of-the-art level performance across both domains. We hope our findings will facilitate unified visual tokenization in the continuous domain. Inference code and model weights are released to benefit community.

cs.CV↗

arXiv2025

The Debate on RLVR Reasoning Capability Boundary: Shrinkage, Expansion, or Both? A Two-Stage Dynamic View

The ongoing debate on whether reinforcement learning with verifiable rewards (RLVR) expands or shrinks the reasoning capabilities of large language models (LLMs) remains unresolved. Some studies contend that RLVR mainly improves sampling efficiency but at the expense of diversity and exploratory capacity, resulting in capability boundary shrinkage. In contrast, others demonstrate that prolonged training can lead to the emergence of novel reasoning strategies, suggesting capability boundary expansion. To reconcile these contradictory findings, we theoretically and empirically show that both perspectives are partially valid-each aligning with a separate phase in an inherent two-stage probability mass dynamic: (1) Exploitation stage: initially, the model primarily samples explored high-reward and low-reward tokens, while rarely selecting the potentially optimal token. Positive advantage estimates increase the probability of high-reward tokens and decrease those of low-reward tokens, yet the optimal token's probability remains largely unchanged during this stage. (2) Exploration stage: as training advances, the growth rate of previously acquired high-reward tokens slows as their probabilities approach saturation. When a potentially optimal token-now receiving positive advantage estimates-is occasionally sampled, its probability increases, while those of the originally high-reward tokens decrease. This dynamic suggests that over-exploitation during the exploitation stage may lead to capability boundary shrinkage, whereas prolonged training into the exploration stage can promote an expansion of the reasoning capability boundary. Building upon our insights, we revisit the potential of only using relative negative gradients for prolonging training, providing a theoretical and empirical foundation for the development of more advanced reasoning capabilities.

cs.LG↗

arXiv2025

EVLF-FM: Explainable Vision Language Foundation Model for Medicine

Despite the promise of foundation models in medical AI, current systems remain limited - they are modality-specific and lack transparent reasoning processes, hindering clinical adoption. To address this gap, we present EVLF-FM, a multimodal vision-language foundation model (VLM) designed to unify broad diagnostic capability with fine-grain explainability. The development and testing of EVLF-FM encompassed over 1.3 million total samples from 23 global datasets across eleven imaging modalities related to six clinical specialties: dermatology, hepatology, ophthalmology, pathology, pulmonology, and radiology. External validation employed 8,884 independent test samples from 10 additional datasets across five imaging modalities. Technically, EVLF-FM is developed to assist with multiple disease diagnosis and visual question answering with pixel-level visual grounding and reasoning capabilities. In internal validation for disease diagnostics, EVLF-FM achieved the highest average accuracy (0.858) and F1-score (0.797), outperforming leading generalist and specialist models. In medical visual grounding, EVLF-FM also achieved stellar performance across nine modalities with average mIOU of 0.743 and Acc@0.5 of 0.837. External validations further confirmed strong zero-shot and few-shot performance, with competitive F1-scores despite a smaller model size. Through a hybrid training strategy combining supervised and visual reinforcement fine-tuning, EVLF-FM not only achieves state-of-the-art accuracy but also exhibits step-by-step reasoning, aligning outputs with visual evidence. EVLF-FM is an early multi-disease VLM model with explainability and reasoning capabilities that could advance adoption of and trust in foundation models for real-world clinical deployment.

cs.CV↗

arXiv2025

LLaDA-MoE: A Sparse MoE Diffusion Language Model

We introduce LLaDA-MoE, a large language diffusion model with the Mixture-of-Experts (MoE) architecture, trained from scratch on approximately 20T tokens. LLaDA-MoE achieves competitive performance with significantly reduced computational overhead by maintaining a 7B-parameter capacity while activating only 1.4B parameters during inference. Our empirical evaluation reveals that LLaDA-MoE achieves state-of-the-art performance among diffusion language models with larger parameters, surpassing previous diffusion language models LLaDA, LLaDA 1.5, and Dream across multiple benchmarks. The instruct-tuned model LLaDA-MoE-7B-A1B-Instruct demonstrates capabilities comparable to Qwen2.5-3B-Instruct in knowledge understanding, code generation, mathematical reasoning, agent and alignment tasks, despite using fewer active parameters. Our results show that integrating a sparse MoE architecture into the training objective of masked diffusion language models still brings out MoE's strengths under efficient inference with few active parameters, and opens ample room for further exploration of diffusion language models. LLaDA-MoE models are available at Huggingface.

cs.CL↗

arXiv2025

Embedding Domain Knowledge for Large Language Models via Reinforcement Learning from Augmented Generation

Large language models (LLMs) often exhibit limited performance on domain-specific tasks due to the natural disproportionate representation of specialized information in their training data and the static nature of these datasets. Knowledge scarcity and temporal lag create knowledge gaps for domain applications. While post-training on domain datasets can embed knowledge into models, existing approaches have some limitations. Continual Pre-Training (CPT) treats all tokens in domain documents with equal importance, failing to prioritize critical knowledge points, while supervised fine-tuning (SFT) with question-answer pairs struggles to develop the coherent knowledge structures necessary for complex reasoning tasks. To address these challenges, we propose Reinforcement Learning from Augmented Generation (RLAG). Our approach iteratively cycles between sampling generations and optimizing the model through calculated rewards, effectively embedding critical and contextually coherent domain knowledge. We select generated outputs with the highest log probabilities as the sampling result, then compute three tailored reward metrics to guide the optimization process. To comprehensively evaluate domain expertise, we assess answer accuracy and the rationality of explanations generated for correctly answered questions. Experimental results across medical, legal, astronomy, and current events datasets demonstrate that our proposed method significantly outperforms baseline approaches. Our code and data are open sourced at https://github.com/ChaojunNie/RLAG.

cs.CL↗

arXiv2025

Vim-F: Visual State Space Model Benefiting from Learning in the Frequency Domain

In recent years, State Space Models (SSMs) with efficient hardware-aware designs, known as the Mamba deep learning models, have made significant progress in modeling long sequences such as language understanding. Therefore, building efficient and general-purpose visual backbones based on SSMs is a promising direction. Compared to traditional convolutional neural networks (CNNs) and Vision Transformers (ViTs), the performance of Vision Mamba (ViM) methods is not yet fully competitive. To enable SSMs to process image data, ViMs typically flatten 2D images into 1D sequences, inevitably ignoring some 2D local dependencies, thereby weakening the model's ability to interpret spatial relationships from a global perspective. We use Fast Fourier Transform (FFT) to obtain the spectrum of the feature map and add it to the original feature map, enabling ViM to model a unified visual representation in both frequency and spatial domains. The introduction of frequency domain information enables ViM to have a global receptive field during scanning. We propose a novel model called Vim-F, which employs pure Mamba encoders and scans in both the frequency and spatial domains. Moreover, we question the necessity of position embedding in ViM and remove it accordingly in Vim-F, which helps to fully utilize the efficient long-sequence modeling capability of ViM. Finally, we redesign a patch embedding for Vim-F, leveraging a convolutional stem to capture more local correlations, further improving the performance of Vim-F. Code is available at: https://github.com/yws-wxs/Vim-F.

cs.CV↗

arXiv2025

Understanding and Mitigating Overrefusal in LLMs from an Unveiling Perspective of Safety Decision Boundary

Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks, yet they often refuse to answer legitimate queries--a phenomenon known as overrefusal. Overrefusal typically stems from over-conservative safety alignment, causing models to treat many reasonable prompts as potentially risky. To systematically understand this issue, we probe and leverage the models' safety decision boundaries to analyze and mitigate overrefusal. Our findings reveal that overrefusal is closely tied to misalignment at these boundary regions, where models struggle to distinguish subtle differences between benign and harmful content. Building on these insights, we present RASS, an automated framework for prompt generation and selection that strategically targets overrefusal prompts near the safety boundary. By harnessing steering vectors in the representation space, RASS efficiently identifies and curates boundary-aligned prompts, enabling more effective and targeted mitigation of overrefusal. This approach not only provides a more precise and interpretable view of model safety decisions but also seamlessly extends to multilingual scenarios. We have explored the safety decision boundaries of various LLMs and construct the MORBench evaluation set to facilitate robust assessment of model safety and helpfulness across multiple languages. Code and datasets are available at https://github.com/Master-PLC/RASS.

cs.AI↗

arXiv2025

RED: Unleashing Token-Level Rewards from Holistic Feedback via Reward Redistribution

Reinforcement learning from human feedback (RLHF) offers a promising approach to aligning large language models (LLMs) with human preferences. Typically, a reward model is trained or supplied to act as a proxy for humans in evaluating generated responses during the reinforcement training phase. However, current reward models operate as sequence-to-one models, allocating a single, sparse, and delayed reward to an entire output sequence. This approach may overlook the significant contributions of individual tokens toward the desired outcome. To this end, we propose a more fine-grained, token-level guidance approach for RL training. Specifically, we introduce RED, a novel reward redistribition method that evaluates and assigns specific credit to each token using an off-the-shelf reward model. Utilizing these fine-grained rewards enhances the model's understanding of language nuances, leading to more precise performance improvements. Notably, our method does not require modifying the reward model or introducing additional training steps, thereby incurring minimal computational costs. Experimental results across diverse datasets and tasks demonstrate the superiority of our approach.

cs.CL↗

arXiv2025

Towards High Data Efficiency in Reinforcement Learning with Verifiable Reward

Recent advances in large reasoning models have leveraged reinforcement learning with verifiable rewards (RLVR) to improve reasoning capabilities. However, scaling these methods typically requires extensive rollout computation and large datasets, leading to high training costs and low data efficiency. To mitigate this issue, we propose DEPO, a Data-Efficient Policy Optimization pipeline that combines optimized strategies for both offline and online data selection. In the offline phase, we curate a high-quality subset of training samples based on diversity, influence, and appropriate difficulty. During online RLVR training, we introduce a sample-level explorability metric to dynamically filter samples with low exploration potential, thereby reducing substantial rollout computational costs. Furthermore, we incorporate a replay mechanism for under-explored samples to ensure adequate training, which enhances the model's final convergence performance. Experiments across five reasoning benchmarks show that DEPO consistently outperforms existing methods in both offline and online data selection scenarios. Notably, using only 20% of the training data, our approach achieves a 1.85 times speed-up on AIME24 and a 1.66 times speed-up on AIME25 compared to GRPO trained on the full dataset.

cs.LG↗

arXiv2025

Iterative Low-rank Network for Hyperspectral Image Denoising

Hyperspectral image (HSI) denoising is a crucial preprocessing step for subsequent tasks. The clean HSI usually reside in a low-dimensional subspace, which can be captured by low-rank and sparse representation, known as the physical prior of HSI. It is generally challenging to adequately use such physical properties for effective denoising while preserving image details. This paper introduces a novel iterative low-rank network (ILRNet) to address these challenges. ILRNet integrates the strengths of model-driven and data-driven approaches by embedding a rank minimization module (RMM) within a U-Net architecture. This module transforms feature maps into the wavelet domain and applies singular value thresholding (SVT) to the low-frequency components during the forward pass, leveraging the spectral low-rankness of HSIs in the feature domain. The parameter, closely related to the hyperparameter of the singular vector thresholding algorithm, is adaptively learned from the data, allowing for flexible and effective capture of low-rankness across different scenarios. Additionally, ILRNet features an iterative refinement process that adaptively combines intermediate denoised HSIs with noisy inputs. This manner ensures progressive enhancement and superior preservation of image details. Experimental results demonstrate that ILRNet achieves state-of-the-art performance in both synthetic and real-world noise removal tasks.

cs.CV↗

arXiv2025

Arrows of Math Reasoning Data Synthesis for Large Language Models: Diversity, Complexity and Correctness

Enhancing the mathematical reasoning of large language models (LLMs) demands high-quality training data, yet conventional methods face critical challenges in scalability, cost, and data reliability. To address these limitations, we propose a novel program-assisted synthesis framework that systematically generates a high-quality mathematical corpus with guaranteed diversity, complexity, and correctness. This framework integrates mathematical knowledge systems and domain-specific tools to create executable programs. These programs are then translated into natural language problem-solution pairs and vetted by a bilateral validation mechanism that verifies solution correctness against program outputs and ensures program-problem consistency. We have generated 12.3 million such problem-solving triples. Experiments demonstrate that models fine-tuned on our data significantly improve their inference capabilities, achieving state-of-the-art performance on several benchmark datasets and showcasing the effectiveness of our synthesis approach.

cs.CL↗

arXiv2025

Patient-Specific Deep Reinforcement Learning for Automatic Replanning in Head-and-Neck Cancer Proton Therapy

Anatomical changes during intensity-modulated proton therapy (IMPT) for head-and-neck cancer (HNC) can shift Bragg peaks, risking tumor underdosing and organ-at-risk overdosing. Treatment replanning is often required to maintain clinically acceptable treatment quality. However, current manual replanning processes are resource-intensive and time-consuming. We propose a patient-specific deep reinforcement learning (DRL) framework for automated IMPT replanning, with a reward-shaping mechanism based on a $150$-point plan quality score addressing competing clinical objectives. We formulate the planning process as a reinforcement learning problem where agents learn control policies to adjust optimization priorities, maximizing plan quality. Unlike population-based approaches, our framework trains agents for each patient using their planning Computed Tomography (CT) and augmented anatomies simulating anatomical changes (tumor progression and regression). This patient-specific approach leverages anatomical similarities along the treatment course, enabling effective plan adaptation. We implemented two DRL algorithms, Deep Q-Network and Proximal Policy Optimization, using dose-volume histograms (DVHs) as state representations and a $22$-dimensional action space of priority adjustments. Evaluation on eight HNC patients using actual replanning CT data showed that both agents improved initial plan scores from $120.78 \pm 17.18$ to $139.59 \pm 5.50$ (DQN) and $141.50 \pm 4.69$ (PPO), surpassing the replans manually generated by a human planner ($136.32 \pm 4.79$). Clinical validation confirms that improvements translate to better tumor coverage and OAR sparing across diverse anatomical changes. This work highlights DRL's potential in addressing geometric and dosimetric complexities of adaptive proton therapy, offering efficient offline adaptation solutions and advancing online adaptive proton therapy.

physics.med-ph↗

arXiv2025

WSM: Decay-Free Learning Rate Schedule via Checkpoint Merging for LLM Pre-training

Recent advances in learning rate (LR) scheduling have demonstrated the effectiveness of decay-free approaches that eliminate the traditional decay phase while maintaining competitive performance. Model merging techniques have emerged as particularly promising solutions in this domain. We present Warmup-Stable and Merge (WSM), a general framework that establishes a formal connection between learning rate decay and model merging. WSM provides a unified theoretical foundation for emulating various decay strategies-including cosine decay, linear decay and inverse square root decay-as principled model averaging schemes, while remaining fully compatible with diverse optimization methods. Through extensive experiments, we identify merge duration-the training window for checkpoint aggregation-as the most critical factor influencing model performance, surpassing the importance of both checkpoint interval and merge quantity. Our framework consistently outperforms the widely-adopted Warmup-Stable-Decay (WSD) approach across multiple benchmarks, achieving significant improvements of +3.5% on MATH, +2.9% on HumanEval, and +5.5% on MMLU-Pro. The performance advantages extend to supervised fine-tuning scenarios, highlighting WSM's potential for long-term model refinement.

cs.CL↗

arXiv2025

Lifting spin degeneracy in rhombohedral trilayer graphene for high magnetoresistance applications

Many exotic properties in rhombohedral (or ABC-stacked) multilayer graphene have recently been reported experimentally. In this Letter, we first reveal the underlying mechanism of spin degeneracy lifting in rhombohedral trilayer graphene. Then, we propose a design concept for all-rhombohedral graphene-based magnetic tunnel junctions (MTJs) by utilizing pristine, back-gated, and top-gated ABC-stacked trilayer graphene, which exhibit semimetallic (conducting), semiconducting (insulating), and half-metallic (ferromagnetic) behavior, respectively. This enables the realization of an "all-in-one" magnetic tunnel junction based entirely on trilayer graphene. This design enables voltage-controlled spintronics (lower power than conventional MTJs) with perfect interfacial matching and sub-nm thickness uniformity across 4-inch wafers. Using first-principles calculations and the non-equilibrium Greens function, we comprehensively study electronic structures and transport properties of these all-graphene MTJs. Furthermore, we demonstrate that their characteristics can be tuned via a perpendicular electric field and electron doping. Our findings offer a new concept for the development of fully graphene-based spintronic devices utilizing the three distinct electronic phases of rhombohedral trilayer graphene.

cond-mat.mtrl-sci↗