arXiv subjects

Ziyue Li

Publications and source records attributed to Ziyue Li.

At least 37 records · Page 2Linked to original sources

arXiv2026

Grokking in LLM Pretraining? Monitor Memorization-to-Generalization without Test

This paper presents the first study of grokking in practical LLM pretraining. Specifically, we investigate when an LLM memorizes the training data, when its generalization on downstream tasks starts to improve, and what happens if there is a lag between the two. Unlike existing works studying when a small model generalizes to limited and specified tasks during thousands epochs' training on algorithmic data, we focus on a practical setting for LLMs, i.e., one-epoch pretraining of next-token prediction on a cross-domain, large-scale corpus, and generalization on diverse benchmark tasks covering math/commonsense reasoning, code generation, and domain-specific retrieval. Our study, for the first time, verifies that grokking still emerges in pretraining mixture-of-experts (MoE) LLMs, though different local data groups may enter their grokking stages asynchronously due to the heterogeneity of their distributions and attributions to others. To find a mechanistic interpretation of this local grokking, we investigate the dynamics of training data's pathways (i.e., expert choices across layers in MoE). Our primary discovery is that the pathways evolve from random, non-smooth across layers, instance-specific to more structured and transferable across samples, despite the converged pretraining loss. This depicts a transition from memorization to generalization. Two novel metrics are developed to quantify these patterns: one computes the pathway similarity between samples, while the other measures the consistency of aggregated experts between subsequent layers for each sample. These training data based metrics induce zero cost but can faithfully track and monitor the generalization of LLMs on downstream tasks, which, in conventional settings, requires costly instruction tuning and benchmark evaluation.

cs.LG↗

arXiv2026

Review of Passenger Flow Modelling Approaches Based on a Bibliometric Analysis

This paper presents a bibliometric analysis of the field of short-term passenger flow forecasting within local public transit, covering 814 publications that span from 1984 to 2024. In addition to common bibliometric analysis tools, a variant of a citation network was developed, and topic modelling was conducted. The analysis reveals that research activity exhibited sporadic patterns prior to 2008, followed by a marked acceleration, characterised by a shift from conventional statistical and machine learning methodologies (e.g., ARIMA, SVM, and basic neural networks) to specialised deep learning architectures. Based on this insight, a connection to more general fields such as machine learning and time series modelling was established. In addition to modelling, spatial, linguistic, and modal biases were identified and findings from existing secondary literature were validated and quantified. This revealed existing gaps, such as constrained data fusion, open (multivariate) data, and underappreciated challenges related to model interpretability, cost-efficiency, and a balance between algorithmic performance and practical deployment considerations. In connection with the superordinate fields, the growth in relevance of foundation models is also noteworthy.

cs.DL↗

arXiv2025

TPTU: Large Language Model-based AI Agents for Task Planning and Tool Usage

With recent advancements in natural language processing, Large Language Models (LLMs) have emerged as powerful tools for various real-world applications. Despite their prowess, the intrinsic generative abilities of LLMs may prove insufficient for handling complex tasks which necessitate a combination of task planning and the usage of external tools. In this paper, we first propose a structured framework tailored for LLM-based AI Agents and discuss the crucial capabilities necessary for tackling intricate problems. Within this framework, we design two distinct types of agents (i.e., one-step agent and sequential agent) to execute the inference process. Subsequently, we instantiate the framework using various LLMs and evaluate their Task Planning and Tool Usage (TPTU) abilities on typical tasks. By highlighting key findings and challenges, our goal is to provide a helpful resource for researchers and practitioners to leverage the power of LLMs in their AI applications. Our study emphasizes the substantial potential of these models, while also identifying areas that need more investigation and improvement.

cs.AI↗

arXiv2025

Relational Mediators: LLM Chatbots as Boundary Objects in Psychotherapy

As large language models (LLMs) are embedded into mental health technologies, they are often framed either as tools assisting therapists or autonomous therapeutic systems. Such perspectives overlook their potential to mediate relational complexities in therapy, particularly for systemically marginalized clients. Drawing on in-depth interviews with 12 therapists and 12 marginalized clients in China, including LGBTQ+ individuals or those from other marginalized backgrounds, we identify enduring relational challenges: difficulties building trust amid institutional barriers, the burden clients carry in educating therapists about marginalized identities, and challenges sustaining authentic self-disclosure across therapy and daily life. We argue that addressing these challenges requires AI systems capable of actively mediating underlying knowledge gaps, power asymmetries, and contextual disconnects. To this end, we propose the Dynamic Boundary Mediation Framework, which reconceptualizes LLM-enhanced systems as adaptive boundary objects that shift mediating roles across therapeutic stages. The framework delineates three forms of mediation: Epistemic (reducing knowledge asymmetries), Relational (rebalancing power dynamics), and Contextual (bridging therapy-life discontinuities). This framework offers a pathway toward designing relationally accountable AI systems that center the lived realities of marginalized users and more effectively support therapeutic relationships.

cs.HC↗

arXiv2025

Language-Instructed Reasoning for Group Activity Detection via Multimodal Large Language Model

Group activity detection (GAD) aims to simultaneously identify group members and categorize their collective activities within video sequences. Existing deep learning-based methods develop specialized architectures (e.g., transformer networks) to model the dynamics of individual roles and semantic dependencies between individuals and groups. However, they rely solely on implicit pattern recognition from visual features and struggle with contextual reasoning and explainability. In this work, we propose LIR-GAD, a novel framework of language-instructed reasoning for GAD via Multimodal Large Language Model (MLLM). Our approach expand the original vocabulary of MLLM by introducing an activity-level token and multiple cluster-specific tokens. We process video frames alongside two specially designed tokens and language instructions, which are then integrated into the MLLM. The pretrained commonsense knowledge embedded in the MLLM enables the token and tokens to effectively capture the semantic information of collective activities and learn distinct representational features of different groups, respectively. Also, we introduce a multi-label classification loss to further enhance the token's ability to learn discriminative semantic representations. Then, we design a Multimodal Dual-Alignment Fusion (MDAF) module that integrates MLLM's hidden embeddings corresponding to the designed tokens with visual features, significantly enhancing the performance of GAD. Both quantitative and qualitative experiments demonstrate the superior performance of our proposed method in GAD taks.

cs.CV↗

arXiv2025

Spatially Dependent Sampling of Component Failures for Power System Preventive Control Against Hurricane

Preventive control is a crucial strategy for power system operation against impending natural hazards, and its effectiveness fundamentally relies on the realism of scenario generation. While most existing studies employ sequential Monte Carlo simulation and assume independent sampling of component failures, this oversimplification neglects the spatial correlations induced by meteorological factors such as hurricanes. In this paper, we identify and address the gap in modeling spatial dependence among component failures under extreme weather. We analyze how the mean, variance, and correlation structure of weather intensity random variables influence the correlation of component failures. To fill this gap, we propose a spatially dependent sampling method that enables joint sampling of multiple component failures by generating correlated meteorological intensity random variables. Comparative studies show that our approach captures long-tailed scenarios and reveals more extreme events than conventional methods. Furthermore, we evaluate the impact of scenario selection on preventive control performance. Our key findings are: (1) Strong spatial correlations in uncertain weather intensity consistently lead to interdependent component failures, regardless of mean value level; (2) The proposed method uncovers more high-severity scenarios that are missed by independent sampling; (3) Preventive control requires balancing load curtailment and over-generation costs under different scenario severities; (4) Ignoring failure correlations results in underestimating risk from high-severity events, undermining the robustness of preventive control strategies.

eess.SY↗

arXiv2025

Routing Manifold Alignment Improves Generalization of Mixture-of-Experts LLMs

Sparse Mixture-of-Experts (MoE) have been widely adopted in recent large language models since it can efficiently scale up the model capability without increasing the inference cost. However, evaluations on broad downstream tasks reveal a consistent suboptimality of the routers in existing MoE LLMs, which results in a severe performance gap (e.g., 10-20% in accuracy) to the optimal routing. In this paper, we show that aligning the manifold of routing weights with that of task embedding can effectively reduce the gap and improve MoE LLMs' generalization performance. Our method, "Routing Manifold Alignment (RoMA)", introduces an additional manifold regularization term in the post-training objective and only requires lightweight finetuning of routers (with other parameters frozen). Specifically, the regularization encourages the routing weights of each sample to be close to those of its successful neighbors (whose routing weights lead to correct answers) in a task embedding space. Consequently, samples targeting similar tasks will share similar expert choices across layers. Building such bindings between tasks and experts over different samples is essential to achieve better generalization. Moreover, RoMA demonstrates the advantage of unifying the task understanding (by embedding models) with solution generation (by MoE LLMs). In experiments, we finetune routers in OLMoE, DeepSeekMoE, and Qwen3-MoE using RoMA. Evaluations on diverse benchmarks and extensive comparisons with baselines show the substantial improvement brought by RoMA.

cs.LG↗

arXiv2025

ColorBench: Can VLMs See and Understand the Colorful World? A Comprehensive Benchmark for Color Perception, Reasoning, and Robustness

Color plays an important role in human perception and usually provides critical clues in visual reasoning. However, it is unclear whether and how vision-language models (VLMs) can perceive, understand, and leverage color as humans. This paper introduces ColorBench, an innovative benchmark meticulously crafted to assess the capabilities of VLMs in color understanding, including color perception, reasoning, and robustness. By curating a suite of diverse test scenarios, with grounding in real applications, ColorBench evaluates how these models perceive colors, infer meanings from color-based cues, and maintain consistent performance under varying color transformations. Through an extensive evaluation of 32 VLMs with varying language models and vision encoders, our paper reveals some undiscovered findings: (i) The scaling law (larger models are better) still holds on ColorBench, while the language model plays a more important role than the vision encoder. (ii) However, the performance gaps across models are relatively small, indicating that color understanding has been largely neglected by existing VLMs. (iii) CoT reasoning improves color understanding accuracies and robustness, though they are vision-centric tasks. (iv) Color clues are indeed leveraged by VLMs on ColorBench but they can also mislead models in some tasks. These findings highlight the critical limitations of current VLMs and underscore the need to enhance color comprehension. Our ColorBenchcan serve as a foundational tool for advancing the study of human-level color understanding of multimodal AI.

cs.CV↗

arXiv2025

Deep Learning to Identify the Spatio-Temporal Cascading Effects of Train Delays in a High-Density Network

The operational efficiency of railway networks, a cornerstone of modern economies, is persistently undermined by the cascading effects of train delays. Accurately forecasting this delay propagation is a critical challenge for real-time traffic management. While recent research has leveraged Graph Neural Networks (GNNs) to model the network structure of railways, a significant gap remains in developing frameworks that provide multi-step autoregressive forecasts at a network-wide scale, while simultaneously offering the live, interpretable explanations needed for decision support. This paper addresses this gap by developing and evaluating a novel XGeoAI framework for live, explainable, multi-step train delay forecasting. The core of this work is a two-stage, autoregressive Graph Attention Network (GAT) model, trained on a real-world dataset covering over 40% of the Dutch railway network. The model represents the system as a spatio-temporal graph of operational events (arrivals and departures) and is enriched with granular features, including platform and station congestion. To test its viability for live deployment, the model is rigorously evaluated using a sequential, k-step-ahead forecasting protocol that simulates real-world conditions where prediction errors can compound. The results demonstrate that while the proposed GATv2 model is challenged on pure error metrics (MAE) by a simpler Persistence baseline, it achieves consistently higher precision in classifying delay events -- a crucial advantage for a reliable decision support tool.

cs.LG↗

arXiv2025

GuiLoMo: Allocating Expert Number and Rank for LoRA-MoE via Bilevel Optimization with GuidedSelection Vectors

Parameter-efficient fine-tuning (PEFT) methods, particularly Low-Rank Adaptation (LoRA), offer an efficient way to adapt large language models with reduced computational costs. However, their performance is limited by the small number of trainable parameters. Recent work combines LoRA with the Mixture-of-Experts (MoE), i.e., LoRA-MoE, to enhance capacity, but two limitations remain in hindering the full exploitation of its potential: 1) the influence of downstream tasks when assigning expert numbers, and 2) the uniform rank assignment across all LoRA experts, which restricts representational diversity. To mitigate these gaps, we propose GuiLoMo, a fine-grained layer-wise expert numbers and ranks allocation strategy with GuidedSelection Vectors (GSVs). GSVs are learned via a prior bilevel optimization process to capture both model- and task-specific needs, and are then used to allocate optimal expert numbers and ranks. Experiments on three backbone models across diverse benchmarks show that GuiLoMo consistently achieves superior or comparable performance to all baselines. Further analysis offers key insights into how expert numbers and ranks vary across layers and tasks, highlighting the benefits of adaptive expert configuration. Our code is available at https://github.com/Liar406/Gui-LoMo.git.

cs.CL↗

arXiv2025

TreeReview: A Dynamic Tree of Questions Framework for Deep and Efficient LLM-based Scientific Peer Review

While Large Language Models (LLMs) have shown significant potential in assisting peer review, current methods often struggle to generate thorough and insightful reviews while maintaining efficiency. In this paper, we propose TreeReview, a novel framework that models paper review as a hierarchical and bidirectional question-answering process. TreeReview first constructs a tree of review questions by recursively decomposing high-level questions into fine-grained sub-questions and then resolves the question tree by iteratively aggregating answers from leaf to root to get the final review. Crucially, we incorporate a dynamic question expansion mechanism to enable deeper probing by generating follow-up questions when needed. We construct a benchmark derived from ICLR and NeurIPS venues to evaluate our method on full review generation and actionable feedback comments generation tasks. Experimental results of both LLM-based and human evaluation show that TreeReview outperforms strong baselines in providing comprehensive, in-depth, and expert-aligned review feedback, while reducing LLM token usage by up to 80% compared to computationally intensive approaches. Our code and benchmark dataset are available at https://github.com/YuanChang98/tree-review.

cs.CL↗

arXiv2025

MedualTime: A Dual-Adapter Language Model for Medical Time Series-Text Multimodal Learning

The recent rapid advancements in language models (LMs) have garnered attention in medical time series-text multimodal learning. However, existing contrastive learning-based and prompt-based LM approaches tend to be biased, often assigning a primary role to time series modality while treating text modality as secondary. We classify these approaches under a temporal-primary paradigm, which may overlook the unique and critical task-relevant information embedded in text modality like clinical reports, thus failing to fully leverage mutual benefits and complementarity of different modalities. To fill this gap, we propose a novel textual-temporal multimodal learning paradigm that enables either modality to serve as the primary while being enhanced by the other, thereby effectively capturing modality-specific information and fostering cross-modal interaction. In specific, we design MedualTime, a language model composed of dual adapters to implement temporal-primary and textual-primary modeling simultaneously. Within each adapter, lightweight adaptation tokens are injected into the top layers of LM to encourage high-level modality fusion. The shared LM pipeline by dual adapters not only achieves adapter alignment but also enables efficient fine-tuning, reducing computational resources. Empirically, MedualTime demonstrates superior performance on medical data, achieving notable improvements of 8% accuracy and 12% F1 in supervised settings. Furthermore, MedualTime's transferability is validated by few-shot label transfer experiments from coarse-grained to fine-grained medical data. https://github.com/start2020/MedualTime

cs.LG↗

arXiv2025

Fishing for Answers: Exploring One-shot vs. Iterative Retrieval Strategies for Retrieval Augmented Generation

Retrieval-Augmented Generation (RAG) based on Large Language Models (LLMs) is a powerful solution to understand and query the industry's closed-source documents. However, basic RAG often struggles with complex QA tasks in legal and regulatory domains, particularly when dealing with numerous government documents. The top-$k$ strategy frequently misses golden chunks, leading to incomplete or inaccurate answers. To address these retrieval bottlenecks, we explore two strategies to improve evidence coverage and answer quality. The first is a One-SHOT retrieval method that adaptively selects chunks based on a token budget, allowing as much relevant content as possible to be included within the model's context window. Additionally, we design modules to further filter and refine the chunks. The second is an iterative retrieval strategy built on a Reasoning Agentic RAG framework, where a reasoning LLM dynamically issues search queries, evaluates retrieved results, and progressively refines the context over multiple turns. We identify query drift and retrieval laziness issues and further design two modules to tackle them. Through extensive experiments on a dataset of government documents, we aim to offer practical insights and guidance for real-world applications in legal and regulatory domains.

cs.IR↗

arXiv2025

HiPlan: Hierarchical Planning for LLM-Based Agents with Adaptive Global-Local Guidance

Large language model (LLM)-based agents have demonstrated remarkable capabilities in decision-making tasks, but struggle significantly with complex, long-horizon planning scenarios. This arises from their lack of macroscopic guidance, causing disorientation and failures in complex tasks, as well as insufficient continuous oversight during execution, rendering them unresponsive to environmental changes and prone to deviations. To tackle these challenges, we introduce HiPlan, a hierarchical planning framework that provides adaptive global-local guidance to boost LLM-based agents'decision-making. HiPlan decomposes complex tasks into milestone action guides for general direction and step-wise hints for detailed actions. During the offline phase, we construct a milestone library from expert demonstrations, enabling structured experience reuse by retrieving semantically similar tasks and milestones. In the execution phase, trajectory segments from past milestones are dynamically adapted to generate step-wise hints that align current observations with the milestone objectives, bridging gaps and correcting deviations. Extensive experiments across two challenging benchmarks demonstrate that HiPlan substantially outperforms strong baselines, and ablation studies validate the complementary benefits of its hierarchical components.

cs.CL↗

arXiv2025

MedSpaformer: a Transferable Transformer with Multi-granularity Token Sparsification for Medical Time Series Classification

Accurate medical time series (MedTS) classification is essential for effective clinical diagnosis, yet remains challenging due to complex multi-channel temporal dependencies, information redundancy, and label scarcity. While transformer-based models have shown promise in time series analysis, most are designed for forecasting tasks and fail to fully exploit the unique characteristics of MedTS. In this paper, we introduce MedSpaformer, a transformer-based framework tailored for MedTS classification. It incorporates a sparse token-based dual-attention mechanism that enables global context modeling and token sparsification, allowing dynamic feature refinement by focusing on informative tokens while reducing redundancy. This mechanism is integrated into a multi-granularity cross-channel encoding scheme to capture intra- and inter-granularity temporal dependencies and inter-channel correlations, enabling progressive refinement of task-relevant patterns in medical signals. The sparsification design allows our model to flexibly accommodate inputs with variable lengths and channel dimensions. We also introduce an adaptive label encoder to extract label semantics and address cross-dataset label space misalignment. Together, these components enhance the model's transferability across heterogeneous medical datasets, which helps alleviate the challenge of label scarcity. Our model outperforms 13 baselines across 7 medical datasets under supervised learning. It also excels in few-shot learning and demonstrates zero-shot capability in both in-domain and cross-domain diagnostics. These results highlight MedSpaformer's robustness and its potential as a unified solution for MedTS classification across diverse settings.

cs.LG↗

arXiv2025

SAMITE: Position Prompted SAM2 with Calibrated Memory for Visual Object Tracking

Visual Object Tracking (VOT) is widely used in applications like autonomous driving to continuously track targets in videos. Existing methods can be roughly categorized into template matching and autoregressive methods, where the former usually neglects the temporal dependencies across frames and the latter tends to get biased towards the object categories during training, showing weak generalizability to unseen classes. To address these issues, some methods propose to adapt the video foundation model SAM2 for VOT, where the tracking results of each frame would be encoded as memory for conditioning the rest of frames in an autoregressive manner. Nevertheless, existing methods fail to overcome the challenges of object occlusions and distractions, and do not have any measures to intercept the propagation of tracking errors. To tackle them, we present a SAMITE model, built upon SAM2 with additional modules, including: (1) Prototypical Memory Bank: We propose to quantify the feature-wise and position-wise correctness of each frame's tracking results, and select the best frames to condition subsequent frames. As the features of occluded and distracting objects are feature-wise and position-wise inaccurate, their scores would naturally be lower and thus can be filtered to intercept error propagation; (2) Positional Prompt Generator: To further reduce the impacts of distractors, we propose to generate positional mask prompts to provide explicit positional clues for the target, leading to more accurate tracking. Extensive experiments have been conducted on six benchmarks, showing the superiority of SAMITE. The code is available at https://github.com/Sam1224/SAMITE.

cs.CV↗

arXiv2025

LLMs Meet Cross-Modal Time Series Analytics: Overview and Directions

Large Language Models (LLMs) have emerged as a promising paradigm for time series analytics, leveraging their massive parameters and the shared sequential nature of textual and time series data. However, a cross-modality gap exists between time series and textual data, as LLMs are pre-trained on textual corpora and are not inherently optimized for time series. In this tutorial, we provide an up-to-date overview of LLM-based cross-modal time series analytics. We introduce a taxonomy that classifies existing approaches into three groups based on cross-modal modeling strategies, e.g., conversion, alignment, and fusion, and then discuss their applications across a range of downstream tasks. In addition, we summarize several open challenges. This tutorial aims to expand the practical application of LLMs in solving real-world problems in cross-modal time series analytics while balancing effectiveness and efficiency. Participants will gain a thorough understanding of current advancements, methodologies, and future research directions in cross-modal time series analytics.

cs.LG↗

arXiv2025

Skip a Layer or Loop it? Test-Time Depth Adaptation of Pretrained LLMs

Can a pretrained neural network adapt its architecture to different inputs without any finetuning? Do we need all layers for simple tasks, and are they adequate for challenging tasks? We found that the layers of a pretrained large language model (LLM) can be manipulated as separate modules to build a better and even shallower model customized for each test sample. In particular, each layer from the pretrained model can be skipped/pruned or repeated multiple times as recurrent neural networks (RNN), and stacked with others in arbitrary orders, yielding a chain-of-layers (CoLa) per sample. This compositional space greatly expands the scope of existing works on looped/recurrent pretrained modules, layer pruning, or early-exit networks. We develop a Monte Carlo Tree Search (MCTS) protocol to explore and identify the optimal CoLa for each sample from math and commonsense reasoning benchmarks. Compared to a static model of a fixed depth, CoLa allows shortcut paths (fast thinking), recurrence of the same layer(s) (slow thinking), and combining both, offering more flexible, dynamic architectures for different inputs. We conduct an extensive analysis of the MCTS-optimized CoLa, which leads to two key findings: (1) For >75% of samples with correct predictions by the original LLM, we can find shorter CoLa, suggesting a large space for improving inference efficiency; (2) For >60% of samples with originally incorrect predictions, we can identify CoLa achieving correct predictions, suggesting a large space of performance enhancement. Our results highlight the shortcomings of using a fixed architecture of pre-trained LLMs for inference on different samples and pave the way to unlock the generalization power of test-time depth adaptation.

cs.LG↗