arXiv subjects

Chenyu Wang

Publications and source records attributed to Chenyu Wang.

At least 73 records · Page 4Linked to original sources

arXiv2025

Learning Diffusion Models with Flexible Representation Guidance

Diffusion models can be improved with additional guidance towards more effective representations of input. Indeed, prior empirical work has already shown that aligning internal representations of the diffusion model with those of pre-trained models improves generation quality. In this paper, we present a systematic framework for incorporating representation guidance into diffusion models. We provide alternative decompositions of denoising models along with their associated training criteria, where the decompositions determine when and how the auxiliary representations are incorporated. Guided by our theoretical insights, we introduce two new strategies for enhancing representation alignment in diffusion models. First, we pair examples with target representations either derived from themselves or arisen from different synthetic modalities, and subsequently learn a joint model over the multimodal pairs. Second, we design an optimal training curriculum that balances representation learning and data generation. Our experiments across image, protein sequence, and molecule generation tasks demonstrate superior performance as well as accelerated training. In particular, on the class-conditional ImageNet $256\times 256$ benchmark, our guidance results in $23.3$ times faster training than the original SiT-XL as well as four times speedup over the state-of-the-art method REPA. The code is available at https://github.com/ChenyuWang-Monica/REED.

cs.LG↗

arXiv2025

Better Together: Leveraging Unpaired Multimodal Data for Stronger Unimodal Models

Traditional multimodal learners find unified representations for tasks like visual question answering, but rely heavily on paired datasets. However, an overlooked yet potentially powerful question is: can one leverage auxiliary unpaired multimodal data to directly enhance representation learning in a target modality? We introduce UML: Unpaired Multimodal Learner, a modality-agnostic training paradigm in which a single model alternately processes inputs from different modalities while sharing parameters across them. This design exploits the assumption that different modalities are projections of a shared underlying reality, allowing the model to benefit from cross-modal structure without requiring explicit pairs. Theoretically, under linear data-generating assumptions, we show that unpaired auxiliary data can yield representations strictly more informative about the data-generating process than unimodal training. Empirically, we show that using unpaired data from auxiliary modalities -- such as text, audio, or images -- consistently improves downstream performance across diverse unimodal targets such as image and audio. Our project page: https://unpaired-multimodal.github.io/

cs.LG↗

arXiv2025

Next Semantic Scale Prediction via Hierarchical Diffusion Language Models

In this paper we introduce Hierarchical Diffusion Language Models (HDLM) -- a novel family of discrete diffusion models for language modeling. HDLM builds on a hierarchical vocabulary where low-level tokens with detailed semantics are surjectively mapped to high-level tokens with coarse-grained meanings. In the forward process, each token is independently perturbed to its higher-level ancestor with more abstract semantics according to the scheduler, while in the reverse process the model progressively predicts the next, more detailed semantics. Taken together, HDLM provides a general time-varying next semantic scale prediction process for language modeling. We derive closed-form expressions for the diffusion Evidence Lower Bound (ELBO), and show that HDLM can be implemented in a flexible manner while including the existing MDLM as a special case. We also propose practical training techniques based on the insights. Extensive text generation experiments validate the effectiveness of HDLM, which demonstrates consistently lower validation and generative perplexity than baselines.

cs.CL↗

arXiv2025

Backdoors in Code Summarizers: How Bad Is It?

Code LLMs are increasingly employed in software development. However, studies have shown that they are vulnerable to backdoor attacks: when a trigger (a specific input pattern) appears in the input, the backdoor will be activated and cause the model to generate malicious outputs. Researchers have designed various triggers and demonstrated the feasibility of implanting backdoors by poisoning a fraction of the training data. Some basic conclusions have been made, such as backdoors becoming easier to implant when more training data is modified. However, existing research has not explored other factors influencing backdoor attacks on Code LLMs, such as training batch size, epoch number, and the broader design space for triggers, e.g., trigger length. To bridge this gap, we use code summarization as an example to perform an empirical study that systematically investigates the factors affecting backdoor effectiveness and understands the extent of the threat posed. Three categories of factors are considered: data, model, and inference, revealing previously overlooked findings. We find that the prevailing consensus -- that attacks are ineffective at extremely low poisoning rates -- is incorrect. The absolute number of poisoned samples matters as well. Specifically, poisoning just 20 out of 454K samples (0.004% poisoning rate -- far below the minimum setting of 0.1% in prior studies) successfully implants backdoors! Moreover, the common defense is incapable of removing even a single poisoned sample from it. Additionally, small batch sizes increase the risk of backdoor attacks. We also uncover other critical factors such as trigger types, trigger length, and the rarity of tokens in the triggers, leading to valuable insights for assessing Code LLMs' vulnerability to backdoor attacks. Our study highlights the urgent need for defense mechanisms against extremely low poisoning rate settings.

cs.CR↗

arXiv2025

SlimPack: Fine-Grained Asymmetric Packing for Balanced and Efficient Variable-Length LLM Training

The efficient distributed training of Large Language Models (LLMs) is severely hampered by the extreme variance in context lengths. This data heterogeneity, amplified by conventional packing strategies and asymmetric forward-backward costs, leads to critical inefficiencies such as cascading workload imbalances and severe hardware underutilization. Existing solutions attempt to mitigate these challenges, but often at the expense of memory or communication efficiency. To address these challenges, we introduce SlimPack, a framework that fundamentally rethinks data packing and scheduling by decomposing samples into fine-grained slices. This slice-level decomposition immediately mitigates critical memory and communication bottlenecks by transforming large, volatile workloads into a stream of smaller, manageable units. This flexibility is then harnessed for our core innovation, Asymmetric Partitioning, which assembles balanced scheduling units uniquely optimized for the different demands of the forward and backward passes. Orchestrated by a two-phase solver and a high-fidelity simulator, SlimPack holistically resolves imbalances across all parallel dimensions. Extensive experiments demonstrate that SlimPack achieves up to a $2.8\times$ training throughput improvement over baselines, breaking the conventional trade-off by delivering both superior balance and high resource efficiency.

cs.AI↗

arXiv2025

Lightweight MSA Design Advances Protein Folding From Evolutionary Embeddings

Protein structure prediction often hinges on multiple sequence alignments (MSAs), which underperform on low-homology and orphan proteins. We introduce PLAME, a lightweight MSA design framework that leverages evolutionary embeddings from pretrained protein language models to generate MSAs that better support downstream folding. PLAME couples these embeddings with a conservation--diversity loss that balances agreement on conserved positions with coverage of plausible sequence variation. Beyond generation, we develop (i) an MSA selection strategy to filter high-quality candidates and (ii) a sequence-quality metric that is complementary to depth-based measures and predictive of folding gains. On AlphaFold2 low-homology/orphan benchmarks, PLAME delivers state-of-the-art improvements in structure accuracy (e.g., lDDT/TM-score), with consistent gains when paired with AlphaFold3. Ablations isolate the benefits of the selection strategy, and case studies elucidate how MSA characteristics shape AlphaFold confidence and error modes. Finally, we show PLAME functions as a lightweight adapter, enabling ESMFold to approach AlphaFold2-level accuracy while retaining ESMFold-like inference speed. PLAME thus provides a practical path to high-quality folding for proteins lacking strong evolutionary neighbors.

cs.LG↗

arXiv2025

Latent Danger Zone: Distilling Unified Attention for Cross-Architecture Black-box Attacks

Black-box adversarial attacks remain challenging due to limited access to model internals. Existing methods often depend on specific network architectures or require numerous queries, resulting in limited cross-architecture transferability and high query costs. To address these limitations, we propose JAD, a latent diffusion model framework for black-box adversarial attacks. JAD generates adversarial examples by leveraging a latent diffusion model guided by attention maps distilled from both a convolutional neural network (CNN) and a Vision Transformer (ViT) models. By focusing on image regions that are commonly sensitive across architectures, this approach crafts adversarial perturbations that transfer effectively between different model types. This joint attention distillation strategy enables JAD to be architecture-agnostic, achieving superior attack generalization across diverse models. Moreover, the generative nature of the diffusion framework yields high adversarial sample generation efficiency by reducing reliance on iterative queries. Experiments demonstrate that JAD offers improved attack generalization, generation efficiency, and cross-architecture transferability compared to existing methods, providing a promising and effective paradigm for black-box adversarial attacks.

cs.LG↗

arXiv2025

Rest2Visual: Predicting Visually Evoked fMRI from Resting-State Scans

Understanding how spontaneous brain activity relates to stimulus-driven neural responses is a fundamental challenge in cognitive neuroscience. While task-based functional magnetic resonance imaging (fMRI) captures localized stimulus-evoked brain activation, its acquisition is costly, time-consuming, and difficult to scale across populations. In contrast, resting-state fMRI (rs-fMRI) is task-free and abundant, but lacks direct interpretability. We introduce Rest2Visual, a conditional generative model that predicts visually evoked fMRI (ve-fMRI) from resting-state input and 2D visual stimuli. It follows a volumetric encoder--decoder design, where multiscale 3D features from rs-fMRI are modulated by image embeddings via adaptive normalization, enabling spatially accurate, stimulus-specific activation synthesis. To enable model training, we construct a large-scale triplet dataset from the Natural Scenes Dataset (NSD), aligning each rs-fMRI volume with stimulus images and their corresponding ve-fMRI activation maps. Quantitative evaluation shows that the predicted activations closely match ground truth across standard similarity and representational metrics, and support successful image reconstruction in downstream decoding. Notably, the predicted maps preserve subject-specific structure, demonstrating the model's capacity to generate individualized functional surrogates. Our results provide compelling evidence that individualized spontaneous neural activity can be transformed into stimulus-aligned representations, opening new avenues for scalable, task-free functional brain modeling.

q-bio.NC↗

arXiv2025

Inpainting-Guided Policy Optimization for Diffusion Large Language Models

Masked diffusion large language models (dLLMs) are emerging as promising alternatives to autoregressive LLMs, offering competitive performance while supporting unique generation capabilities such as inpainting. We explore how inpainting can inform RL algorithm design for dLLMs. Aligning LLMs with reinforcement learning faces an exploration challenge: sparse reward signals and sample waste when models fail to discover correct solutions. While this inefficiency affects LLMs broadly, dLLMs offer a distinctive opportunity--their inpainting ability can guide exploration. We introduce IGPO (Inpainting Guided Policy Optimization), an RL framework that strategically inserts partial ground-truth reasoning traces during online sampling. Unlike providing full solutions, inpainting steers exploration toward promising trajectory spaces while preserving self-generated reasoning, bridging supervised fine-tuning and reinforcement learning. We apply IGPO to group-based optimization methods such as GRPO, where exploration failures cause zero advantages and gradients. IGPO restores meaningful gradients while improving sample efficiency. We also propose supervised fine-tuning on synthetically rewritten concise traces that better align with dLLM generation patterns. With additional techniques including entropy-based filtering, our training recipe yields substantial gains across three mathematical benchmarks--GSM8K, Math500, and AMC--achieving new state-of-the-art results for full-attention masked dLLMs.

cs.LG↗

arXiv2025

From Long to Short: LLMs Excel at Trimming Own Reasoning Chains

O1/R1 style large reasoning models (LRMs) signal a substantial leap forward over conventional instruction-following LLMs. By applying test-time scaling to generate extended reasoning paths, they establish many SOTAs across a wide range of complex reasoning tasks. However, recent studies show that LRMs are prone to suffer from overthinking -- the tendency to overcomplicate simple problems, leading to excessive strategy switching and long, convoluted reasoning traces that hinder their interpretability. To mitigate this issue, we conduct a systematic investigation into the reasoning efficiency of a broad set of LRMs and uncover a common dilemma: the difficulty in balancing multiple generation objectives such as correctness and brevity. Based on this discovery, we propose a test-time scaling method, EDIT (Efficient Dynamic Inference Trimming), which efficiently guides LRMs to identify the shortest correct reasoning paths at test time. EDIT employs constraint-guided generation while jointly tracking length and answer distributions under varying constraints, allowing it to select responses that strike an optimal balance between conciseness and correctness. Extensive experiments across diverse models and datasets show that EDIT substantially enhance the reasoning efficiency, producing compact yet informative outputs that improve readability and user experience.

cs.AI↗

arXiv2025

From Promise to Practical Reality: Transforming Diffusion MRI Analysis with Fast Deep Learning Enhancement

Fiber orientation distribution (FOD) is an advanced diffusion MRI modeling technique that represents complex white matter fiber configurations, and a key step for subsequent brain tractography and connectome analysis. Its reliability and accuracy, however, heavily rely on the quality of the MRI acquisition and the subsequent estimation of the FODs at each voxel. Generating reliable FODs from widely available clinical protocols with single-shell and low-angular-resolution acquisitions remains challenging but could potentially be addressed with recent advances in deep learning-based enhancement techniques. Despite advancements, existing methods have predominantly been assessed on healthy subjects, which have proved to be a major hurdle for their clinical adoption. In this work, we validate a newly optimized enhancement framework, FastFOD-Net, across healthy controls and six neurological disorders. This accelerated end-to-end deep learning framework enhancing FODs with superior performance and delivering training/inference efficiency for clinical use ($60\times$ faster comparing to its predecessor). With the most comprehensive clinical evaluation to date, our work demonstrates the potential of FastFOD-Net in accelerating clinical neuroscience research, empowering diffusion MRI analysis for disease differentiation, improving interpretability in connectome applications, and reducing measurement errors to lower sample size requirements. Critically, this work will facilitate the more widespread adoption of, and build clinical trust in, deep learning based methods for diffusion MRI enhancement. Specifically, FastFOD-Net enables robust analysis of real-world, clinical diffusion MRI data, comparable to that achievable with high-quality research acquisitions.

cs.CV↗

arXiv2025

"What" x "When" working memory representations using Laplace Neural Manifolds

Working memory - the ability to remember recent events as they recede continuously into the past - requires the ability to represent any stimulus at any time delay. This property requires neurons coding working memory to show mixed selectivity, with conjunctive receptive fields (RFs) for stimuli and time, forming a representation of 'what' x 'when'. We study the properties of such a working memory in simple experiments where a single stimulus must be remembered for a short time. The requirement of conjunctive receptive fields allows the covariance matrix of the network to decouple neatly, allowing an understanding of the low-dimensional dynamics of the population. Different choices of temporal basis functions lead to qualitatively different dynamics. We study a specific choice - a Laplace space with exponential basis functions for time coupled to an "Inverse Laplace" space with circumscribed basis functions in time. We refer to this choice with basis functions that evenly tile log time as a Laplace Neural Manifold. Despite the fact that they are related to one another by a linear projection, the Laplace population shows a stable stimulus-specific subspace whereas the Inverse Laplace population shows rotational dynamics. The growth of the rank of the covariance matrix with time depends on the density of the temporal basis set; logarithmic tiling shows good agreement with data. We sketch a continuous attractor CANN that constructs a Laplace Neural Manifold. The attractor in the Laplace space appears as an edge; the attractor for the inverse space appears as a bump. This work provides a map for going from more abstract cognitive models of WM to circuit-level implementation using continuous attractor neural networks, and places constraints on the types of neural dynamics that support working memory.

q-bio.NC↗

arXiv2025

CAD-MLLM: Unifying Multimodality-Conditioned CAD Generation With MLLM

This paper aims to design a unified Computer-Aided Design (CAD) generation system that can easily generate CAD models based on the user's inputs in the form of textual description, images, point clouds, or even a combination of them. Towards this goal, we introduce the CAD-MLLM, the first system capable of generating parametric CAD models conditioned on the multimodal input. Specifically, within the CAD-MLLM framework, we leverage the command sequences of CAD models and then employ advanced large language models (LLMs) to align the feature space across these diverse multi-modalities data and CAD models' vectorized representations. To facilitate the model training, we design a comprehensive data construction and annotation pipeline that equips each CAD model with corresponding multimodal data. Our resulting dataset, named Omni-CAD, is the first multimodal CAD dataset that contains textual description, multi-view images, points, and command sequence for each CAD model. It contains approximately 450K instances and their CAD construction sequences. To thoroughly evaluate the quality of our generated CAD models, we go beyond current evaluation metrics that focus on reconstruction quality by introducing additional metrics that assess topology quality and surface enclosure extent. Extensive experimental results demonstrate that CAD-MLLM significantly outperforms existing conditional generative methods and remains highly robust to noises and missing points. The project page and more visualizations can be found at: https://cad-mllm.github.io/

cs.CV↗

arXiv2025

High-gain optical parametric amplification with a continuous-wave pump using a domain-engineered thin-film lithium niobate waveguide

While thin film lithium niobate (TFLN) is known for efficient signal generation, on-chip signal amplification remains challenging from fully integrated optical communication circuits. Here we demonstrate the continuous-wave-pump optical parametric amplification (OPA) using an x-cut domain-engineered TFLN waveguide, with high gain over the telecom band up to 13.9 dB, and test it for high signal-to-noise ratio signal amplification using a commercial optical communication module pair. Fabricated in wafer scale using common process as devices including modulators, this OPA device marks an important step in TFLN photonic integration.

physics.optics↗

arXiv2025

CellFlux: Simulating Cellular Morphology Changes via Flow Matching

Building a virtual cell capable of accurately simulating cellular behaviors in silico has long been a dream in computational biology. We introduce CellFlux, an image-generative model that simulates cellular morphology changes induced by chemical and genetic perturbations using flow matching. Unlike prior methods, CellFlux models distribution-wise transformations from unperturbed to perturbed cell states, effectively distinguishing actual perturbation effects from experimental artifacts such as batch effects -- a major challenge in biological data. Evaluated on chemical (BBBC021), genetic (RxRx1), and combined perturbation (JUMP) datasets, CellFlux generates biologically meaningful cell images that faithfully capture perturbation-specific morphological changes, achieving a 35% improvement in FID scores and a 12% increase in mode-of-action prediction accuracy over existing methods. Additionally, CellFlux enables continuous interpolation between cellular states, providing a potential tool for studying perturbation dynamics. These capabilities mark a significant step toward realizing virtual cell modeling for biomedical research. Project page: https://yuhui-zh15.github.io/CellFlux/.

q-bio.QM↗

arXiv2025

Deep Learning and Explainable AI: New Pathways to Genetic Insights

Deep learning-based AI models have been extensively applied in genomics, achieving remarkable success across diverse applications. As these models gain prominence, there exists an urgent need for interpretability methods to establish trustworthiness in model-driven decisions. For genetic researchers, interpretable insights derived from these models hold significant value in providing novel perspectives for understanding biological processes. Current interpretability analyses in genomics predominantly rely on intuition and experience rather than rigorous theoretical foundations. In this review, we systematically categorize interpretability methods into input-based and model-based approaches, while critically evaluating their limitations through concrete biological application scenarios. Furthermore, we establish theoretical underpinnings to elucidate the origins of these constraints through formal mathematical demonstrations, aiming to assist genetic researchers in better understanding and designing models in the future. Finally, we provide feasible suggestions for future research on interpretability in the field of genetics.

q-bio.GN↗

arXiv2025

Seeing from Another Perspective: Evaluating Multi-View Understanding in MLLMs

Multi-view understanding, the ability to reconcile visual information across diverse viewpoints for effective navigation, manipulation, and 3D scene comprehension, is a fundamental challenge in Multi-Modal Large Language Models (MLLMs) to be used as embodied agents. While recent MLLMs have shown impressive advances in high-level reasoning and planning, they frequently fall short when confronted with multi-view geometric consistency and cross-view correspondence. To comprehensively evaluate the challenges of MLLMs in multi-view scene reasoning, we propose All-Angles Bench, a benchmark of over 2,100 human carefully annotated multi-view question-answer pairs across 90 diverse real-world scenes. Our six tasks (counting, attribute identification, relative distance, relative direction, object manipulation, and camera pose estimation) specifically test model's geometric correspondence and the capacity to align information consistently across views. Our extensive experiments, benchmark on 27 representative MLLMs including Gemini-2.0-Flash, Claude-3.7-Sonnet, and GPT-4o against human evaluators reveals a substantial performance gap, indicating that current MLLMs remain far from human-level proficiency. Through in-depth analysis, we show that MLLMs are particularly underperforming under two aspects: (1) cross-view correspondence for partially occluded views and (2) establishing the coarse camera poses. These findings highlight the necessity of domain-specific refinements or modules that embed stronger multi-view awareness. We believe that our All-Angles Bench offers valuable insights and contribute to bridging the gap between MLLMs and human-level multi-view understanding. The project and benchmark are publicly available at https://danielchyeh.github.io/All-Angles-Bench/.

cs.CV↗

arXiv2025

MLLM-Tool: A Multimodal Large Language Model For Tool Agent Learning

Recently, the astonishing performance of large language models (LLMs) in natural language comprehension and generation tasks triggered lots of exploration of using them as central controllers to build agent systems. Multiple studies focus on bridging the LLMs to external tools to extend the application scenarios. However, the current LLMs' ability to perceive tool use is limited to a single text query, which may result in ambiguity in understanding the users' real intentions. LLMs are expected to eliminate that by perceiving the information in the visual- or auditory-grounded instructions. Therefore, in this paper, we propose MLLM-Tool, a system incorporating open-source LLMs and multi-modal encoders so that the learned LLMs can be conscious of multi-modal input instruction and then select the function-matched tool correctly. To facilitate the evaluation of the model's capability, we collect a dataset featuring multi-modal input tools from HuggingFace. Another essential feature of our dataset is that it also contains multiple potential choices for the same instruction due to the existence of identical functions and synonymous functions, which provides more potential solutions for the same query. The experiments reveal that our MLLM-Tool is capable of recommending appropriate tools for multi-modal instructions. Codes and data are available at https://github.com/MLLM-Tool/MLLM-Tool.

cs.CV↗