arXiv subjects

Zeyu Dong

Publications and source records attributed to Zeyu Dong.

At least 19 recordsLinked to original sources

arXiv2026

Outcome-Guided Distillation: A Teacher-Student Framework to Advance VLM Reasoning in Autonomous Driving

End-to-end (E2E) autonomous driving aims to learn a direct mapping from visual observations to control actions. However, these E2E models often act as black boxes and struggle with complex scenarios. To address this, recent works incorporate Vision-Language Models (VLMs) to provide explicit reasoning, enhancing both interpretability and driving robustness. These approaches typically rely on pre-generated annotations, which suffer from potentially flawed labels and require costly human labor. In this work, we propose a new framework that integrates structured reasoning and geometric precision through a teacher-student architecture. The teacher model introduces reflective reasoning, where the VLM generates logical explanations and then reflectively refines the reasoning under the supervision of ground-truth action. This enhances zero-shot generalization without intermediate labels. The student model distills the teacher's reasoning capabilities via supervised fine-tuning. We also design a separate waypoint decoder that interprets textual reasoning into continuous trajectories. Our proposed solution integrates two goals: providing explicit reasoning for interpretability and delivering robust and accurate driving performance. It leverages the synergy between these two objectives within a staged inference engine to enhance driving performance and explicitly uses the reasoning to guide driving prediction. Evaluated on Waymo benchmarks, our framework outperforms classical reasoning-based baselines in zero-shot reasoning, waypoint accuracy, and inference efficiency. Our experiments validate this design, demonstrating that the reasoning text makes a significant contribution to driving inference, resulting in around a 24% improvement in performance compared to an identical model that lacks reasoning. Our work advances reasoning-driven autonomous driving toward interpretable and deployable systems.

cs.RO↗

arXiv2026

What Can I Edit? Open-Ended Strategy Discovery and the Emotion Editability Landscape

Emotional image editing requires more than applying affective filters or modifying predefined visual factors: an effective edit must identify what a particular image can afford for a target emotion. Existing affective image manipulation methods, including recent agentic variants, largely operate within bounded strategy spaces based on predefined factor taxonomies, knowledge libraries, or conventional editing templates, and therefore often miss image-specific, context-grounded strategies. We introduce EmoScope, a multi-agent framework that reframes the task from "how should I edit?" to "what can I edit?" EmoScope first discovers an image-specific editable space through emotion-conditioned affordance reasoning, then uses a semantic hierarchy of anchors, variables, and context to balance content consistency and emotional expressiveness before executing and verifying the edit. Because its plans are expressed as image-specific affordances rather than retrieved templates, EmoScope also exposes the editing strategy as an interactive surface for user refinement at the plan level. In a large-scale human evaluation covering all eight Mikels emotion categories, with 4,693 valid responses across 1,824 pairwise questions, participants preferred EmoScope over two competitive baselines by 88.1% on average. Attribution analysis further shows that EmoScope selects target-emotion-adaptive strategies rather than applying a uniform template. The same affordance-level plan also supports lightweight user refinement in an interactive pilot. Finally, we show that classifier-based metrics exhibit emotion-conditional blind spots toward non-stereotypical, context-grounded edits, and present a relative content-emotion preference-affinity landscape showing that EmoScope's advantage varies systematically across image-emotion combinations.

cs.CV↗

arXiv2026

Iterative Thresholding Pursuit with Continuation for $\ell_{1-2}$-Regularized Sparse Recovery

Sparse recovery aims to reconstruct sparse signals from underdetermined and possibly noisy linear measurements. Existing $\ell_{1-2}$ iterative thresholding schemes are first-order methods. We propose an iterative thresholding pursuit method with continuation (ITP-C) for $\ell_{1-2}$-regularized sparse recovery. The method goes beyond first-order thresholding by combining the active-set identification capability of the $\ell_{1-2}$ proximal step with a restricted least-squares pursuit step that provides a second-order update on the identified support. The support is generated adaptively by the thresholding update, and no prior knowledge of the true sparsity level is required. To control the possible instability of the pursuit step while preserving the descent structure of the continuation scheme, we impose a strict descent check with respect to the dynamic objective. We establish convergence of the generated sequence under the Kurdyka-Lojasiewicz framework and prove a local oracle-type property after correct support identification. Numerical experiments on synthetic sparse recovery and image reconstruction illustrate the descent preservation of the proposed safeguard and demonstrate the improved recovery performance of ITP-C over the state-of-the-art baselines.

math.NA↗

arXiv2026

Measuring Iterative Temporal Reasoning with Time Puzzles

Tool use, such as web search, has become a standard capability even in freely available large language models (LLMs). However, existing benchmarks evaluate temporal reasoning mainly in static, non-tool-using settings, which poorly reflect how LLMs perform temporal reasoning in practice. We introduce Time Puzzles, a constraint-based date inference task for evaluating iterative temporal reasoning with tools. Each puzzle combines factual temporal anchors with (cross-cultural) calendar relations and may admit one or multiple valid dates. The puzzles are algorithmically generated, enabling controlled and continual evaluation. Across 13 LLMs, even the best model (GPT-5) achieves only 55.3% accuracy without tools, despite using easily searchable facts. While web search improves performance, models perform substantially better when constraints are rewritten with explicit dates, removing the need for factual lookup. These results reveal a gap in reliable tool use for iterative temporal reasoning.

cs.CL↗

arXiv2026

Adaptive Randomized Extended Bregman-Kaczmarz Method for Combined Optimization Problems

Combined optimization problems that couple data-fidelity and regularization terms arise naturally in a wide range of inverse problems. In this paper, we study an adaptive randomized averaging block extended Bregman-Kaczmarz (aRABEBK) method for solving such problems. The proposed method incorporates iteration-wise relaxation parameters that are automatically adjusted using residual information, allowing for more aggressive step sizes without additional manual tuning. We establish a convergence theory for the proposed framework and derive expected linear convergence rate guarantees. Numerical experiments on both synthetic and real data sets for sparse and minimum-norm least-squares problems demonstrate that our aRABEBK method achieves faster convergence and improved robustness compared with state-of-the-art extended Kaczmarz and Bregman-Kaczmarz-type algorithms, including its nonadaptive counterpart.

math.NA↗

arXiv2026

Adaptive Model-Based Reinforcement Learning for Orbit Feedback Control in NSLS-II Storage Ring

The National Synchrotron Light Source II (NSLS-II) uses highly stable electron beam to produce high-quality X-ray beams with high brightness and low-emittance synchrotron radiation. The traditional algorithm to stabilize the beam applies singular value decomposition (SVD) on the orbit response matrix to remove noise and extract actions. Supervised learning has been studied on NSLS-II storage ring stabilization and other accelerator facilities recently. Several problems, for example, machine status drifting, environment noise, and non-linear accelerator dynamics, remain unresolved in the SVD-based and supervised learning algorithms. To address these problems, we propose an adaptive training framework based on model-based reinforcement learning. This framework consists of two types of optimizations: trajectory optimization attempts to minimize the expected total reward in a differentiable environment, and online model optimization learns non-linear machine dynamics through the agent-environment interaction. Through online training, this framework tracks the internal status drifting in the electron beam ring. Simulation and real in-facility experiments on NSLS-II reveal that our method stabilizes the beam position and minimizes the alignment error, defined as the root mean square (RMS) error between adjusted beam positions and the reference position, down to ~1$μ$m.

eess.SY↗

arXiv2026

FROST-Drive: Scalable and Efficient End-to-End Driving with a Frozen Vision Encoder

End-to-end (E2E) models in autonomous driving aim to directly map sensor inputs to control commands, but their ability to generalize to novel and complex scenarios remains a key challenge. The common practice of fully fine-tuning the vision encoder on driving datasets potentially limits its generalization by causing the model to specialize too heavily in the training data. This work challenges the necessity of this training paradigm. We propose FROST-Drive, a novel E2E architecture designed to preserve and leverage the powerful generalization capabilities of a pretrained vision encoder from a Vision-Language Model (VLM). By keeping the encoder's weights frozen, our approach directly transfers the rich, generalized world knowledge from the VLM to the driving task. Our model architecture combines this frozen encoder with a transformer-based adapter for multimodal fusion and a GRU-based decoder for smooth waypoint generation. Furthermore, we introduce a custom loss function designed to directly optimize for Rater Feedback Score (RFS), a metric that prioritizes robust trajectory planning. We conduct extensive experiments on Waymo Open E2E Dataset, a large-scale datasets deliberately curated to capture the long-tail scenarios, demonstrating that our frozen-encoder approach significantly outperforms models that employ full fine-tuning. Our results provide substantial evidence that preserving the broad knowledge of a capable VLM is a more effective strategy for achieving robust, generalizable driving performance than intensive domain-specific adaptation. This offers a new pathway for developing vision-based models that can better handle the complexities of real-world application domains.

cs.CV↗

arXiv2025

A Relaxed Randomized Averaging Block Extended Bregman-Kaczmarz Method for Combined Optimization Problems

Randomized Kaczmarz-type methods are widely used for their simplicity and efficiency in solving large-scale linear systems and optimization problems. However, their applicability is limited when dealing with inconsistent systems or incorporating structural information such as sparsity. In this work, we propose a \emph{relaxed randomized averaging block extended Bregman-Kaczmarz} (rRABEBK) method for solving a broad class of combined optimization problems. The proposed method integrates an averaging block strategy with two relaxation parameters to accelerate convergence and enhance numerical stability. We establish a rigorous convergence theory showing that rRABEBK achieves linear convergence in expectation, with explicit constants that quantify the effect of the relaxation mechanism, and a provably faster rate than the classical randomized extended Bregman-Kaczmarz method. Our method can be readily adapted to sparse least-squares problems and extended to both consistent and inconsistent systems without modification. Complementary numerical experiments corroborate the theoretical findings and demonstrate that rRABEBK significantly outperforms the existing Kaczmarz-type algorithms in terms of both iteration complexity and computational efficiency, highlighting both its practical and theoretical advantages.

math.NA↗

arXiv2025

SGLP: A Similarity Guided Fast Layer Partition Pruning for Compressing Large Deep Models

Layer pruning has emerged as a potent approach to remove redundant layers in the pre-trained network on the purpose of reducing network size and improve computational efficiency. However, existing layer pruning methods mostly overlook the intrinsic connections and inter-dependencies between different layers within complicated deep neural networks. This oversight can result in pruned models that do not preserve the essential characteristics of the pre-trained network as effectively as desired. To address these limitations, we propose a Similarity-Guided Layer Partition (SGLP) Pruning, a novel pruning framework that exploits representation similarity to guide efficient and informed layer removal for compressing large deep models. Our method begins by employing Centered Kernel Alignment (CKA) to quantify representational similarity between layers, uncovering structural patterns within the network. We then apply Fisher Optimal Segmentation on the similarity matrix to partition the network into semantically coherent layer segments. This segmentation allows pruning decisions to respect layer interdependencies and preserve essential knowledge. Within each segment, we introduce a fine-tuning-free importance evaluation using GradNorm, identifying and removing redundant layers in a targeted, segment-wise manner. Experimental results on both image classification tasks and large language models (LLMs) demonstrate that our proposed SGLP outperforms the state-of-the-art methods in accuracy and efficiency. Our approach achieves significant model compression with minimal performance degradation, making it well-suited for deployment in resource-limited environments.

cs.LG↗

arXiv2025

PanoLora: Bridging Perspective and Panoramic Video Generation with LoRA Adaptation

Generating high-quality 360° panoramic videos remains a significant challenge due to the fundamental differences between panoramic and traditional perspective-view projections. While perspective videos rely on a single viewpoint with a limited field of view, panoramic content requires rendering the full surrounding environment, making it difficult for standard video generation models to adapt. Existing solutions often introduce complex architectures or large-scale training, leading to inefficiency and suboptimal results. Motivated by the success of Low-Rank Adaptation (LoRA) in style transfer tasks, we propose treating panoramic video generation as an adaptation problem from perspective views. Through theoretical analysis, we demonstrate that LoRA can effectively model the transformation between these projections when its rank exceeds the degrees of freedom in the task. Our approach efficiently fine-tunes a pretrained video diffusion model using only approximately 1,000 videos while achieving high-quality panoramic generation. Experimental results demonstrate that our method maintains proper projection geometry and surpasses previous state-of-the-art approaches in visual quality, left-right consistency, and motion diversity.

cs.CV↗

arXiv2025

AMMKD: Adaptive Multimodal Multi-teacher Distillation for Lightweight Vision-Language Models

The success of large-scale visual language pretraining (VLP) models has driven widespread adoption of image-text retrieval tasks. However, their deployment on mobile devices remains limited due to large model sizes and computational complexity. We propose Adaptive Multi-Modal Multi-Teacher Knowledge Distillation (AMMKD), a novel framework that integrates multi-modal feature fusion, multi-teacher distillation, and adaptive optimization to deliver lightweight yet effective retrieval models. Specifically, our method begins with a feature fusion network that extracts and merges discriminative features from both the image and text modalities. To reduce model parameters and further improve performance, we design a multi-teacher knowledge distillation framework to pre-train two CLIP teacher models. We decouple modalities by pre-computing and storing text features as class vectors via the teacher text encoder to enhance efficiency. To better align teacher and student outputs, we apply KL scatter for probability distribution matching. Finally, we design an adaptive dynamic weighting scheme that treats multi-teacher distillation as a multi-objective optimization problem. By leveraging gradient space diversity, we dynamically adjust the influence of each teacher, reducing conflicts and guiding the student toward more optimal learning directions. Extensive experiments on three benchmark datasets demonstrate that AMMKD achieves superior performance while significantly reducing model complexity, validating its effectiveness and flexibility.

cs.CV↗

arXiv2025

Frequency-Aligned Knowledge Distillation for Lightweight Spatiotemporal Forecasting

Spatiotemporal forecasting tasks, such as traffic flow, combustion dynamics, and weather forecasting, often require complex models that suffer from low training efficiency and high memory consumption. This paper proposes a lightweight framework, Spectral Decoupled Knowledge Distillation (termed SDKD), which transfers the multi-scale spatiotemporal representations from a complex teacher model to a more efficient lightweight student network. The teacher model follows an encoder-latent evolution-decoder architecture, where its latent evolution module decouples high-frequency details and low-frequency trends using convolution and Transformer (global low-frequency modeler). However, the multi-layer convolution and deconvolution structures result in slow training and high memory usage. To address these issues, we propose a frequency-aligned knowledge distillation strategy, which extracts multi-scale spectral features from the teacher's latent space, including both high and low frequency components, to guide the lightweight student model in capturing both local fine-grained variations and global evolution patterns. Experimental results show that SDKD significantly improves performance, achieving reductions of up to 81.3% in MSE and in MAE 52.3% on the Navier-Stokes equation dataset. The framework effectively captures both high-frequency variations and long-term trends while reducing computational complexity. Our codes are available at https://github.com/itsnotacie/SDKD

cs.LG↗

arXiv2025

SRKD: Towards Efficient 3D Point Cloud Segmentation via Structure- and Relation-aware Knowledge Distillation

3D point cloud segmentation faces practical challenges due to the computational complexity and deployment limitations of large-scale transformer-based models. To address this, we propose a novel Structure- and Relation-aware Knowledge Distillation framework, named SRKD, that transfers rich geometric and semantic knowledge from a large frozen teacher model (>100M) to a lightweight student model (<15M). Specifically, we propose an affinity matrix-based relation alignment module, which distills structural dependencies from the teacher to the student through point-wise similarity matching, enhancing the student's capability to learn contextual interactions. Meanwhile, we introduce a cross-sample mini-batch construction strategy that enables the student to perceive stable and generalized geometric structure. This aligns across diverse point cloud instances of the teacher, rather than within a single sample. Additionally, KL divergence is applied to align semantic distributions, and ground-truth supervision further reinforces accurate segmentation. Our method achieves state of the art performance with significantly reduced model complexity, demonstrating its effectiveness and efficiency in real-world deployment scenarios. Our Code is available at https://github.com/itsnotacie/SRKD.

cs.CV↗

arXiv2025

SepPrune: Structured Pruning for Efficient Deep Speech Separation

Although deep learning has substantially advanced speech separation in recent years, most existing studies continue to prioritize separation quality while overlooking computational efficiency, an essential factor for low-latency speech processing in real-time applications. In this paper, we propose SepPrune, the first structured pruning framework specifically designed to compress deep speech separation models and reduce their computational cost. SepPrune begins by analyzing the computational structure of a given model to identify layers with the highest computational burden. It then introduces a differentiable masking strategy to enable gradient-driven channel selection. Based on the learned masks, SepPrune prunes redundant channels and fine-tunes the remaining parameters to recover performance. Extensive experiments demonstrate that this learnable pruning paradigm yields substantial advantages for channel pruning in speech separation models, outperforming existing methods. Notably, a model pruned with SepPrune can recover 85% of the performance of a pre-trained model (trained over hundreds of epochs) with only one epoch of fine-tuning, and achieves convergence 36$\times$ faster than training from scratch. Code is available at https://github.com/itsnotacie/SepPrune.

cs.SD↗

arXiv2025

Learning to Help in Multi-Class Settings

Deploying complex machine learning models on resource-constrained devices is challenging due to limited computational power, memory, and model retrainability. To address these limitations, a hybrid system can be established by augmenting the local model with a server-side model, where samples are selectively deferred by a rejector and then sent to the server for processing. The hybrid system enables efficient use of computational resources while minimizing the overhead associated with server usage. The recently proposed Learning to Help (L2H) model trains a server model given a fixed local (client) model, differing from the Learning to Defer (L2D) framework, which trains the client for a fixed (expert) server. In both L2D and L2H, the training includes learning a rejector at the client to determine when to query the server. In this work, we extend the L2H model from binary to multi-class classification problems and demonstrate its applicability in a number of different scenarios of practical interest in which access to the server may be limited by cost, availability, or policy. We derive a stage-switching surrogate loss function that is differentiable, convex, and consistent with the Bayes rule corresponding to the 0-1 loss for the L2H model. Experiments show that our proposed methods offer an efficient and practical solution for multi-class classification in resource-constrained environments.

cs.LG↗

arXiv2024

Renaissance of Literate Programming in the Era of LLMs: Enhancing LLM-Based Code Generation in Large-Scale Projects

Large Language Models (LLMs) have helped programmers increase efficiency through code generation, comprehension, and repair. However, their application to large-scale projects remains challenging due to complex interdependencies and the extensive size of modern codebases. Although Knuth's concept of Literate Programming (LP) combines code and natural language to convey logic and intent, its potential for enhancing relationships in large projects has not been fully explored. In this study, we introduce the idea of Interoperable LP (ILP), which leverages literate programming principles to enhance the development of both small-scale documents and large-scale projects with LLMs. We investigate how LLMs perform under ILP-style instructions for both document-oriented tasks and entire projects. Recognizing that many researchers rely on well-structured templates to guide LLMs, we propose a concise prompt engineering method to write LP documents so LLMs can better be involved in code generation. We also examine the capacity of various LLMs to generate Scheme and Python code on the RepoBench benchmark, illustrating the advantages of our approach. Our findings indicate that ILP with LLMs can enhance LLM-based code generation in large-scale project development.

cs.SE↗

arXiv2024

Generalizing End-To-End Autonomous Driving In Real-World Environments Using Zero-Shot LLMs

Traditional autonomous driving methods adopt a modular design, decomposing tasks into sub-tasks. In contrast, end-to-end autonomous driving directly outputs actions from raw sensor data, avoiding error accumulation. However, training an end-to-end model requires a comprehensive dataset; otherwise, the model exhibits poor generalization capabilities. Recently, large language models (LLMs) have been applied to enhance the generalization capabilities of end-to-end driving models. Most studies explore LLMs in an open-loop manner, where the output actions are compared to those of experts without direct feedback from the real world, while others examine closed-loop results only in simulations. This paper proposes an efficient architecture that integrates multimodal LLMs into end-to-end driving models operating in closed-loop settings in real-world environments. In our architecture, the LLM periodically processes raw sensor data to generate high-level driving instructions, effectively guiding the end-to-end model, even at a slower rate than the raw sensor data. This architecture relaxes the trade-off between the latency and inference quality of the LLM. It also allows us to choose from a wide variety of LLMs to improve high-level driving instructions and minimize fine-tuning costs. Consequently, our architecture reduces data collection requirements because the LLMs do not directly output actions; we only need to train a simple imitation learning model to output actions. In our experiments, the training data for the end-to-end model in a real-world environment consists of only simple obstacle configurations with one traffic cone, while the test environment is more complex and contains multiple obstacles placed in various positions. Experiments show that the proposed architecture enhances the generalization capabilities of the end-to-end model even without fine-tuning the LLM.

cs.RO↗

arXiv2024

EAPCR: A Universal Feature Extractor for Scientific Data without Explicit Feature Relation Patterns

Conventional methods, including Decision Tree (DT)-based methods, have been effective in scientific tasks, such as non-image medical diagnostics, system anomaly detection, and inorganic catalysis efficiency prediction. However, most deep-learning techniques have struggled to surpass or even match this level of success as traditional machine-learning methods. The primary reason is that these applications involve multi-source, heterogeneous data where features lack explicit relationships. This contrasts with image data, where pixels exhibit spatial relationships; textual data, where words have sequential dependencies; and graph data, where nodes are connected through established associations. The absence of explicit Feature Relation Patterns (FRPs) presents a significant challenge for deep learning techniques in scientific applications that are not image, text, and graph-based. In this paper, we introduce EAPCR, a universal feature extractor designed for data without explicit FRPs. Tested across various scientific tasks, EAPCR consistently outperforms traditional methods and bridges the gap where deep learning models fall short. To further demonstrate its robustness, we synthesize a dataset without explicit FRPs. While Kolmogorov-Arnold Network (KAN) and feature extractors like Convolutional Neural Networks (CNNs), Graph Convolutional Networks (GCNs), and Transformers struggle, EAPCR excels, demonstrating its robustness and superior performance in scientific tasks without FRPs.

cs.LG↗