Searcharxiv⌕ Search

arXiv subjects

Yifei Wang

Publications and source records attributed to Yifei Wang.

At least 55 records · Page 3Linked to original sources
arXiv2026

MARS: Efficient, Adaptive Co-Scheduling for Heterogeneous Agentic Systems

Large language models (LLMs) are increasingly deployed as the execution core of autonomous agents rather than as standalone text generators. Agentic workloads induce a temporal shift from single-turn inference to multi-turn LLM-tool loops, and a spatial shift from chat-scale, GPU-only execution to repository-scale, GPU-CPU co-located execution. Consequently, coordinating heterogeneous resource demands of agentic execution has emerged as a critical system challenge. We design and implement MARS, an efficient and adaptive co-scheduling system that globally coordinates heterogeneous agentic workloads under coupled GPU-CPU resource pressure. By establishing holistic visibility across GPU inference and CPU tool execution via a unified information stream, an external control plane in MARS decouples admission from execution to prevent heterogeneous resource oversubscription. An internal agent-centric scheduler further minimizes the end-to-end critical path by prioritizing latency-sensitive continuations and adaptively retaining KV cache state only when warm resumption yields a latency benefit. Our evaluations show that MARS reduces end-to-end latency by up to 5.94x while maintaining nearly maximal system throughput. We further integrate MARS as the serving backend for the OpenHands coding agent framework, demonstrating its real-world effectiveness by accelerating end-to-end task completion time by up to 1.87x. Our source code is publicly available at https://github.com/Afterglow231/MARS_preview .

cs.OS↗
arXiv2026

SurroundNEXO: Ego-Centric Metric Bridging for Spatially Consistent Geometry in Autonomous Driving

Modern autonomous driving depends on accurate metric 3D understanding for perception, reconstruction, and planning, which in turn requires reliable multi-camera depth prediction. However, the outward-facing nature of vehicle-mounted surround-view camera rigs inherently limits visual overlap across views, challenging the correspondence-based assumptions that underpin conventional multi-view geometry. To bridge this gap, we present SurroundNEXO, named after the Spanish word nexo for a geometric link, a low-overlap multi-camera metric depth framework that grounds cross-view reasoning in ego-centric geometry rather than dense visual correspondences. Instead of directly enforcing early global fusion, SurroundNEXO first assigns image tokens globally comparable ego-frame viewing directions through Ego-Ray Positional Encoding, then uses sparse LiDAR measurements as metric anchors to propagate absolute scale cues, and finally expands feature interaction progressively from view-local modeling to decomposed spatio-temporal reasoning and global integration. This design enables metric-scale depth prediction with improved spatial consistency across weakly overlapping cameras. Across low-overlap autonomous driving benchmarks, including NuScenes, Waymo and DDAD, SurroundNEXO reduces single-view error by 33.2%, improves cross-view consistency by 10.5%, and enhances metric reconstruction quality by 25.6% compared with SOTA methods. It further remains robust under extremely sparse depth prompts and exhibits strong zero-shot generalization to unseen camera layouts.

cs.CV↗
arXiv2026

Advancing Ligand-based Virtual Screening and Molecular Generation with Pretrained Molecular Embedding Distance

Molecular similarity plays a central role in ligand-based drug discovery, such as virtual screening, analog searching, and goal-directed molecular generation. However, traditional similarity measures, ranging from fingerprint-based Tanimoto coefficients to 3D shape overlays, are often computationally expensive at scale or rely on hand-crafted molecular descriptors. Meanwhile, many deep learning approaches to similarity-aware design still depend on similarity-specific supervision or costly data curation, limiting their generality across targets. In this work, we propose pretrained embedding distance (PED) as an effective alternative, computed directly from pretrained molecular models without task-specific training. Experimental results show that PED exhibits distinct correlations with traditional similarity metrics, and performs effectively in both ranking molecules for virtual screening and guiding molecular generation via reward design. These findings suggest that pretrained molecular embeddings capture rich structural information and can serve as a promising and scalable similarity measurement for modern AI-aided drug discovery.

cs.LG↗
arXiv2026

Drive-KD: Multi-Teacher Distillation for VLMs in Autonomous Driving

Autonomous driving is an important and safety-critical task, and recent advances in LLMs/VLMs have opened new possibilities for reasoning and planning in this domain. However, large models demand substantial GPU memory and exhibit high inference latency, while conventional supervised fine-tuning (SFT) often struggles to bridge the capability gaps of small models. To address these limitations, we propose Drive-KD, a framework that decomposes autonomous driving into a "perception-reasoning-planning" triad and transfers these capabilities via knowledge distillation. We identify layer-specific attention as the distillation signal to construct capability-specific single-teacher models that outperform baselines. Moreover, we unify these single-teacher settings into a multi-teacher distillation framework and introduce asymmetric gradient projection to mitigate cross-capability gradient conflicts. Extensive evaluations validate the generalization of our method across diverse model families and scales. Experiments show that our distilled InternVL3-1B model, with ~42 times less GPU memory and ~11.4 times higher throughput, achieves better overall performance than the pretrained 78B model from the same family on DriveBench, and surpasses GPT-5.1 on the planning dimension, providing insights toward efficient autonomous driving VLMs.

cs.AI↗
arXiv2026

Data-Driven Parameter Identification for Tumor Growth Models

Modeling tumor growth accurately is essential for understanding cancer progression and informing treatment strategies. To estimate the parameters in the tumor growth model described by a nonlinear PDE, we adopt Physics-Informed Neural Networks (PINNs) and DeepONet, which show advantages especially when the observation data is scarce and contains noise. With the help of real-life lab data, we have demonstrated the potential of applying deep learning tools to address data-driven modeling for tumor growth in biology.

math.AP↗
arXiv2026

Drive-P2D: A Progressive Perception-to-Decision Benchmark for VLMs in Autonomous Driving

Autonomous driving requires reliable perception and safe decision-making in complex scenarios. Recent vision-language models (VLMs) demonstrate reasoning and generalization abilities, opening new possibilities for autonomous driving; however, existing benchmarks often evaluate perception and decision-making separately, limit failure analysis with choice-only formats, or introduce evaluation bias through LLM-scored long-form outputs. To address these issues, we present Drive-P2D, a progressive perception-to-decision benchmark with 6,650 questions across Object, Scene, and Decision levels. Drive-P2D adopts a separated reasoning-and-answer protocol: final answers are scored objectively, while reasoning is analyzed to identify error modes exposed along the progressive perception-to-decision chain. We evaluate mainstream VLMs across all and high-risk scenarios, and further characterize the perception-to-decision capability boundary through correlation analysis and similar-scene robustness testing. Reasoning further exposes failure modes such as logical reasoning errors and semantic feature omissions, and we train a lightweight analyzer model to automate large-scale error-mode annotation of reasoning. Together, these designs provide practical insights for building safer and more reliable VLMs for real-world autonomous driving.

cs.AI↗
arXiv2026

Data Difficulty and the Generalization--Extrapolation Tradeoff in LLM Fine-Tuning

Data selection during supervised fine-tuning (SFT) can critically change the behavior of large language models (LLMs). Although existing work has studied the effect of selecting data based on heuristics such as perplexity, difficulty, or length, the reported findings are often inconsistent or context-dependent. In this work, we systematically study the role of data difficulty in fine-tuning from both empirical and theoretical perspectives, and find that there is no universally optimal difficulty level; rather, its effectiveness depends on the dataset size. We show that for a fixed data budget, there exists an optimal data difficulty for SFT, and that this optimal difficulty shifts toward harder data as the data budget increases. To explain this phenomenon, we conduct controlled synthetic experiments that reveal a simple underlying mechanism: the interplay between the (in-distribution) generalization gap and the extrapolation gap. We further support this mechanism through a theoretical analysis using PAC-Bayesian generalization bounds. Overall, our results clarify how data size and difficulty jointly affect the trade-off between generalization and extrapolation in SFT, providing guidance for difficulty-based data selection under certain model and data conditions.

cs.LG↗
arXiv2026

Certifying Quantum States with Uniform Measurements

Qubit-resolved operations and measurements are required for most current quantum information processing schemes. However, these operations can be experimentally costly due to the need for local addressing, demanding significant classical control. A more resource-efficient alternative to extract information is uniform measurement, where a site-independent rotation of qubits is performed before measuring in the computational basis. This operation can be performed in parallel, or globally, in atom- and ion-based platforms, reducing resource cost and increasing fidelity. In this work, we initiate the exploration of the utility of this operation in quantum information processing. In particular, we demonstrate that uniform measurements can certify certain graph states, a family of highly entangled and broadly useful quantum states. We provide a sample-efficient certification algorithm with a proved performance guarantee, together with an experimental scheme based on analog-mode Rydberg atom arrays. Uniform measurements, therefore, allow direct and efficient characterization of quantum states on quantum platforms in a hitherto unexplored manner.

quant-ph↗
arXiv2026

Optimus: Elastic Decoding for Efficient Diffusion LLM Serving

Large language model (LLM) serving is fundamentally limited by inefficient hardware utilization. Autoregressive (AR) decoding underutilizes GPUs due to its strictly sequential execution, while diffusion LLMs (DLLMs) improve throughput by decoding multiple tokens per iteration. However, fixed block-size diffusion decoding exhibits strong load sensitivity: large blocks exploit idle GPU resources under low load, but saturate early and incur substantial redundant computation under high load. As a result, throughput gains vanish beyond saturation, and no single decoding granularity performs well across dynamic serving workloads. We present Optimus, a serving system that enables elastic decoding for diffusion LLMs by dynamically adapting decoding granularity to runtime load. The key idea is to treat decoding granularity as a runtime control variable, balancing GPU utilization and token efficiency. Optimus combines chunked decoding, which enables fine-grained execution without retraining, with saturation-aware scheduling, a closed-loop mechanism that selects chunk sizes based on runtime conditions. Together with system-level optimizations and customized attention kernels, Optimus achieves significant performance improvements while preserving model accuracy. Experiments show that Optimus delivers up to 6.1x throughput improvement over AR decoding and 4.3x improvement over fixed-block diffusion LLM, while maintaining stable performance across diverse load regimes and improving end-to-end serving capacity under latency constraints. The source code is available at https://github.com/dubcyfor3/Optimus.

cs.DC↗
arXiv2026

QUEST: Training Frontier Deep Research Agents with Fully Synthetic Tasks

Deep research agents extend the role of search engines from retrieving keyword-matched pages to synthesizing knowledge, fundamentally changing how humans interact with information. However, frontier systems remain proprietary, while existing open agents often generalize poorly across different task types, leaving unclear how to train a broadly capable deep research agent. We release QUEST, a family of open models (ranging from 2B to 35B) that serve as general-purpose deep research agents designed to handle a wide range of long-horizon search tasks, with strong capabilities in fact seeking, citation grounding, and report synthesis. To build QUEST, we propose an effective training recipe combining mid-training, supervised fine-tuning, and reinforcement learning. Central to this recipe is a curated data synthesis pipeline based on unified rubric trees, which applies to different task types and enables synthesizing training data with verifiable rewards without human annotation. In addition, QUEST incorporates a built-in context management mechanism that enables effective long-horizon reasoning and knowledge synthesis. Using only 8K synthesized tasks, QUEST approaches or even surpasses frontier closed-source agents across eight deep research benchmarks spanning diverse task types, and achieves the best overall performance among recent open-weight agents. We released everything: models, data, and training scripts.

cs.CL↗
arXiv2026

SSL4RL: Revisiting Self-supervised Learning as Intrinsic Reward for Visual-Language Reasoning

Vision-language models (VLMs) have shown remarkable abilities by integrating large language models with visual inputs. However, they often fail to utilize visual evidence adequately, either depending on linguistic priors in vision-centric tasks or resorting to textual shortcuts during reasoning. Although reinforcement learning (RL) can align models with desired behaviors, its application to VLMs has been hindered by the lack of scalable and reliable reward mechanisms. To overcome this challenge, we propose SSL4RL, a novel framework that leverages self-supervised learning (SSL) tasks as a source of verifiable rewards for RL-based fine-tuning. Our approach reformulates SSL objectives-such as predicting image rotation or reconstructing masked patches-into dense, automatic reward signals, eliminating the need for human preference data or unreliable AI evaluators. Experiments show that SSL4RL substantially improves performance on both vision-centric and vision-language reasoning benchmarks. Furthermore, through systematic ablations, we identify key factors-such as task difficulty, model scale, and semantic alignment with the target domain-that influence the effectiveness of SSL4RL tasks, offering new design principles for future work. We also demonstrate the framework's generality by applying it to graph learning, where it yields significant gains. SSL4RL establishes a versatile and effective paradigm for aligning multimodal models using verifiable, self-supervised objectives.

cs.CV↗
arXiv2026

Taming Outlier Tokens in Diffusion Transformers

We study outlier tokens in Diffusion Transformers (DiTs) for image generation. Prior work has shown that Vision Transformers (ViTs) can produce a small number of high-norm tokens that attract disproportionate attention while carrying limited local information, but their role in generative models remains underexplored. We show that this phenomenon appears in both the encoder and denoiser of modern Representation Autoencoder (RAE)-DiT pipelines: pretrained ViT encoders can produce outlier representations, and DiTs themselves can develop internal outlier tokens, especially in intermediate layers. Moreover, simply masking high-norm tokens does not improve performance, indicating that the problem is not only caused by a few extreme values, but is more closely related to corrupted local patch semantics. To address this issue, we introduce Dual-Stage Registers (DSR), a register-based intervention for both components: trained registers when available, recursive test-time registers otherwise, and diffusion registers for the denoiser. Across ImageNet and large-scale text-to-image generation, these interventions consistently reduce outlier artifacts and improve generation quality. Our results highlight outlier-token control as an important ingredient in building stronger DiTs.

cs.CV↗
arXiv2026

Resolving spurious topological entanglement entropy in stabilizer codes

Topological entanglement entropy (TEE) is a key diagnostic of long-range entanglement in two-dimensional gapped phases of matter, but it can suffer from spurious contributions that overestimate the total quantum dimension of the underlying topological order. In this work, we identify the microscopic origin of spurious TEE and introduce a concave partition for computing the Levin-Wen TEE of translation-invariant stabilizer codes of prime-dimensional qudits. We rigorously prove that this prescription is free of spurious contributions. As a complementary probe, we study bivariate bicycle codes on a bipartite cylinder and show that the entanglement entropy depends sensitively on the cylinder circumference, revealing topological frustration of the underlying anyons.

quant-ph↗
arXiv2026

Non-Abelian Particle-Loop, Fracton, and Planon Condensation in Cage-Net Models

We present a framework for non-Abelian p-loop, fracton, and planon condensation in 3+1 dimensions by constructing extended cage-net fracton models using decoupled layers of the Hu-Geer-Wu (HGW) string-net model. These cage-net models extend the conventional cage-net models based on the Levin-Wen (LW) string-net model in the sense that they inherit the tail degrees of freedom of the HGW models, which are essential for completely describing the internal spaces of quasiparticles. This approach allows us to explicitly derive the quasiparticle spectra of the cage-net models by projecting those of the parent 2D HGW layers. Utilizing this framework, we can condense the p-loops formed by non-Abelian anyons within a fracton phase. Specifically, we construct the condensation projector for $(σ\barσ, 1)$-loops within the extended Ising Cage-Net (ICN) model. We demonstrate that condensing these non-Abelian loops drives a phase transition that maps the ICN model to the X-cube (XC) model defined on a truncated cubic lattice, a process that explicitly reveals the splitting of non-Abelian planons into distinct sub-dimensional excitations. Furthermore, our framework extends to the condensation of fractons and planons: we demonstrate that in the ICN model fracton condensation drives the decoupling of the 3D fracton order back into isolated 2D topological order layers, while planon condensation collapses the system entirely into a trivial phase. Our results establish a concrete Hamiltonian mechanism for phase transitions between distinct fracton orders and provide a generalizable method for analyzing the evolution of sub-dimensional excitations.

cond-mat.str-el↗
arXiv2026

Beyond Cross-Modal Alignment: Measuring and Leveraging Modality Gap in Vision-Language Models

The success of vision-language models is primarily attributed to effective alignment across modalities such as vision and language. However, modality gaps persist in existing alignment algorithms and appear necessary for human perception as evidenced by modality-specific phenomena like visual texture and linguistic tone. These observations motivate us to computationally measure and leverage modality gaps to improve downstream tasks. We first introduce the Modality Dominance Score (MDS), which attributes multimodal features to specific modalities by categorizing them into three classes: vision-dominant features, language-dominant features, and cross-modal features. We then propose automatic interpretability metrics to evaluate these modality-specific features in a scalable manner. Finally, we demonstrate that the training-free model editing enhances multiple downstream tasks, including mitigating bias in gender classification, generating cross-modal adversarial examples, and enabling modality-specific control in text-to-image generation. Combined with task-agnostic interpretability tools, our work offers insights for systematic analysis and lightweight editing of multimodal models.

cs.CV↗
arXiv2026

R2Code: A Self-Reflective LLM Framework for Requirements-to-Code Traceability

Accurate requirement-to-code traceability is crucial for software maintenance. However, existing IR- and embedding-based methods are heavily dependent on lexical similarity, often yielding incomplete or inconsistent links across projects and languages and incurring high cost from long-context retrieval and prompting. This paper presents R2Code, an LLM-based semantic traceability framework designed to improve trace link accuracy while reducing inference cost. R2Code integrates three components: 1) a decomposition-enhanced Bidirectional Alignment Network (BAN) that aligns four-layer requirement semantics with corresponding code structures to support cross-level semantic matching; 2) a Self-Reflective Consistency Verification (SRCV) module that conducts explanation-guided consistency checking to calibrate link reliability; and 3) a Dynamic Context-Adaptive Retrieval (DCAR) mechanism that adjusts retrieval granularity and filters contexts using semantic-overlap weighting for efficient context utilization. Experiments on five public datasets spanning multiple domains and two programming languages demonstrate that R2Code consistently outperforms the strongest baselines, achieving an average F1 gain of 7.4%, while reducing token consumption by up to 41.7% through adaptive context control.

cs.SE↗
arXiv2026

Lightweight Real-Time ALADIN for Distributed Optimization

This paper presents a real-time computational framework for multi-node distributed optimization by extending the Augmented Lagrangian Alternating Direction Inexact Newton (ALADIN) algorithm. Our approach integrates adjoint sequential quadratic programming (SQP) techniques to enable efficient approximation of Jacobian information within the ALADIN embedded quadratic program, thereby reducing communication overhead. Furthermore, to decrease computational complexity, we design an event-triggered update strategy that avoids updating Hessian and Jacobian matrices at every iteration. The proposed method achieves local convergence and enhanced communication efficiency, making it well suited for time-critical applications. Numerical experiments demonstrate that our approach achieves competitive performance while exhibiting superior computational efficiency in real-time scenarios, validating its practical applicability for time-sensitive distributed optimization challenges.

math.OC↗
arXiv2026

Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges

Reinforcement Learning from Human Feedback (RLHF) and related alignment paradigms have become central to steering large language models (LLMs) and multimodal large language models (MLLMs) toward human-preferred behaviors. However, these approaches introduce a systemic vulnerability: reward hacking, where models exploit imperfections in learned reward signals to maximize proxy objectives without fulfilling true task intent. As models scale and optimization intensifies, such exploitation manifests as verbosity bias, sycophancy, hallucinated justification, benchmark overfitting, and, in multimodal settings, perception--reasoning decoupling and evaluator manipulation. Recent evidence further suggests that seemingly benign shortcut behaviors can generalize into broader forms of misalignment, including deception and strategic gaming of oversight mechanisms. In this survey, we propose the Proxy Compression Hypothesis (PCH) as a unifying framework for understanding reward hacking. We formalize reward hacking as an emergent consequence of optimizing expressive policies against compressed reward representations of high-dimensional human objectives. Under this view, reward hacking arises from the interaction of objective compression, optimization amplification, and evaluator--policy co-adaptation. This perspective unifies empirical phenomena across RLHF, RLAIF, and RLVR regimes, and explains how local shortcut learning can generalize into broader forms of misalignment, including deception and strategic manipulation of oversight mechanisms. We further organize detection and mitigation strategies according to how they intervene on compression, amplification, or co-adaptation dynamics. By framing reward hacking as a structural instability of proxy-based alignment under scale, we highlight open challenges in scalable oversight, multimodal grounding, and agentic autonomy.

cs.LG↗