arXiv subjects

Xue Li

Publications and source records attributed to Xue Li.

At least 37 records · Page 2Linked to original sources

arXiv2026

Modality Relevance is not Modality Utility: Post-hoc Selective Modality Escalation for Cost-Aware Multimodal RAG

Multimodal retrieval-augmented generation (RAG) grounds a generator in evidence drawn from heterogeneous modalities -- text, tables, and images. The dominant deployment choice is binary and made before the model has tried to answer: either run a cheap text(+table) pipeline, or pay for an expensive vision-language model (VLM) over every image. Recent adaptive systems improve on this by selecting the modality or fidelity pre-retrieval, from a question-conditioned predictor of which modality will be needed. We show that this is the wrong decision point. Through an oracle headroom analysis on MultiModalQA, we find that the relevance of a modality to a question is a weak predictor of whether that modality is actually needed to answer correctly: a large fraction of questions whose gold support includes an image are nonetheless answerable from text and tables alone, and a pre-retrieval router that escalates on apparent visual relevance over-escalates substantially relative to an oracle. We propose \textbf{post-hoc selective modality escalation}: answer cheaply from text and tables, run a verifier on the (query, draft answer, evidence) tuple that localizes which modality is missing, and pay for VLM evidence only there. A calibrated value-of-escalation router then decides whether the expected accuracy gain justifies the visual cost. On MultiModalQA, our router recovers the accuracy of an always-on VLM pipeline while issuing far fewer visual calls, and closes most of the gap to the oracle escalation rate. The result extends a routing-signal hierarchy established for retrieval depth and reasoning hops to a third axis -- modality -- under a single cost-aware selective-escalation view.

cs.IR↗

arXiv2026

MedP-CLIP: Medical CLIP with Region-Aware Prompt Integration

Contrastive Language-Image Pre-training (CLIP) has demonstrated outstanding performance in global image understanding and zero-shot transfer through large-scale text-image alignment. However, the core of medical image analysis often lies in the fine-grained understanding of specific anatomical structures or lesion regions. Therefore, precisely comprehending region-of-interest (RoI) information provided by medical professionals or perception models becomes crucial. To address this need, we propose MedP-CLIP, a region-aware medical vision-language model (VLM). MedP-CLIP innovatively integrates medical prior knowledge and designs a feature-level region prompt integration mechanism, enabling it to flexibly respond to various prompt forms (e.g., points, bounding boxes, masks) while maintaining global contextual awareness when focusing on local regions. We pre-train the model on a meticulously constructed large-scale dataset (containing over 6.4 million medical images and 97.3 million region-level annotations), equipping it with cross-disease and cross-modality fine-grained spatial semantic understanding capabilities. Experiments demonstrate that MedP-CLIP significantly outperforms baseline methods in various medical tasks, including zero-shot recognition, interactive segmentation, and empowering multimodal large language models. This model provides a scalable, plug-and-play visual backbone for medical AI, combining holistic image understanding with precise regional analysis.

cs.CV↗

arXiv2026

Multi-Adapter Representation Interventions via Energy Calibration

Representation intervention has emerged as a promising paradigm for aligning large language models toward desired behaviors without modifying model weights. Existing methods typically apply a fixed intervention uniformly across all inputs. However, we find that the appropriate intervention direction and strength vary substantially across samples, and such indiscriminate intervention leads to degradation of general capabilities on benign inputs. To address these challenges, we propose Multi-Adapter Representation Interventions via Energy Calibration (MARI). Specifically, we introduce a competitive multi-adapter mechanism in which specialized experts capture non-linear correction patterns and adaptively determine the appropriate intervention direction and strength for different samples. Furthermore, we design an energy-based gating module that leverages internal propagation dynamics to distinguish inputs that are applicable for intervention. Extensive experiments across diverse model families and parameter scales demonstrate that MARI achieves state-of-the-art alignment performance. Our method significantly improves performance on TruthfulQA, BBQ, and safety benchmarks, while maintaining and even improving general capabilities on tasks such as MMLU and ARC. Our code is available at https://github.com/V1centNevwake/MARI.

cs.AI↗

arXiv2026

Termination-Dependent Surface States and Magnetic Fingerprints of Chiral Helimagnet Cr1/3TaS2

Chiral helimagnets based on intercalated transition-metal dichalcogenides, characterized by nano-scale spin ordering, provide a powerful route to engineer chiral spin textures (e.g. the topologically protected magnetic solitons) and emergent electronic functionality at reduced dimensions, where surface and interface states often dominate device operation. However, despite growing interest, direct experimental studies of termination-dependent surface electronic structures and their temperature-driven magnetic evolution remain largely unexplored, hindering a microscopic understanding of the electronic states that is crucial for the development of low-dimensional spintronic devices. Here, for the first time, taking Cr1/3TaS2 as a representative example, we systematically investigate the termination-dependent surface electronic states of the chiral helimagnets and uncover their distinct temperature evolution across the magnetic transition (TC~142K) by combining high-resolution ARPES with a micro-focused beam and surface-state-resolved first-principles calculations. The TaS2-terminated surface hosts folded monolayer-like TaS2 bands under the $\sqrt3\times\sqrt3$ superlattice potential and a shallow triangular electron pocket at the superlattice $\bar K$ point arising from Cr-Ta orbital hybridization. In contrast, the Cr-terminated surface exhibits reconstructed hole pockets with pronounced magnetic band splitting. This splitting disappears above TC and closely follows the chiral helimagnetic order parameter, providing a direct spectroscopic fingerprint of chiral helimagnetic order. In addition, multiple ultranarrow Cr-d-derived surface flat bands are resolved. These findings establish Cr1/3TaS2 as a model system in which surface electronic states are strongly coupled to chiral magnetism, opening new opportunities for chiral spintronic and valleytronic micro/nanodevices.

cond-mat.mtrl-sci↗

arXiv2026

RCTEA: Richness-guided Co-training for Temporal Entity Alignment

Temporal Entity Alignment (TEA), which aims to identify equivalent entities across Temporal Knowledge Graphs (TKGs), is crucial for integrating knowledge facts from multiple sources. However, existing TEA models often fail to capture the orthogonal yet complementary effects between structural and temporal features, and typically overlook the importance of information richness, a key factor for effective message passing in neural feature encoders. To address these limitations, we propose the RCTEA framework, which jointly models both structural and temporal aspects of TKGs for entity alignment. Specifically, we design a richness-guided attention mechanism along with an adaptive weighting strategy to facilitate effective feature fusion. To ensure robust alignment despite noisy entity contexts, we introduce a dual-view neighborhood consensus algorithm that jointly refines the feature encoders to enforce local structural consistency of the predicted alignments. Extensive experiments demonstrate the superiority of RCTEA, achieving state-of-the-art performance on public TEA benchmarks.

cs.IR↗

arXiv2026

Pythia: Exploiting Workflow Predictability for Efficient Agent-Native LLM Serving

As LLM applications grow more complex, developers are increasingly adopting multi-agent architectures to decompose workflows into specialized, collaborative components, introducing structure that constrains agent behavior and exposes useful semantic predictability. Unlike traditional LLM serving, which operates under highly dynamic and uncertain conditions, this structured topology enables opportunities to reduce runtime uncertainty$\unicode{x2015}$yet existing systems fail to exploit it, treating agentic workloads as generic traffic and incurring significant inefficiencies. Our analysis of production traces from an agent-serving platform and an internal coding assistant reveals key bottlenecks, including low prefix cache hit rates, severe resource contention from long-context requests, and substantial queuing delays due to suboptimal scaling. To address these challenges, we propose Pythia, a multi-agent serving system that captures workflow semantics through a simple interface at the serving layer, unlocking new optimization opportunities and substantially improving throughput and job completion time over state-of-the-art baselines.

cs.MA↗

arXiv2026

ServeGen: Workload Characterization and Generation of Large Language Model Serving in Production

With the widespread adoption of Large Language Models (LLMs), serving LLM inference requests has become an increasingly important task, attracting active research advancements. Practical workloads play an essential role in this process: they are critical for motivating and benchmarking serving techniques and systems. However, the existing understanding of real-world LLM serving workloads is limited due to the lack of a comprehensive workload characterization. Prior analyses remain insufficient in scale and scope, thus failing to fully capture intricate workload characteristics. In this paper, we fill the gap with an in-depth characterization of LLM serving workloads collected from our worldwide cloud inference serving service, covering not only language models but also emerging multimodal and reasoning models, and unveiling important new findings in each case. Moreover, based on our findings, we propose ServeGen, a principled framework for generating realistic LLM serving workloads by composing them on a per-client basis. A practical use case in production validates that ServeGen avoids 50% under-provisioning compared to naive workload generation, demonstrating ServeGen's advantage in performance benchmarking. ServeGen is available at https://github.com/alibaba/ServeGen.

cs.DC↗

arXiv2026

STAR: Decode-Phase Rescheduling for LLM Inference

Large Language Model (LLM) inference has emerged as a fundamental paradigm, however, variations in output length cause severe workload imbalance in the decode phase, particularly for long-output reasoning tasks. Existing systems, such as PD disaggregation architectures, rely on static prefill-to-decode scheduling, which often results in SLO violations and OOM failures under evolving decode workloads. In this paper, we propose STAR, a decode rescheduling system powered by length prediction to anticipate future workloads. Our core contributions include: (1) A lightweight and continuous LLM-native prediction method that leverages LLM hidden state to model remaining generation length with high precision (reducing MAE by 49.42%) and low overhead (cutting predictor parameters by 93.28%); (2) A rescheduling solution in decode phase with a dynamic balancing mechanism that integrates current and predicted workloads, reducing P99 TPOT by 75.1% and achieving 2.63 times higher goodput.

cs.DC↗

arXiv2026

RV and TTV Measurements of Two Transiting Long-Period Giants around TOI-4600

TOI-4600b and c, originally identified by the Transiting Exoplanet Survey Satellite (TESS) and reported by I. Mireles et al. (2023), are a rare pair of transiting long-period giant planets ($\rm P_b=82.7$ days, $\rm P_c=482.8$ days) orbiting an early K dwarf. In this work, we refine the orbital parameters of the TOI-4600 system by combining new TESS photometry, ground-based transit follow-up, and radial velocity (RV) observations from MAROON-X. We obtain improved constraints on planetary masses and eccentricities, and update other parameters, such as the stellar age. For TOI-4600b, we measure a mass of $M_p = 74.7^{+4.7}_{-4.4}\,M_{\oplus}$ and an eccentricity of $e=0.153^{+0.020}_{-0.018}$, and $M_p = 212.53^{+13.26}_{-13.03}\,M_{\oplus}$ and $e=0.219^{+0.015}_{-0.018}$ for TOI-4600c. We find significant transit timing variations (TTV) in both planets, with semi-amplitudes of approximately $1$\,hr. We derive Transit Spectroscopy Metric values of 16.87 for TOI-4600b and 10.09 for TOI-4600c, indicating that both planets are promising JWST targets for studying the atmospheres of temperate and cold Jupiters, a relatively poorly characterized sample thus far. These updated parameters and TTV ephemerides are important for planning and interpreting future photometric, spectroscopic, and dynamical studies of the TOI-4600 system.

astro-ph.EP↗

arXiv2026

Amoeba: Runtime Tensor Parallel Transformation for LLM Inference Services

In Large Language Model (LLM) inference services, it is challenging to make a parallelism strategy configuration, to efficiently process the requests of variance context lengths. Requests of long context require high degree of parallelism to provide more memory for Key-Value (KV) Cache, while requests of short context prefer low degree of parallelism to increase concurrency, thus improving throughput. To maintain high throughput while supporting large context lengths on demand, we propose Amoeba, a runtime Tensor Parallel (TP) transformation for online LLM inference services, which adaptively adjusts the TP of running instances to align with the dynamics of incoming requests. Evaluations using real-world traces show that Amoeba improves throughput by 1.75x-6.57x compared to state-of-the-art solutions.

cs.DC↗

arXiv2026

CoDec: Prefix-Shared Decoding Kernel for LLMs

Prefix-sharing among multiple prompts presents opportunities to combine the operations of the shared prefix, while attention computation in the decode stage, which becomes a critical bottleneck with increasing context lengths, is a memory-intensive process requiring heavy memory access on the key-value (KV) cache of the prefixes. Therefore, in this paper, we explore the potential of prefix-sharing in the attention computation of the decode stage. However, the tree structure of the prefix-sharing mechanism presents significant challenges for attention computation in efficiently processing shared KV cache access patterns while managing complex dependencies and balancing irregular workloads. To address the above challenges, we propose a dedicated attention kernel to combine the memory access of shared prefixes in the decoding stage, namely CoDec. CoDec delivers two key innovations: a novel shared-prefix attention kernel that optimizes memory hierarchy and exploits both intra-block and inter-block parallelism, and a comprehensive workload balancing mechanism that efficiently estimates cost, divides tasks, and schedules execution. Experimental results show that CoDec achieves an average $1.9\times$ speedup and $120.9\times$ memory access reduction compared to the state-of-the-art FlashDecoding kernel regarding attention computation in the decode stage and $3.8\times$ end-to-end time per output token compared to the vLLM.

cs.LG↗

arXiv2026

Are LLMs Smarter Than Chimpanzees? An Evaluation on Perspective Taking and Knowledge State Estimation

Cognitive anthropology suggests that the distinction of human intelligence lies in the ability to infer other individuals' knowledge states and understand their intentions. In comparison, our closest animal relative, chimpanzees, lack the capacity to do so. With this paper, we aim to evaluate LLM performance in estimating other individuals' knowledge states and their potential actions. We design two tasks to test (1) if LLMs can predict story characters' next actions based on their own knowledge vs. improperly using information unavailable from their perspective, and (2) if LLMs can detect when story characters, through their actions, demonstrate knowledge they should not possess. Results reveal that most current state-of-the-art LLMs achieve near-random performance on both tasks, and are substantially inferior to humans. We argue future LLM research should place more weight on the abilities of knowledge estimation and intention understanding.

cs.AI↗

arXiv2026

LEXA: Legal Case Retrieval via Graph Contrastive Learning with Contextualised LLM Embeddings

Legal case retrieval (LCR) is a specialised information retrieval task aimed at identifying relevant cases given a query case. LCR holds pivotal significance in facilitating legal practitioners to locate legal precedents. Existing LCR methods predominantly rely on traditional lexical models or language models; however, they typically overlook the domain-specific structural information embedded in legal documents. Our previous work CaseGNN successfully harnesses text-attributed graphs and graph neural networks to incorporate structural legal information. Nonetheless, three key challenges remain in enhancing the representational capacity of CaseGNN: (1) The under-utilisation of rich edge information in text-attributed case graph (TACG). (2) The insufficiency of training signals for graph contrastive learning. (3) The lack of contextualised legal information in node and edge features. In this paper, the LEXA model, an extension of CaseGNN, is proposed to overcome these limitations by jointly leveraging rich edge information, enhanced training signals, and contextualised embeddings derived from large language models (LLMs). Specifically, an edge-updated graph attention layer (EUGAT) is proposed to comprehensively update node and edge features during graph modelling, resulting in a full utilisation of structural information of legal cases. Moreover, LEXA incorporates a novel graph contrastive learning objective with graph augmentation to provide additional training signals, thereby strengthening the model's legal comprehension capabilities. What's more, LLMs are employed to generate node and edge features for TACG. Extensive experiments on two benchmark datasets demonstrate that LEXA not only significantly improves CaseGNN but also achieves supreme performance compared to state-of-the-art LCR methods.

cs.IR↗

arXiv2026

Varying core-envelope coupling efficiency identified from stellar rotation--activity relation

Core-envelope coupling provides a reasonable explanation of the spin-down stalling of stars in open clusters, which was not predicted by classical gyrochronology. However, it remains an open question whether the coupling efficiency is constant or variable. M dwarfs, possessing thicker convective envelopes and thus longer coupling timescales than other late-type stars, are ideal objects for this investigation. In this work, based on the $R_{\rm{HK}}^{'}$ measurements from LAMOST and DESI spectra, we construct new rotation--activity relations for M dwarfs. Unlike the traditional picture, we suggest that the new relation consists of three distinct regimes of fast, intermediate, and slow rotation, closely matching the three sequences of gyrochronology, namely the ``Convective'' sequence, ``Gap'', and ``Interface'' sequence. Our study reveals, for the first time, a variable activity decay rate in the intermediate-rotation regime (i.e., the ``Gap'' region). This implies a varying core-envelope coupling efficiency, peaking towards the end of this region. It also coincides with the well-known stage of stalled stellar spin-down.

astro-ph.SR↗

arXiv2026

Robust Test-time Video-Text Retrieval: Benchmarking and Adapting for Query Shifts

Modern video-text retrieval (VTR) models excel on in-distribution benchmarks but are highly vulnerable to real-world query shifts, where the distribution of query data deviates from the training domain, leading to a sharp performance drop. Existing image-focused robustness solutions are inadequate to handle this vulnerability in video, as they fail to address the complex spatio-temporal dynamics inherent in these shifts. To systematically evaluate this vulnerability, we first introduce a comprehensive benchmark featuring 12 distinct types of video perturbations across five severity degrees. Analysis on this benchmark reveals that query shifts amplify the hubness phenomenon, where a few gallery items become dominant "hubs" that attract a disproportionate number of queries. To mitigate this, we then propose HAT-VTR (Hubness Alleviation for Test-time Video-Text Retrieval), as our baseline test-time adaptation framework designed to directly counteract hubness in VTR. It leverages two key components: a Hubness Suppression Memory to refine similarity scores, and multi-granular losses to enforce temporal feature consistency. Extensive experiments demonstrate that HAT-VTR substantially improves robustness, consistently outperforming prior methods across diverse query shift scenarios, and enhancing model reliability for real-world applications.

cs.IR↗

arXiv2026

Refined DFT recipe and renormalisation of band-edge parameters for electrons in monolayer MoS$_2$ informed by the measured spin-orbit splitting

Conduction band-edge spin-orbit splitting (SOS) in monolayer transition metal dichalcogenides determines a competition between bright and dark excitons and sets conditions for spintronics applications of these semiconductors. Here, we report the SOS measurement for electrons in monolayer MoS$_2$, found from the threshold density, $n_*$, for the upper spin-orbit-split band population, which exceeds by an order of magnitude the values expected from the conventional density functional theory (DFT). Theoretically, half of the observed value can be attributed to the exchange enhancement of SOS in a finite-density electron gas, but explaining the rest requires refining the DFT approach. As the conduction band SOS in MoS$_2$ is set by a delicate balance between the contribution of sulphur $p_x$ and $p_y$ orbitals and $d_{z^2}-d_{xz}$ and $d_{z^2}-d_{yz}$ mixing in molybdenum, we use a DFT+U+V framework for fine-tuning the orbital composition of the relevant band-edge states. An optimised choice of Hubbard U/V parameters produces close agreement with the experimentally observed conduction band SOS in MoS$_2$, simultaneously resulting in the valence-band SOS and the quasi-particle band gap which are closer to their values established in the earlier-published experiments.

cond-mat.mes-hall↗

arXiv2026

Attention-weighted Centered Kernel Alignment for Knowledge Distillation in Large Audio-Language Models Applied to Speech Emotion Recognition

The emergence of Large Audio-Language Models (LALMs) has advanced Speech Emotion Recognition (SER), but their size limits deployment in resource-constrained environments. While Knowledge Distillation is effective for LALM compression, existing methods remain underexplored in distilling the cross-modal projection module (Projector), and often struggle with alignment due to differences in feature dimensions. We propose PL-Distill, a KD framework that combines Projector-Level Distillation (PDist) to align audio embeddings and Logits-Level Distillation (LDist) to align output logits. PDist introduces Attention-weighted Centered Kernel Alignment, a novel approach we propose to highlight important time steps and address dimension mismatches. Meanwhile, LDist minimizes the Kullback-Leibler divergence between teacher and student logits from audio and text modalities. On IEMOCAP, RAVDESS, and SAVEE, PL-Distill compresses an 8.4B-parameter teacher to a compact 1.1B-parameter student, consistently outperforming the teacher, state-of-the-art pretrained models, and other KD baselines across all metrics.

cs.SD↗

arXiv2026

DART: Diffusion-Inspired Speculative Decoding for Fast LLM Inference

Speculative decoding is an effective and lossless approach for accelerating LLM inference. However, existing widely adopted model-based draft designs, such as EAGLE3, improve accuracy at the cost of multi-step autoregressive inference, resulting in high drafting latency and ultimately rendering the drafting stage itself a performance bottleneck. Inspired by diffusion-based large language models (dLLMs), we propose DART, which leverages parallel generation to reduce drafting latency. DART predicts logits for multiple future masked positions in parallel within a single forward pass based on hidden states of the target model, thereby eliminating autoregressive rollouts in the draft model while preserving a lightweight design. Based on these parallel logit predictions, we further introduce an efficient tree pruning algorithm that constructs high-quality draft token trees with N-gram-enforced semantic continuity. DART substantially reduces draft-stage overhead while preserving high draft accuracy, leading to significantly improved end-to-end decoding speed. Experimental results demonstrate that DART achieves a 2.03x--3.44x wall-clock time speedup across multiple datasets, surpassing EAGLE3 by 30% on average and offering a practical speculative decoding framework. Code is released at https://github.com/fvliang/DART.

cs.CL↗