Table reasoning with large language models (LLMs) plays a critical role in building intelligent systems capable of understanding and analyzing tabular data. Despite recent progress, existing methods still face key limitations: their reasoning processes lacks depth and explicit multi-step reasoning, often relying solely on implicit language model understanding. In addition, their reasoning processes suffer from instability, primarily caused by model uncertainty. In this work, we propose STaR, a novel slow-thinking model that can achieve effective and stable table reasoning. To enable effective multi-step reasoning, we design a two-stage training framework consisting of supervised fine-tuning (SFT) warm-up followed by reinforced fine-tuning (RFT). Specifically, in the SFT stage, we construct a high-quality dataset through automatic self-verification. In the RFT stage, we introduce a difficulty-aware reinforcement learning mechanism to further enhance reasoning capabilities. Furthermore, to improve reasoning stability, we introduce trajectory-level uncertainty quantification, which fuses token-level confidence with answer-level consistency, enabling the selection of better reasoning trajectories. Extensive experiments demonstrate that STaR-8B achieves state-of-the-art performance on in-domain benchmarks and exhibits strong generalization to out-of-domain datasets, highlighting its potential for enhancing both effectiveness and stability in table reasoning.
论文检索
输入标题、作者或关键词，从 13,734 篇学术成果中精准定位
Using Web Mining and Content Analysis to find and understand clues from the massive amount of unstructured text online is very important for predicting future events and providing early risk warnings in important fields like finance and public safety. While Large Language Models (LLMs) exhibit potential in processing and understanding text, current text-based event prediction faces two primary challenges: first, an insufficient utilization of potential information within the text, such as causal relationships and latent associations, and second, limited predictive reliability constrained by issues like the LLM's own ability and hallucinations. To address these challenges, we propose a novel event prediction framework, Bidirectional Reasoning with Self-Correction (BRSC). BRSC comprises two complementary reasoning dimensions: temporal deductive reasoning, which analyzes the trajectory of historical events along the timeline to enable accurate trend extrapolation, and synchronic associative reasoning, which deeply mines details and latent connections from documents within a specific time window to extended semantic information. In addition, we use a self-correction mechanism that identifies and rectifies potential hallucinations and errors during the reasoning process. Extensive experiments on international relations event prediction demonstrate that BRSC achieves significant improvements over several leading LLM-based methods.
Temporal Point Processes (TPPs) have been widely used for modeling event sequences on the Web, such as user reviews, social media posts, and online transactions. However, traditional TPP models often struggle to effectively incorporate the rich textual descriptions that accompany these events, while Large Language Models (LLMs), despite their remarkable text processing capabilities, lack mechanisms for handling the temporal dynamics inherent in Web-based event sequences. To bridge this gap, we introduce Language-TPP, a unified framework that seamlessly integrates TPPs with LLMs for enhanced Web event sequence modeling. Our key innovation is a novel temporal encoding mechanism that converts continuous time intervals into specialized byte-tokens, enabling direct integration with standard language model architectures for TPP modeling without requiring TPP-specific modifications. This approach allows Language-TPP to achieve state-of-the-art performance across multiple TPP benchmarks, including event time prediction and type prediction, on real-world Web datasets spanning e-commerce reviews, social media and online Q&A platforms. More importantly, we demonstrate that our unified framework unlocks new capabilities for TPP research: incorporating temporal information improves the quality of generated event descriptions, as evidenced by enhanced ROUGE-L scores, and better aligned sentiment distributions. Through comprehensive experiments, including qualitative analysis of learned distributions and scalability evaluations on long sequences, we show that Language-TPP effectively captures both temporal dynamics and textual patterns in Web user behavior, with important implications for content generation, user behavior understanding, and Web platform applications. Code is available at https://github.com/qykong/Language-TPP.
Logs serve as a primary source of information for engineers to diagnose failures in large-scale online service systems. Log parsing, which extracts structured events from massive unstructured log data, is a critical first step for downstream tasks like anomaly detection and failure diagnosis. With advances in large language models (LLMs), leveraging their strong text understanding capabilities has proven effective for accurate log parsing. However, existing LLM-based log parsers all focus on the constant part of logs, ignoring the potential contribution of the variable part to log parsing. This constant-centric strategy brings four key problems. First, inefficient log grouping and sampling with only constant information. Second, a relatively large number of LLM invocations due to constant-based cache, leading to low log parsing accuracy and efficiency. Third, a relatively large number of consumed constant tokens in prompts leads to high LLM invocation costs. At last, these methods only retain placeholders in the results, losing the system visibility brought by variable information in logs. Facing these problems, we propose a variable-centric log parsing strategy named VarParser. Through variable contribution sampling, variable-centric parsing cache, and adaptive variable-aware in-context learning, our approach can efficiently capture the variable parts of logs and leverage their contributions to parsing. By introducing variable units, we preserve rich variable information, enhancing the integrity of log parsing results. Extensive evaluations on large-scale datasets demonstrate that VarParser achieves higher accuracy compared to existing methods, significantly improving parsing efficiency while reducing the LLM invocation costs.
Recommender systems have been shown to exhibit popularity bias by over-recommending popular items and under-recommending relevant niche items. We seek to understand niche users in benchmark recommendation datasets as a step toward mitigating popularity bias. We find that, compared to mainstream users, niche-preferring users exhibit a longer-tailed activity-level distribution, indicating the existence of users who both prefer niche items and exhibit high activity levels on platforms. We partition users along two axes: (1) activity level (''power'' vs. ''light'') and (2) item-popularity preference (''mainstream'' vs. ''niche''), and show that in three benchmark datasets, the number of power-niche users (high activity and niche preference) is statistically significantly larger than expected. We also find that interaction data from power-niche users is especially valuable for improving recommendations for not only niche but also mainstream users. In contrast, many existing popularity bias mitigation methods have focused on upweighting niche users regardless of activity level. Motivated by the value of power-niche user data, we propose PAIR (Popularity-and-Activity-Informed Reweighting), a framework for reweighting the Bayesian Personalized Ranking (BPR) loss that simultaneously reweights based on user activity level and item popularity, upweighting power-niche users the most. We instantiate the framework on both deep and shallow collaborative filtering models, and experiments on benchmark datasets show that PAIR reduces popularity bias and can increase overall performance. Although existing popularity-bias mitigation methods yield a trade-off between performance and bias, our results suggest that considering both user activity level and popularity preference leads to Pareto-dominant performance.
The performance of Recommender Systems (RS) varies significantly across users, yet the underlying reasons for this variance remain poorly understood. This paper introduces a unified framework to analyze and explain this performance gap by quantifying user profile characteristics. We propose two novel, information-theoretic measures: Mean Surprise (𝑆(𝑢)), which captures a user's deviation from popular items and is closely related to popularity bias, and Mean Conditional Surprise (𝐂𝑆(𝑢)), which measures the internal coherence of a user's interactions in a domain-agnostic manner. Through extensive experiments on 7 algorithms and 9 datasets, we demonstrate that these measures are strong predictors of recommendation performance. Our analysis reveals that performance gains from complex models are concentrated on ''coherent'' users, while all algorithms perform poorly on ''incoherent'' users. We show how these measures provide practical utility for the Web community by: (1) enabling robust, stratified evaluation to identify model weaknesses; (2) facilitating a novel analysis of the behavioral alignment of recommendations; and (3) guiding targeted system design, which we validate by training a specialized model on a segment of ''coherent'' users that achieves superior performance for that group with significantly less data. This work provides a new lens for understanding user behavior and offers practical tools for building more robust and efficient large-scale recommender systems.
Search-based recommendation is one of the most critical application scenarios in e-commerce platforms. Users' complex search contexts—such as spatiotemporal factors, historical interactions, and current query's information—constitute an essential part of their decision-making, reflecting implicit preferences that complement explicit query terms. Modeling such rich contextual signals and their intricate associations with candidate items remains a key challenge. Although numerous efforts have been devoted to building more effective search methods, existing approaches still show limitations in integrating contextual information, which hinders their ability to fully capture user intent. To address these challenges, we propose a context-aware reasoning-enhanced generative search framework for better understanding the complicated context. Specifically, the framework first unifies heterogeneous user and item contexts into textual representations or text-based semantic identifiers and aligns them. To overcome the lack of explicit reasoning trajectories, we introduce a self-evolving post-training paradigm that iteratively combines supervised fine-tuning and reinforcement learning to progressively enhance the model's reasoning capability. In addition, we identify potential biases in existing RL algorithms when applied to search scenarios and present a debiased variant of GRPO to improve ranking performance. Extensive experiments on search log data collected from a real-world e-commerce platform demonstrate that our approach achieves superior performance compared with strong baselines, validating its effectiveness for search-based recommendation.
Scaling laws have enabled large language models(LLMs) to achieve remarkable performance and strong generalization across diverse language understanding tasks, including few-shot, in-context, and zero-shot learning. While prior studies in large-scale collaborative filtering(CF) have revealed clear relationships between model performance and scaling factors such as data size and model capacity, little attention has been given to how heterogeneous datasets can be synergistically combined for recommender systems(RS). In particular, it remains unclear whether systematically integrating diverse recommendation datasets can yield scaling behaviors analogous to those observed in LLMs, while simultaneously addressing challenges such as cold-start recommendation and cross-domain transfer. In this paper, we present RecCLIP, a multimodal framework that reformulates user--item interactions as visual representations compatible with vision--language models(VLMs). RecCLIP compresses interaction signals and employs prompt-based ranking to enable unified representation across heterogeneous data sources. Extensive experiments reveal consistent power-law scaling trends with respect to data size, and demonstrate that RecCLIP achieves superior performance in both cold-start and cross-domain transfer scenarios. Our findings underscore the importance of data-centric design in recommender systems and provide practical insights into scaling them effectively.The code for replication is available at https://github.com/jinliwei-1/RecCLIP.
Online game advertising is a prominent class of Web-mediated interactive services, where understanding and predicting player Lifetime Value (LTV) is a core scientific challenge in Web-scale user modeling, personalization, and digital economy optimization. However, the LTV prediction task poses severe challenges to traditional methods, which include data sparsity and complex distribution characteristics (such as zero-inflation, long tail, multimodal distribution, and cross-game). Existing methods struggle to capture the realistic and complex LTV distributions and exhibit limitations in leveraging cross-game data. We propose the first cross-game dynamic mixture framework with weighted sub-distributions for LTV prediction, DynaMoLTV. DynaMoLTV primarily models complex distributions via a zero-inflated mixture of lognormal (ZIMLN) loss, incorporates a game expert for cross-game data adaptation, employs a hierarchical payment classifier to capture consumption pattern variations, and integrates coarse and fine-grained losses to balance high-value user identification with LTV prediction accuracy. We conduct comprehensive experiments. The results demonstrate that DynaMoLTV achieves the best performance compared to five state-of-the-art baselines across metrics, including paid user identification, high-value user recall and LTV prediction accuracy. Specifically on three gaming datasets, DynaMoLTV reduces RMSE by 0.76%–46.65%, improves AUC by 0.94%–7.11%, and improves Norm-GINI by 0.63%–11.77% compared to five state-of-the-art baselines. DynaMoLTV also significantly improves ranking capabilities, with Recall@50K increasing by 17.64%–577.78%. We validate DynaMoLTV's effectiveness through two online A/B tests: (1) In the scenario of churned user re-engagement, DynaMoLTV increases online LTV by 20.3%-142.6% and downloads by 22.6%-37.7%. (2) In the scenario of online game advertising, DynaMoLTV increases GMV by 1.89% and GMV(ROI) by 27.31%. Our method has been fully deployed in a Web-based online game advertising platform, which ensures that LTV predictions remain personalized for online gaming ad delivery, supporting smarter and more inclusive decision-making on the Web.
From Prediction to Understanding: Leveraging Reasoning in Large Language Model-based Recommendations
Recently, large language models (LLMs) have shown great promise in sequential recommendation. Existing methods typically transform users' historical interactions into textual sequences and then feed these sequences into LLMs to generate recommended items. However, since the output of LLMs is limited to recommendations, the models must analyze user interactions and infer user preferences implicitly. This implicit approach constrains the expressive reasoning abilities of LLMs. Moreover, as the output of LLMs consists solely of recommended items, the resulting recommendations lack explainability. To address these issues, we propose RE2, which enables LLMs to explicitly generate reasoning content before providing recommendations. Making the reasoning process explicit helps elicit the reasoning abilities of LLMs and simultaneously enhances the explainability of recommendation results. RE2 consists of three steps: (1) Reasoning Collection, which collects reasoning data by guiding LLMs to generate both reasoning content and recommended items for each interaction sequence. (2) Pattern Imitation, which leverages the collected data to train LLMs via supervised fine-tuning to imitate the pattern of first generating reasoning content and then providing recommendations. (3) Pattern Internalization, which further internalizes this reasoning-and-recommendation pattern and enhances both recommendation performance and the rationality of the reasoning content through reinforcement learning. RE2 can be implemented under both the self-distillation and teacher-distillation frameworks, without requiring external user metadata such as user reviews. Experimental results demonstrate the effectiveness of RE2 in improving recommendation performance and in generating high-quality reasoning content. Furthermore, we show that RE2 can mitigate popularity bias while maintaining recommendation accuracy to some extent. Our data and code are available at https://github.com/zhiyuanc2001/RE2.
Large Language Models (LLMs) with powerful reasoning capabilities, offer new opportunities for recommendation systems (RS) , especially in understanding user preference. However, the direct application of LLMs to this task faces two major challenges: (1) limited context windows struggle to process the extensive user behavior sequences in real-world scenarios; (2) inherent hallucination effect can lead LLMs to infer spurious preferences that contradict true user intent, thereby degrading recommendation quality. To address this, we propose Rec 2, a Reinforcement learning-constrained segmented user modeling framework for recommendation. The framework segments lengthy behavior sequences into multiple segments and generates user preferences in a cascaded fashion. To suppress hallucinations and align with user intent, we introduce reinforcement learning, which uses the subsequent behavior segment as a supervisory signal to constrain and reward the preference inference process on the current segment, thereby generating more predictive and high-fidelity user preferences. Finally, the aligned LLMs infer preferences for each behavior segment; these are then processed by a specially-designed dynamic preference learning module to model preference evolution and are ultimately aggregated into a unified, dynamic long-term user preference embedding. This representation can be integrated into any recommendation model to boost its performance. Extensive experiments demonstrate that Rec2 significantly enhances recommendation performance by effectively capturing dynamic and authentic user preferences.
Recommendation systems aim to learn user interests from historical behaviors and deliver relevant items. Recent methods leverage large language models (LLMs) to construct and integrate semantic representations of users and items for capturing user interests. However, user behavior theories suggest that truly understanding user interests requires not only semantic integration but also semantic reasoning from explicit individual interests to implicit group interests. To this end, we propose an Iterative Semantic Reasoning Framework (ISRF) for generative recommendation. ISRF leverages LLMs to bridge explicit individual interests and implicit group interests in three steps. First, we perform multi-step bidirectional reasoning over item attributes to infer semantic item features and build a semantic interaction graph capturing users' explicit interests. Second, we generate semantic user features based on the semantic item features and construct a similarity-based user graph to infer the implicit interests of similar user groups. Third, we adopt an iterative batch optimization strategy, where individual explicit interests directly guide the refinement of group implicit interests, while group implicit interests indirectly enhance individual modeling. This iterative process ensures consistent and progressive interest reasoning, enabling more accurate and comprehensive user interest learning. Extensive experiments on the Sports, Beauty, and Toys datasets demonstrate that ISRF outperforms state-of-the-art baselines. The code is available at https://github.com/htired/ISRF.
Optimizing recommender systems for objectives beyond accuracy, such as diversity, novelty, and personalization, is crucial for long-term user satisfaction. To this end, the industry has accumulated vast amounts of structured domain knowledge, which we term human priors (e.g., item taxonomies, temporal patterns). This knowledge is typically applied through post-hoc adjustments during ranking or post-ranking. However, this approach remains decoupled from the core model learning, which is particularly undesirable as the industry shifts to end-to-end generative recommendation foundation models. On the other hand, many methods targeting these beyond-accuracy objectives often require architecture-specific modifications and discard these valuable human priors by learning user intent in a fully unsupervised manner. Instead of discarding the human priors accumulated over years of practice, we introduce a backbone-agnostic framework that seamlessly integrates these human priors directly into the end-to-end training of generative recommenders. With lightweight, prior-conditioned adapter heads inspired by efficient LLM decoding strategies, our approach guides the model to disentangle user intent along human-understandable axes (e.g., interaction types, long- vs. short-term interests). We also introduce a hierarchical composition strategy for modeling complex interactions across different prior types. Extensive experiments on three large-scale datasets demonstrate that our method significantly enhances both accuracy and beyond-accuracy objectives. We also show that human priors allow the backbone model to more effectively leverage longer context lengths and larger model sizes.
Recent advances in large language models (LLMs) have enabled more semantic-aware recommendations through natural language generation. Existing LLM for recommendation (LLM4Rec) methods mostly operate in a System 1-like manner, relying on superficial features to match similar items based on click history, rather than reasoning through deeper behavioral logic. This often leads to superficial and erroneous recommendations. Inspired by this, we propose ThinkRec, a thinking-based framework that shifts LLM4Rec from an intuitive system to a rational system. First, ThinkRec introduces a thinking activation mechanism by injecting synthetic reasoning traces, making the recommendation process resemble the Chain of Thought (CoT) reasoning of LLMs. This mechanism analyzes interaction histories, identifies user preferences, and makes decisions based on target items. Furthermore, considering the highly diverse distribution of recommendation data, we propose an instance-wise expert fusion mechanism to reduce the reasoning difficulty. By dynamically assigning weights to expert models based on users' latent features, ThinkRec adapts its reasoning path to individual users, thereby enhancing precision and personalization. Extensive experiments on various real-world web user behavior preference datasets demonstrate that ThinkRec significantly outperforms baselines in terms of recommendation accuracy and interpretability, providing superior recommendations based on a deeper understanding of user intent and a more rigorous reasoning process. Code is available in https://github.com/Yu-Qi-hang/ThinkRec.
Time series anomaly detection (TSAD) has been a long-standing pillar problem in Web-scale systems and online infrastructures, such as service reliability monitoring, system fault diagnosis, and performance optimization. Large language models (LLMs) have demonstrated unprecedented capabilities in time series analysis, the potential of multimodal LLMs (MLLMs), particularly vision-language models, in TSAD remains largely under-explored. One natural way for humans to detect time series anomalies is through visualization and textual description. It motivates our research question: Can multimodal LLMs perform time series anomaly detection? Existing studies often oversimplify the problem by treating point-wise anomalies as special cases of range-wise ones or by aggregating point anomalies to approximate range-wise scenarios. They limit our understanding for realistic scenarios such as multi-granular anomalies and irregular time series. To address the gap, we build a VisualTimeAnomaly benchmark to comprehensively investigate zero-shot capabilities of MLLMs for TSAD, progressively from point-, range-, to variate-wise anomalies, and extends to irregular sampling conditions. Our study reveals several key insights. 1) MLLMs and traditional TSAD methods are complementary: MLLMs excel at coarse-grained anomalies while traditional methods are effective at fine-grained anomalies. 2) MLLMs are resilient to irregular time series. 3) Input time series modality changing from text to image makes information focus shift from quantitative variations to qualitative patterns while significantly reducing hallucinations. Built on the findings, we propose a MLLMs-based multi-agent framework TSAD-Agents to achieve automatic TSAD. Our framework comprises scanning, planning, detection, and checking agents that synergistically collaborate to reason, plan, and self-reflect to enable automatic TSAD. These agents adaptively invoke tools such as traditional methods and MLLMs and dynamically switch between text and image modalities to optimize detection performance.
Web service administrators must ensure the stability of multiple systems by promptly detecting anomalies in Key Performance Indicators (KPIs). Achieving the goal of ''train once, infer across scenarios'' remains a fundamental challenge for time series anomaly detection models. Beyond improving zero-shot generalization, such models must also flexibly handle sequences of varying lengths during inference—ranging from one hour to one week—without retraining. Conventional approaches rely on sliding-window encoding and self-supervised learning, which restrict inference to fixed length inputs. Large Language Models (LLMs) have demonstrated remarkable zero-shot capabilities across general domains. However, when applied to time series data, they face inherent limitations due to context length. To address this issue, we propose ViTs, a Vision-Language Model (VLM)-based framework that converts time series curves into visual representations. By rescaling time series images, temporal dependencies are preserved while maintaining a consistent input size, thereby enabling efficient processing of arbitrarily long sequences without context constraints. Training VLMs for this purpose introduces unique challenges, primarily due to the scarcity of aligned time series image–text data. To overcome this, we employ an evolutionary algorithm to automatically generate thousands of high-quality image–text pairs and design a three-stage training pipeline consisting of: (1) time series knowledge injection, (2) anomaly detection enhancement, and (3) anomaly reasoning refinement. Extensive experiments demonstrate that ViTs substantially enhance the ability of VLMs to understand and detect anomalies in time series data. All datasets and code will be publicly released at: https://github.com/Uni-WangZexin/ViTs.
Cross-view geo-localization (CVGL) establishes correspondences between ground-level and satellite images of the same geographic location, serving as a fundamental technology for smart city applications, including autonomous navigation, urban planning, and location-based services. Current CVGL approaches fall into two categories: feature-based methods achieve superior performance through 2D representation learning but lack interpretability. Spatial-based methods provide geometric understanding and interpretable matching but suffer from limited spatial modeling and weak cross-view alignment, leading to lower performance. We reformulate CVGL from a spatial perspective and propose an auxiliary task-enhanced network. The network captures spatial semantics and provides explicit alignment processes with visualizable results. We introduce an auxiliary spatial semantic alignment (SSA) task that learns spatial structure via vision foundation models (VFM) and BEV transformation to enhance the primary CVGL task. The primary task captures visual semantics, including texture and appearance. Within this unified framework with shared encoders, the primary task enriches the learned embeddings by fusing spatial structure with visual semantics, yielding spatially complete representations. Extensive experiments on three standard CVGL benchmarks demonstrate that our method significantly surpasses previous spatial-based approaches while maintaining competitive performance with state-of-the-art (SOTA) feature-based methods, achieving 98.48% R@1 on CVUSA and 71.05% R@1 on CVACT\_test. We provide comprehensive analyses through pixel-level activation maps and feature-space UMAP visualizations to validate both effectiveness and interpretability.
Trace analysis is essential for understanding system behaviors, detecting anomalies, and diagnosing faults in complex microservice-based web applications. Existing trace analysis approaches face several challenges in industrial microservice-based systems, including high manual overhead, limited functionality, unfriendly interaction mechanisms, and difficulties in deployment and integration. The strong capabilities of large language models (LLMs) in natural language understanding, reasoning, and multi-task generalization provide new opportunities for a more intelligent and flexible trace analysis approach. However, the trace analysis capabilities of LLMs remain underexplored and underdeveloped. To bridge this gap, we conduct the first comprehensive evaluation on the trace analysis capabilities of LLMs. In particular, we construct the first instruction&response benchmark dataset for trace analysis, named TraceBench. It involves a wide range of trace analysis tasks, allowing us to systematically evaluate the capabilities of LLMs in this area. Experimental results show that LLMs have potential in handling trace analysis tasks, but there leaves room for improvement. To this end, we propose TraceLLM, an approach that significantly enhances the capabilities of LLMs via fine-tuning, outperforming the open-source LLMs by 34.77% on average in terms of accuracy, and outperforming the closed-source model by 21.66% in the best case. The generalization and robustness of TraceLLM are also confirmed in our experiments. To the best of our knowledge, TraceLLM is the first LLM which is specialized for handling various types of trace analysis tasks. This work provides a foundation for future research to further explore the trace analysis capabilities of LLMs.
Traffic prediction serves as a cornerstone for systems and network services such as the Web of Vehicles (WoV), online navigation, and smart city applications. Despite the proliferation of model architectures in recent years, existing approaches often suffer from highly customized structures and weak transferability, making it difficult to cope with increasing task heterogeneity and modeling complexity. To address these challenges, we propose ST-LEGO, a modular assembly framework driven by large language models (LLMs) that supports flexible structural composition and automated code generation. ST-LEGO employs a multi-agent collaborative system comprising a Prompt Agent, Assemble Agent, and Code Agent, which are responsible for understanding task requirements, dynamically assembling structural modules, and automatically generating executable PyTorch code. By introducing a standardized module library and an intermediate structural description language (DSL), the framework enables controllable generation, reusable composition, and cross-task generalization of model architectures. Empirical results on multiple real-world traffic datasets demonstrate that models generated by ST-LEGO achieve superior accuracy, structural diversity, and convergence compared to a wide range of manually designed baselines. These results highlight the unique potential and scalability of LLMs as structural architects for traffic prediction, offering a new paradigm for integrating language models into web-interactive intelligent transportation systems.
Large-scale vision-language models enable powerful cross-modal understanding and generation, driving rapidly growing demand for online inference services. However, cloud-centric serving often suffers from high latency, rising costs, and network dependency, while purely on-device deployment is constrained by limited memory and reduced accuracy on complex tasks. To address this accuracy–latency–cost trilemma, we propose ShiftVL, a task-aware end–cloud serving framework that shifts suitable execution to the end device with a cloud fallback. ShiftVL serves high-frequency requests on an end-side small VLM enhanced with ViTexLoRA, a modality-disentangled parameter-efficient tuning method that preserves cross-modal alignment, while routing low-frequency or complex requests to a cloud-hosted large VLM for higher accuracy. Under tight device budgets, ShiftVL employs a predictive adapter scheduler that combines LRU-style caching with imitation learning to pre-load task-specific adapters. Experiments with InternVL models show that ShiftVL reduces cloud cost by up to 76.3% and latency by up to 42.9% while maintaining high multi-task accuracy, demonstrating its practicality for real-world vision-language model serving.