arXiv subjects

Yifei Wang

Publications and source records attributed to Yifei Wang.

At least 127 records · Page 7Linked to original sources

arXiv2025

What is Wrong with Perplexity for Long-context Language Modeling?

Handling long-context inputs is crucial for large language models (LLMs) in tasks such as extended conversations, document summarization, and many-shot in-context learning. While recent approaches have extended the context windows of LLMs and employed perplexity (PPL) as a standard evaluation metric, PPL has proven unreliable for assessing long-context capabilities. The underlying cause of this limitation has remained unclear. In this work, we provide a comprehensive explanation for this issue. We find that PPL overlooks key tokens, which are essential for long-context understanding, by averaging across all tokens and thereby obscuring the true performance of models in long-context scenarios. To address this, we propose \textbf{LongPPL}, a novel metric that focuses on key tokens by employing a long-short context contrastive method to identify them. Our experiments demonstrate that LongPPL strongly correlates with performance on various long-context benchmarks (e.g., Pearson correlation of -0.96), significantly outperforming traditional PPL in predictive accuracy. Additionally, we introduce \textbf{LongCE} (Long-context Cross-Entropy) loss, a re-weighting strategy for fine-tuning that prioritizes key tokens, leading to consistent improvements across diverse benchmarks. In summary, these contributions offer deeper insights into the limitations of PPL and present effective solutions for accurately evaluating and enhancing the long-context capabilities of LLMs. Code is available at https://github.com/PKU-ML/LongPPL.

cs.CL↗

arXiv2025

Nitrogen-Vacancy Magnetometry of Edge Magnetism in WS2 Flakes

Two-dimensional (2D) magnets are of significant interest both as a platform for exploring novel fundamental physics and for their potential in spintronic and optoelectronic devices. Recent bulk magnetometry studies have indicated a weak ferromagnetic response in WS2, and theoretical predictions suggest edge-localized magnetization in flakes with partial hydrogenation. Here, we use room-temperature wide-field quantum diamond magnetometry to image pristine and Fe-implanted WS2 flakes of varying thicknesses (45-160 nm), exfoliated from bulk crystals and transferred to NV-doped diamond substrates. We observe direct evidence of edge-localized stray magnetic fields, which scale linearly with applied external magnetic field (4.4-220 mT), reaching up to 4.7 uT. The edge signal shows a limited dependence on the flake thickness, consistent with dipolar field decay and sensing geometry. Magnetic simulations using five alternative models favor the presence of edge magnetization aligned along an axis slightly tilted from the normal to the WS2 flake plane, consistent with spin canting in antiferromagnetically coupled edge states. Our findings establish WS2 as a promising platform for edge-controlled 2D spintronics.

cond-mat.mes-hall↗

arXiv2025

Beyond Isolated Capabilities: Bridging Long CoT Reasoning and Long-Context Understanding

Reasoning distillation has emerged as an effective approach to enhance the reasoning capabilities of smaller language models. However, the impact of large-scale reasoning distillation on other critical abilities, particularly in-context retrieval and reasoning, remains unexplored. This gap in understanding is particularly significant given the increasing importance of Retrieval-Augmented Generation (RAG) systems, where efficient acquisition and utilization of contextual information are paramount for generating reliable responses. Motivated by the need to understand how the extended long-CoT process influences long-context comprehension, we conduct a comprehensive investigation using a series of open-source models distilled from Deepseek-R1, renowned for its exceptional reasoning capabilities. Our study focuses on evaluating these models' performance in extracting and integrating relevant information from extended contexts through multi-document question and answering tasks. Through rigorous experimentation, we demonstrate that distilled reasoning patterns significantly improve long-context understanding. Our analysis reveals that distillation fosters greater long-context awareness by promoting more detailed and explicit reasoning processes during context analysis and information parsing. This advancement effectively mitigates the persistent "lost in the middle" issue that has hindered long-context models.

cs.CL↗

arXiv2025

STF: Spatial Temporal Fusion for Trajectory Prediction

Trajectory prediction is a challenging task that aims to predict the future trajectory of vehicles or pedestrians over a short time horizon based on their historical positions. The main reason is that the trajectory is a kind of complex data, including spatial and temporal information, which is crucial for accurate prediction. Intuitively, the more information the model can capture, the more precise the future trajectory can be predicted. However, previous works based on deep learning methods processed spatial and temporal information separately, leading to inadequate spatial information capture, which means they failed to capture the complete spatial information. Therefore, it is of significance to capture information more fully and effectively on vehicle interactions. In this study, we introduced an integrated 3D graph that incorporates both spatial and temporal edges. Based on this, we proposed the integrated 3D graph, which considers the cross-time interaction information. In specific, we design a Spatial-Temporal Fusion (STF) model including Multi-layer perceptions (MLP) and Graph Attention (GAT) to capture the spatial and temporal information historical trajectories simultaneously on the 3D graph. Our experiment on the ApolloScape Trajectory Datasets shows that the proposed STF outperforms several baseline methods, especially on the long-time-horizon trajectory prediction.

cs.CV↗

arXiv2025

Long-Short Alignment for Effective Long-Context Modeling in LLMs

Large language models (LLMs) have exhibited impressive performance and surprising emergent properties. However, their effectiveness remains limited by the fixed context window of the transformer architecture, posing challenges for long-context modeling. Among these challenges, length generalization -- the ability to generalize to sequences longer than those seen during training -- is a classical and fundamental problem. In this work, we propose a fresh perspective on length generalization, shifting the focus from the conventional emphasis on input features such as positional encodings or data structures to the output distribution of the model. Specifically, through case studies on synthetic tasks, we highlight the critical role of \textbf{long-short alignment} -- the consistency of output distributions across sequences of varying lengths. Extending this insight to natural language tasks, we propose a metric called Long-Short Misalignment to quantify this phenomenon, uncovering a strong correlation between the metric and length generalization performance. Building on these findings, we develop a regularization term that promotes long-short alignment during training. Extensive experiments validate the effectiveness of our approach, offering new insights for achieving more effective long-context modeling in LLMs. Code is available at https://github.com/PKU-ML/LongShortAlignment.

cs.CL↗

arXiv2025

A Signed Graph Approach to Understanding and Mitigating Oversmoothing in GNNs

Deep graph neural networks (GNNs) often suffer from oversmoothing, where node representations become overly homogeneous with increasing depth. While techniques like normalization, residual connections, and edge dropout have been proposed to mitigate oversmoothing, they are typically developed independently, with limited theoretical understanding of their underlying mechanisms. In this work, we present a unified theoretical perspective based on the framework of signed graphs, showing that many existing strategies implicitly introduce negative edges that alter message-passing to resist oversmoothing. However, we show that merely adding negative edges in an unstructured manner is insufficient-the asymptotic behavior of signed propagation depends critically on the strength and organization of positive and negative edges. To address this limitation, we leverage the theory of structural balance, which promotes stable, cluster-preserving dynamics by connecting similar nodes with positive edges and dissimilar ones with negative edges. We propose Structural Balanced Propagation (SBP), a plug-and-play method that assigns signed edges based on either labels or feature similarity to explicitly enhance structural balance in the constructed signed graphs. Experiments on nine benchmarks across both homophilic and heterophilic settings demonstrate that SBP consistently improves classification accuracy and mitigates oversmoothing, even at depths of up to 300 layers. Our results provide a principled explanation for prior oversmoothing remedies and introduce a new direction for signed message-passing design in deep GNNs.

cs.LG↗

arXiv2025

When More is Less: Understanding Chain-of-Thought Length in LLMs

Large Language Models (LLMs) employ Chain-of-Thought (CoT) reasoning to deconstruct complex problems. While longer CoTs are often presumed superior, this paper challenges that notion, arguing that longer is not always better. Drawing on combined evidence from real-world observations, controlled experiments, and theoretical analysis, we demonstrate that task accuracy typically follows an inverted U-shaped curve with CoT length, where performance initially improves but eventually decreases as the number of CoT steps increases. With controlled experiments, we further uncover the scaling behaviors of the optimal CoT length: it increases with task difficulty but decreases with model capability, exposing an inherent simplicity bias where more capable models favor shorter, more efficient CoT reasoning. This bias is also evident in Reinforcement Learning (RL) training, where models gravitate towards shorter CoTs as their accuracy improves. To have a deep understanding of these dynamics, we establish a simple theoretical model that formally proves these phenomena, including the optimal length's scaling laws and the emergence of simplicity bias during RL. Guided by this framework, we demonstrate significant practical benefits from training with optimally-lengthed CoTs and employing length-aware filtering at inference. These findings offer both a principled understanding of the "overthinking" phenomenon and multiple practical guidelines for CoT calibration, enabling LLMs to achieve optimal reasoning performance with adaptive CoTs tailored to task complexity and model capability.

cs.AI↗

arXiv2025

Uncertainty Unveiled: Can Exposure to More In-context Examples Mitigate Uncertainty for Large Language Models?

Recent advances in handling long sequences have facilitated the exploration of long-context in-context learning (ICL). While much of the existing research emphasizes performance improvements driven by additional in-context examples, the influence on the trustworthiness of generated responses remains underexplored. This paper addresses this gap by investigating how increased examples influence predictive uncertainty, an essential aspect in trustworthiness. We begin by systematically quantifying the uncertainty of ICL with varying shot counts, analyzing the impact of example quantity. Through uncertainty decomposition, we introduce a novel perspective on performance enhancement, with a focus on epistemic uncertainty (EU). Our results reveal that additional examples reduce total uncertainty in both simple and complex tasks by injecting task-specific knowledge, thereby diminishing EU and enhancing performance. For complex tasks, these advantages emerge only after addressing the increased noise and uncertainty associated with longer inputs. Finally, we explore the evolution of internal confidence across layers, unveiling the mechanisms driving the reduction in uncertainty.

cs.CL↗

arXiv2025

Beyond Matryoshka: Revisiting Sparse Coding for Adaptive Representation

Many large-scale systems rely on high-quality deep representations (embeddings) to facilitate tasks like retrieval, search, and generative modeling. Matryoshka Representation Learning (MRL) recently emerged as a solution for adaptive embedding lengths, but it requires full model retraining and suffers from noticeable performance degradations at short lengths. In this paper, we show that sparse coding offers a compelling alternative for achieving adaptive representation with minimal overhead and higher fidelity. We propose Contrastive Sparse Representation (CSR), a method that sparsifies pre-trained embeddings into a high-dimensional but selectively activated feature space. By leveraging lightweight autoencoding and task-aware contrastive objectives, CSR preserves semantic quality while allowing flexible, cost-effective inference at different sparsity levels. Extensive experiments on image, text, and multimodal benchmarks demonstrate that CSR consistently outperforms MRL in terms of both accuracy and retrieval speed-often by large margins-while also cutting training time to a fraction of that required by MRL. Our results establish sparse coding as a powerful paradigm for adaptive representation learning in real-world applications where efficiency and fidelity are both paramount. Code is available at https://github.com/neilwen987/CSR_Adaptive_Rep

cs.LG↗

arXiv2025

Advancing Drug Discovery with Enhanced Chemical Understanding via Asymmetric Contrastive Multimodal Learning

The versatility of multimodal deep learning holds tremendous promise for advancing scientific research and practical applications. As this field continues to evolve, the collective power of cross-modal analysis promises to drive transformative innovations, opening new frontiers in chemical understanding and drug discovery. Hence, we introduce Asymmetric Contrastive Multimodal Learning (ACML), a specifically designed approach to enhance molecular understanding and accelerate advancements in drug discovery. ACML harnesses the power of effective asymmetric contrastive learning to seamlessly transfer information from various chemical modalities to molecular graph representations. By combining pre-trained chemical unimodal encoders and a shallow-designed graph encoder with 5 layers, ACML facilitates the assimilation of coordinated chemical semantics from different modalities, leading to comprehensive representation learning with efficient training. We demonstrate the effectiveness of this framework through large-scale cross-modality retrieval and isomer discrimination tasks. Additionally, ACML enhances interpretability by revealing chemical semantics in graph presentations and bolsters the expressive power of graph neural networks, as evidenced by improved performance in molecular property prediction tasks from MoleculeNet and Therapeutics Data Commons (TDC). Ultimately, ACML exemplifies its potential to revolutionize molecular representational learning, offering deeper insights into the chemical semantics of diverse modalities and paving the way for groundbreaking advancements in chemical research and drug discovery.

cs.LG↗

arXiv2025

Rethinking Invariance in In-context Learning

In-Context Learning (ICL) has emerged as a pivotal capability of auto-regressive large language models, yet it is hindered by a notable sensitivity to the ordering of context examples regardless of their mutual independence. To address this issue, recent studies have introduced several variant algorithms of ICL that achieve permutation invariance. However, many of these do not exhibit comparable performance with the standard auto-regressive ICL algorithm. In this work, we identify two crucial elements in the design of an invariant ICL algorithm: information non-leakage and context interdependence, which are not simultaneously achieved by any of the existing methods. These investigations lead us to the proposed Invariant ICL (InvICL), a methodology designed to achieve invariance in ICL while ensuring the two properties. Empirically, our findings reveal that InvICL surpasses previous models, both invariant and non-invariant, in most benchmark datasets, showcasing superior generalization capabilities across varying input lengths. Code is available at https://github.com/PKU-ML/InvICL.

cs.CL↗

arXiv2025

Hybrid Privacy Policy-Code Consistency Check using Knowledge Graphs and LLMs

The increasing concern in user privacy misuse has accelerated research into checking consistencies between smartphone apps' declared privacy policies and their actual behaviors. Recent advances in Large Language Models (LLMs) have introduced promising techniques for semantic comparison, but these methods often suffer from low accuracies and expensive computational costs. To address this problem, this paper proposes a novel hybrid approach that integrates 1) knowledge graph-based deterministic checking to ensure higher accuracy, and 2) LLMs exclusively used for preliminary semantic analysis to save computational costs. Preliminary evaluation indicates this hybrid approach not only achieves 37.63% increase in precision and 23.13% increase F1-score but also consumes 93.5% less tokens and 87.3% shorter time.

cs.CR↗

arXiv2025

PSG-MAE: Robust Multitask Sleep Event Monitoring using Multichannel PSG Reconstruction and Inter-channel Contrastive Learning

Polysomnography (PSG) signals are essential for studying sleep processes and diagnosing sleep disorders. Analyzing PSG data through deep neural networks (DNNs) for automated sleep monitoring has become increasingly feasible. However, the limited availability of datasets for certain sleep events often leads to DNNs focusing on a single task with a single-sourced training dataset. As a result, these models struggle to transfer to new sleep events and lack robustness when applied to new datasets. To address these challenges, we propose PSG-MAE, a mask autoencoder (MAE) based pre-training framework. By performing self-supervised learning on a large volume of unlabeled PSG data, PSG-MAE develops a robust feature extraction network that can be broadly applied to various sleep event monitoring tasks. Unlike conventional MAEs, PSG-MAE generates complementary masks across PSG channels, integrates a multichannel signal reconstruction method, and employs a self-supervised inter-channel contrastive learning (ICCL) strategy. This approach enables the encoder to capture temporal features from each channel while simultaneously learning latent relationships between channels, thereby enhancing the utilization of multichannel information. Experimental results show that PSG-MAE effectively captures both temporal details and inter-channel information from PSG signals. When the encoder pre-trained through PSG-MAE is fine-tuned with downstream feature decomposition networks, it achieves an accuracy of 83.7% for sleep staging and 90.45% for detecting obstructive sleep apnea, which highlights the framework's robustness and broad applicability.

cs.LG↗

arXiv2025

Demystifying Issues, Causes and Solutions in LLM Open-Source Projects

With the advancements of Large Language Models (LLMs), an increasing number of open-source software projects are using LLMs as their core functional component. Although research and practice on LLMs are capturing considerable interest, no dedicated studies explored the challenges faced by practitioners of LLM open-source projects, the causes of these challenges, and potential solutions. To fill this research gap, we conducted an empirical study to understand the issues that practitioners encounter when developing and using LLM open-source software, the possible causes of these issues, and potential solutions. We collected all closed issues from 15 LLM open-source projects and labelled issues that met our requirements. We then randomly selected 994 issues from the labelled issues as the sample for data extraction and analysis to understand the prevalent issues, their underlying causes, and potential solutions. Our study results show that (1) Model Issue is the most common issue faced by practitioners, (2) Model Problem, Configuration and Connection Problem, and Feature and Method Problem are identified as the most frequent causes of the issues, and (3) Optimize Model is the predominant solution to the issues. Based on the study results, we provide implications for practitioners and researchers of LLM open-source projects.

cs.SE↗

arXiv2025

Cat-Eye Inspired Active-Passive-Composite Aperture-Shared Sub-Terahertz Meta-Imager for Non-Interactive Concealed Object Detection

Within the feline eye, a distinctive tapetum lucidum as a mirror resides posterior to the retina, reflecting the incident rays to simulate light source emission. This secondary emission property enables felines to be highly sensitive to light, possessing remarkable visual capabilities even in dark settings. Drawing inspiration from this natural phenomenon, we propose an active-passive-composite sub-terahertz meta-imager integrated with a bifocus metasurface, a high-sensitivity radiometer, and a low-power signal hidden radiation source. Benefiting from its aperture-shared advantage, this advanced fusion imaging system, enabled to be deployed by a simplified portable hardware platform, allows for the concurrent acquisition of active and passive electromagnetic properties to extend the target detection category and realize multi-mode fusion perception. Notably, it also enables the extraction of radiation and reflection characteristics without additional calibration modules. Experiments demonstrate the multi-target fusion imaging and localized information decoupling with the tailored field of view and emission energy. This compact and multi-mode fusion imaging system may have plenty of potential for airplane navigation positioning, abnormal monitoring, and non-interactive concealed security checks.

physics.optics↗

arXiv2025

A Novel Underwater Vehicle With Orientation Adjustable Thrusters: Design and Adaptive Tracking Control

Autonomous underwater vehicles (AUVs) are essential for marine exploration and research. However, conventional designs often struggle with limited maneuverability in complex, dynamic underwater environments. This paper introduces an innovative orientation-adjustable thruster AUV (OATAUV), equipped with a redundant vector thruster configuration that enables full six-degree-of-freedom (6-DOF) motion and composite maneuvers. To overcome challenges associated with uncertain model parameters and environmental disturbances, a novel feedforward adaptive model predictive controller (FFAMPC) is proposed to ensure robust trajectory tracking, which integrates real-time state feedback with adaptive parameter updates. Extensive experiments, including closed-loop tracking and composite motion tests in a laboratory pool, validate the enhanced performance of the OAT-AUV. The results demonstrate that the OAT-AUV's redundant vector thruster configuration enables 23.8% cost reduction relative to common vehicles, while the FF-AMPC controller achieves 68.6% trajectory tracking improvement compared to PID controllers. Uniquely, the system executes composite helical/spiral trajectories unattainable by similar vehicles.

cs.RO↗

arXiv2025

Magic teleportation with generalized lattice surgery

We propose a novel, distillation-free scheme for the fault-tolerant implementation of non-Clifford gates at the logical level, thereby completing the universal gate set. Our approach exploits generalized lattice surgery to integrate two quantum error-correcting (QEC) codes. Specifically, non-Clifford gates are executed transversally on one QEC code and then teleported to the main circuit via a logical-level joint measurement that connects two distinct QEC codes. In contrast to conventional magic state distillation (MSD) combined with gate teleportation, our method obviates the need for concatenating separate codes for distillation and logical qubits, thus reducing the total overhead from multiplicative to additive scaling. We illustrate our approach by explicitly demonstrating its implementation for a 3D color code interfaced with a surface code of the same code distance, and comment on its potential advantage over the conventional MSD-teleportation scheme.

quant-ph↗

arXiv2025

HEATS: A Hierarchical Framework for Efficient Autonomous Target Search with Mobile Manipulators

Utilizing robots for autonomous target search in complex and unknown environments can greatly improve the efficiency of search and rescue missions. However, existing methods have shown inadequate performance due to hardware platform limitations, inefficient viewpoint selection strategies, and conservative motion planning. In this work, we propose HEATS, which enhances the search capability of mobile manipulators in complex and unknown environments. We design a target viewpoint planner tailored to the strengths of mobile manipulators, ensuring efficient and comprehensive viewpoint planning. Supported by this, a whole-body motion planner integrates global path search with local IPC optimization, enabling the mobile manipulator to safely and agilely visit target viewpoints, significantly improving search performance. We present extensive simulated and real-world tests, in which our method demonstrates reduced search time, higher target search completeness, and lower movement cost compared to classic and state-of-the-art approaches. Our method will be open-sourced for community benefit.

cs.RO↗