Searcharxiv⌕ Search

arXiv subjects

Kai Liu

Publications and source records attributed to Kai Liu.

At least 19 recordsLinked to original sources
arXiv2026

G$^2$TR: Generation-Guided Visual Token Reduction for Separate-Encoder Unified Multimodal Models

The development of separate-encoder Unified multimodal models (UMMs) comes with a rapidly growing inference cost due to dense visual token processing. In this paper, we focus on understanding-side visual token reduction for improving the efficiency of separate-encoder UMMs. While this topic has been widely studied for MLLMs, existing methods typically rely on attention scores, text-image similarity and so on, implicitly assuming that the final objective is discriminative reasoning. This assumption does not hold for UMMs, where understanding-side visual tokens must also preserve the model's capabilities for editing images. We propose G$^2$TR, a generation-guided visual token reduction framework for separate-encoder UMMs. Our key insight is that the generation branch provides a task-agnostic signal for identifying understanding-side visual tokens that are not only semantically relevant but also important for latent-space image reconstruction and generation. G$^2$TR estimates token importance from consistency with VAE latent, performs balanced token selection, and merges redundant tokens into retained representatives to reduce information loss. The method is training-free, plug-and-play, and applied only after the understanding encoding stage, making it compatible with existing UMM inference pipelines. Experiments on image understanding and editing benchmarks show that G$^2$TR substantially reduces visual tokens and prefill computation by 1.94x while maintaining both reasoning accuracy and editing quality, outperforming baselines on almost all benchmarks. Code is at: https://github.com/lijunxian111/G2TR.

cs.CV↗
arXiv2026

VOICE: A Vision-Omics Foundation Model Integrating Direct and Retrieval-Based Prediction of In-situ Single-Cell Gene Expression

Spatial transcriptomics can resolve gene expression at single-cell resolution, but it is costly, limited to targeted panels of a few hundred to a few thousand genes, and applicable to only a small number of samples. H&E imaging, by contrast, is cheap and collected routinely at scale. This makes predicting single-cell expression directly from morphology a practical way to bring molecular analysis to large tissue archives. We therefore present VOICE, a multimodal foundation model that predicts single-cell gene expression from H&E images using paired Xenium data. VOICE first aligns cell centered H&E morphology from a pathology foundation model with single-cell expression embeddings from a transcriptome foundation model, trained using contrastive learning over 23 million cells. Next it predicts expression through two branches. One branch directly regresses expression from morphology. The other branch retrieves measured expression from similar reference cells, recovering genes that do not have morphological signal. Because genes vary in morphological predictability, VOICE fuses the two branches with a per-gene weight. After training, VOICE generalizes to heldout patients, slides, and partially overlapping gene panels from Xenium, and it consistently outperforms prior single-cell expression prediction methods on seven metrics.

cs.CV↗
arXiv2026

DRScaffold: Boosting Dense-Scene Reasoning in Lightweight Vision Language Models

Lightweight vision-language models perform competitively on standard benchmarks yet fail systematically in dense-scene reasoning, where multiple objects, attributes, and relations must be jointly grounded and resolved through multi-step inference. Such capability is critical for real-world applications where models must reliably interpret cluttered environments. Yet existing training signals provide no explicit grounding between reasoning steps and the underlying visual entities and relations, leaving lightweight models free to generate fluent but visually unanchored reasoning chains. To address this gap, we first introduce DRBench, a benchmark of 14,573 questions across 2,943 images, organized into five task categories spanning three progressive reasoning layers. Building on DRBench, we propose DRScaffold, a supervised fine-tuning framework that decomposes the supervision target into four causally ordered stages, promoting grounded reasoning without architectural modification. Experiments on three lightweight VLMs demonstrate substantial gains on DRBench while preserving or improving performance on general-purpose benchmarks. Notably, Qwen2.5-VL-3B trained with DRScaffold surpasses the frozen Qwen2.5-VL-32B under zero-shot direct-answer prompting on DRBench, demonstrating that structured supervision can substitute for a significant portion of model scale in dense-scene reasoning. Our code is available at https://github.com/irene-shi/DRScaffold.

cs.CV↗
arXiv2026

Magnetoelectric Exchange Spring for Spin-Wave Logic with Ultralow Energy Consumption

Spin waves (magnons) are intrinsically free of Joule heating during their propagation in magnets. Exploring, understanding, and manipulating such quasiparticles not only helps reveal the exotic magnonic phenomena in condensed matter physics but also underpins the development of information devices with low energy consumption. However, achieving low-power and scalable control of spin-wave propagation remains a critical challenge. Herein, a magnetoelectric exchange spring has been artificially designed and constructed across an interface between ferroelectric BaTiO3 and ferromagnetic (La,Sr)MnO3 with low magnetic damping, in which polarization-modulated orbital reconstruction tailors local magnetic anisotropy and enables reversible control of spin-wave propagation. By successfully deviating from the Janovec-Kay-Dunn law, reversal polarization of engineered BaTiO3 (Ps~1.2 uC cm^-2, Vc~0.4 V) is sufficient to switch the coherent magnon transport in the (La,Sr)MnO3, indicating the magnetoelectric heterointerface may serve as a promising candidate for atto-joule control of spin-wave propagation (normalized to a 10*10 nm^2 device). Furthermore, taking advantage of amplitude, phase, and frequency to encode spin waves, all 16 compact Boolean logics with ultralow energy consumption have been designed. The artificial magnetoelectric exchange spring can be broadly applicable to various material choices, offering a versatile platform for developing energy-efficient and integrated magnonic devices.

cond-mat.mtrl-sci↗
arXiv2026

PlanViz: Evaluating Planning-Oriented Image Generation and Editing for Computer-Use Tasks

Unified multimodal models (UMMs) have shown impressive capabilities in complex natural image generation. However, their potential in computer-use planning tasks, which are closely related to our lives, remains rarely explored. Image generation and editing in computer-use tasks require capabilities like fine-grained perception and procedural understanding, and it is still unknown whether UMMs can finish these tasks well. Therefore, we propose PlanViz, a new benchmark designed to evaluate image generation and editing for computer-use tasks. We focus on sub-tasks which frequently involve in daily life and require planning steps. Specifically, three representative sub-tasks are designed: route planning, work diagramming, and web\&UI displaying. Data quality is ensured by including large human efforts on question, reference and criteria annotation. Additionally, a task-adaptive score, PlanScore, is proposed for comprehensive and exact evaluation. PlanScore helps understanding the correctness, visual quality and efficiency of generated images. Through experiments, we highlight key limitations and opportunities for future research on this topic. Codes and supplementary materials are in https://github.com/lijunxian111/PlanViz.

cs.CV↗
arXiv2026

Architecture-Aware Reinforcement Learning Makes Sliding-Window Attention Competitive in Math Reasoning

The rapid progress of reasoning and agentic large language models (LLMs) has increased the demand for long-context inference, but self-attention (SA) scales quadratically with context length. To address this, we study SWARR (Sliding-Window Attention with Reinforced Adaptation for Math Reasoning), a practical recipe for adapting SWA models to mathematical reasoning. SWARR has two stages: (1) efficient conversion from a pretrained SA model to SWA with supervised fine-tuning (SFT), which avoids pretraining a new base model, and (2) policy adaptation with reinforcement learning (RL). We find that SWA still underperforms SA after SFT, and we hypothesize that this gap is caused in part by a data-architecture mismatch: most SFT data are prepared for SA models and may contain long-range dependencies that are difficult for SWA to model. Because on-policy RL optimizes self-generated trajectories under the SWA constraint, it can adapt trajectories to better match SWA. Experiments on mathematical reasoning benchmarks show that this recipe substantially narrows the gap between SWA and SA, recovering much of the accuracy lost during SWA conversion while preserving the efficiency benefits of linear-complexity attention. Our central contribution is the empirical finding that RL changes the conclusion one would draw from conversion and SFT alone about SWA's viability for math reasoning.

cs.AI↗
arXiv2026

WAM-TTT: Steering World-Action Models by Watching Human Play at Test Time

Steering robot foundation models (RFMs) toward new task variants or user-preferred behaviors remains challenging, often requiring additional robot demonstrations, task-specific fine-tuning, or long-context conditioning. We present WAM-TTT, a test-time training framework for steering world action models from raw human videos. Rather than treating human videos as trajectories to imitate, WAM-TTT absorbs them into a lightweight adaptive memory inside a frozen WAM through self-supervised video prediction. To make this memory useful for control, we introduce a meta-training stage that aligns human demonstrations with robot behaviors using paired human-robot data and a key--value memory reconstruction objective. At test time, only unlabeled human videos are required to adapt the memory, while the pretrained WAM remains frozen. This enables efficient and reusable steering without robot actions, human-side annotations, or task-specific fine-tuning, while preserving the generalization ability of the foundation model. Extensive experiments show that WAM-TTT consistently outperforms in-context human-video conditioning baselines across diverse manipulation tasks and generalization settings.

cs.RO↗
arXiv2026

RECAP: Reciprocal Sparse Attention and Hierarchical Caching for Efficient Video Diffusion

Diffusion transformers generate high-quality videos but remain computationally expensive. Sparse attention and temporal caching reduce computation within and across denoising steps, respectively, offering a natural composition. However, predictions produced with sparse attention become sources for temporal reuse, coupling their errors along the sampling trajectory. We observe that improving the accuracy of these sources recovers fidelity without changing the reuse schedule, and that sparsification errors remain directionally aligned across nearby steps, suggesting that correction information can be shared rather than recomputed. These observations motivate RECAP, a training-free method that couples sparse attention with hierarchical caching through reciprocal correction, adaptively coordinated by online error feedback. Cached dense refinement corrects omitted interactions in current sparse estimates, while fresh sparse statistics track input changes and regulate the aging correction. The resulting predictions then renew the sources for whole-step prediction reuse, allowing prediction history to advance between dense refreshes. We evaluate RECAP on multiple video diffusion models against sparse-attention methods, caching methods, and their direct compositions. On HunyuanVideo at 720p 5-second generation, RECAP achieves a 3.05x speedup and outperforms all tested direct compositions in both fidelity and inference speed.

cs.CV↗
arXiv2026

OMEGA-Avatar: One-shot Modeling of 360° Gaussian Avatars

Creating high-fidelity, animatable 3D avatars from a single image remains a formidable challenge. We identified three desirable attributes of avatar generation: 1) the method should be feed-forward, 2) model a 360° full-head, and 3) should be animation-ready. However, current work addresses only two of the three points simultaneously. To address these limitations, we propose OMEGA-Avatar, the first feed-forward framework that simultaneously generates a generalizable, 360°-complete, and animatable 3D Gaussian head from a single image. Starting from a feed-forward and animatable framework, we address the 360° full-head avatar generation problem with two novel components. First, to overcome poor hair modeling in full-head avatar generation, we introduce a semantic-aware mesh deformation module that integrates multi-view normals to optimize a FLAME head with hair while preserving its topology structure. Second, to enable effective feed-forward decoding of full-head features, we propose a multi-view feature splatting module that constructs a shared canonical UV representation from features across multiple views through differentiable bilinear splatting, hierarchical UV mapping, and visibility-aware fusion. This approach preserves both global structural coherence and local high-frequency details across all viewpoints, ensuring 360° consistency without per-instance optimization. Extensive experiments demonstrate that OMEGA-Avatar achieves state-of-the-art performance, significantly outperforming existing baselines in 360° full-head completeness while robustly preserving identity across different viewpoints.

cs.GR↗
arXiv2026

GRASP: Plan-Guided Graph Retrieval with Adaptive Fusion and Reranking on Semi-Structured Knowledge Bases

Retrieval over semi-structured knowledge bases (SKBs) must handle both textual descriptions and relational constraints. Text-based retrievers capture the overall semantics of a query but struggle with multi-hop relations, whereas graph-based retrievers follow relational paths precisely but miss the broader meaning. We propose GRASP, a hybrid retriever that learns to fuse the two. A small LLM planner translates the query into an executable graph plan in a single call, its graph ranking is fused with a text ranking, and a fine-tuned reranker reorders the fused candidates. Instead of training the planner to be accurate on its own, we optimize it with reinforcement learning to maximize the gain its graph ranking brings to the fused ranking. On three STaRK benchmarks, GRASP achieves the best results on all metrics, improving the average Hit@1 by 18.6 points over the strongest baseline, while taking only 2.47 seconds per query.

cs.IR↗
arXiv2026

TwinPrune: Density-Aware Two-Phase Token Pruning for Vision-Language Models

We propose a unified spatial-density signal to score token importance for both inference phases: bipartite merge for prefill and KV eviction for decode, yielding measurable gains in each stage. To isolate the two phases, we split benchmarks by scoring logic. Prefill probes (MMBench, POPE, ScienceQA) evaluate only the first decoded token; their accuracy reflects prefill logits argmax before any decode KV eviction. Full-pipeline probes (TextVQA, DocVQA, MM-Vet) assess the complete generated sequence and capture both phases. We identify a critical evaluation pitfall: decode pruning claims validated solely on prefill probes measure prefill quality, not decode performance. Experiments show discarding all visual KV during decode barely affects prefill-probe accuracy but degrades TextVQA and MM-Vet drastically; this insensitivity stems from the scoring rule, not compression intensity. Evaluated on seven model variants (InternVL-3.5, Qwen3-VL, LLaVA-1.5) and three full-pipeline benchmarks, our method achieves a three-way Pareto frontier across compute, memory and accuracy. Density-based prefill compression maintains strong accuracy, scales linearly with visual token count, and delivers 100x faster token selection at high resolution. For decode, density top-k KV eviction consistently outperforms strong baselines H2O, SnapKV and StreamingLLM on TextVQA and MM-Vet across three retention ratios. At 0.125 keep ratio on TextVQA, it surpasses H2O by 9.6 points; InternVL-3.5 models (2B to 38B) maintain a 7 to 10 point advantage. Compression scales favorably with model size: every measured 8B to 38B model transition reduces accuracy loss instead of worsening it. We suggest MM-Vet or DocVQA as mandatory full-pipeline validation for any decode-phase claims. Our code is available at https://github.com/Kai-Liu001/TwinPrune

cs.CV↗
arXiv2026

Intermediate memory costs of quantum channel synthesis with a protected subsystem

A quantum memory can preserve a subsystem perfectly while discarding information needed by a later operation. We study the total intermediate dimension required for channel synthesis when the first stage protects a $b$-dimensional subsystem and has a prescribed noncommutative graph imposing full confusability on an $a$-dimensional control register. For a finite family of classically controlled branches, exact synthesis is possible if and only if the target graph contains the prescribed graph of the first stage. For every feasible family, we construct a single processor that implements all branches exactly. The processor uses pairwise nonorthogonal pure program states in a register of dimension at most $a$. A target graph may be realizable with a carrier of dimension $b$, while synthesizing a channel with that graph can require dimension $ab$. We also characterize the closure of the attainable channels, distinguishing exact synthesis from arbitrarily accurate approximation. Finally, for a family built from the trine communication task, a four-dimensional carrier permits exact synthesis with quantum programs under output depolarization. At the same total carrier dimension and without free shared randomness, classical programs have optimal normalized diamond error $(1-η)/4$, where $0\leqη\leq1$ is the depolarization strength. For $η<1$, the gap persists under sufficiently small perturbations of the target branches.

quant-ph↗
arXiv2026

Ultrasmall (3,0) carbon nanotube: a superconducting unit with strong electron-phonon coupling

Based on density-functional first-principles calculations, we study the electronic structures, lattice dynamics, and electron-phonon coupling (EPC) properties of the ultrasmall one-dimensional (3,0) carbon nanotube (CNT). Our studies reveal that the (3,0) CNT hosts characteristic collective phonon modes, in which the corresponding atomic vibrations can significantly modulate the electronic self-consistent potential and induce strong EPC. Our results further show that the (3,0) CNT exhibits robust phonon-mediated superconductivity under ambient pressure, with a superconducting transition temperature reaching $\sim$ 33 K ($μ^*$ = 0.10). We also find that although hole doping enhances the density of states (DOS) near the Fermi level, it simultaneously hardens the strongly coupled phonon modes and suppresses the total EPC. This result indicates that phonons play a more crucial role in the superconductivity of ultrasmall CNTs rather than the DOS alone. Therefore, we suggest that the ultrasmall (3,0) CNT can serve as a superconducting unit hosting phonon modes with strong EPC, and propose a strategy for discovering new phonon-mediated superconductors by identifying superconducting units.

cond-mat.supr-con↗
arXiv2026

FrameLite: Decision-Oriented Aesthetic Cropping with a Compact Vision-Language Model

Aesthetic image cropping improves composition by balancing subjects, context, and image boundaries. Large vision-language models bring semantic knowledge to this task, but their scale limits lightweight photography assistance. We propose FrameLite, a compact cropping model built on Qwen3.5-0.8B. Our approach is motivated by the observation that a small model can generate better crops than its first answer suggests. We develop three-stage training to turn this potential into stronger direct predictions. Spatial and composition training establishes localization and framing preferences. Quality-Guided Online Supervision (QOS) then learns the highest-quality complete answers among the current policy's generated candidates. Greedy-Referenced Optimization (GRO) further refines the policy using positive and negative quality differences relative to its current greedy answer. Both online stages use annotation-derived feedback, while inference returns one coordinate answer without search or an external scorer. Experiments show that FrameLite outperforms the compared large-VLM croppers across public benchmarks with a substantially smaller backbone. Controlled studies confirm that selecting high-quality training answers improves the final model beyond online sampling alone. We will release our code and model weights at https://github.com/JianzeLi-114/FrameLite .

cs.CV↗
arXiv2026

KV-Skill: Forging Expertise in the Model's Native Language

Task knowledge is commonly stored either as text in the prompt or as an update to model weights. Text is modular but must be interpreted on every use, while weight adaptation makes the resulting capability difficult to load, remove, or share independently. We introduce KV-Skill, a design space of external factorized operators that a frozen language model reads through a lightweight interface. KV-Skill supports two complementary paths. Registration converts an authored text skill into a text-derived operator and trains a shared per-backbone interface. Reward learning develops a compact latent operator directly from task outcomes, with or without an authored skill. Neither path adds positions to the prompt. Across ten benchmarks and four backbones from three model families, converting text to a KV-Skill consistently makes the same procedural knowledge more effective on mathematical reasoning and retrieval tasks. On Qwen3.5-4B LiveMath, registration reaches 77.2 accuracy, compared with 52.0 for the same skill as text (SkillOpt) and 64.5 for SoftSkill. Under matched reward training and parameter budgets, KV-Skill gives the best result in seven of eight matched settings against soft prefixes, prefix tuning, and LoRA. A post-hoc rank analysis further shows that text-derived operators retain nearly all of their benefit with one task-aligned direction per injection layer, while matched random directions fail. Finally, a single shared interface serves eleven diverse skills at once with no loss in average accuracy. These results show that task knowledge can be acquired from text or experience, compressed into an external operator, and deployed separately from the backbone. Code is available at: https://github.com/shawnzhg/KV-Skill

cs.LG↗
arXiv2026

Vesta: Queueing-Aware LLM Scheduling for Real-World Workloads

Agentic applications generate multi-turn sequences of LLM requests, with each new turn extending the context accumulated from prior interactions and tool calls. Prefix caching reuses the KV cache of earlier turns and avoids redundant prefill, but routing every turn to its cache-affine instance can create severe local queues while capacity remains idle elsewhere. Remote KV cache access relaxes this tension, yet migrating execution adds transfer work and is harmful when the cluster is already congested. Thus, deciding where to execute a request requires reasoning jointly about cache locality, queueing, and the current operating point of the cluster. We present Vesta, a queueing-aware scheduler for distributed agentic LLM serving. Vesta assigns each session a stable holder for its authoritative KV cache while independently selecting an executor for every request. Grounded in classical queueing theory, its objective accounts for backlog, reusable prefixes, KV cache transfer, and the queueing externality of additional service demand; it migrates execution to recover idle capacity when beneficial and favors placements that add the least prefill and transfer work as congestion rises. Vesta also migrates reusable prefixes across length buckets to avoid cold starts. On real-world agentic workloads, Vesta improves throughput by 14.0\% and reduces mean TTFT by 57.8\% over the state-of-the-art scheduler in a PD-colocated deployment. In PD-disaggregated deployments, it reduces mean TTFT by up to 18.1\% and P99 TTFT by up to 29.0\% at matched throughput.

cs.DC↗
arXiv2026

What Visual Representations Do World Action Models Need?

World Action Models (WAMs) inherit powerful priors from pretrained video generation models, but their variational autoencoder (VAE) representations are optimized for visual reconstruction rather than manipulation. External visual representations can improve manipulation, but existing methods couple representation choices with different integration mechanisms, obscuring their individual effects. We systematically compare visual representations, five integration mechanisms, and feature combinations under controlled settings. Our study yields three main findings. First, strong general-purpose visual representations provide broader robustness gains and can match or outperform representations pretrained for specific visual capabilities. Second, channel-wise joint prediction achieves the highest average success rate; further analyses suggest that future-feature prediction and the action expert's effective use of added visual information contribute to its gains. Third, additional sources should be selected for the task-relevant information they contribute beyond existing representations, as the same source can help one combination but harm another. However, even the best evaluated composition provides a smaller additional gain than the initial improvement from adding a strong general-purpose representation to the VAE-only baseline. We instantiate the resulting Representation Recipe as R$^2$-WAM. It achieves state-of-the-art performance on RoboCasa GR1 Tabletop with an 84.7% success rate and reaches 92.8% on RoboTwin 2.0, improving over the VAE-only baselines by 5.9 and 2.2 percentage points, respectively. Taken together, these findings suggest that future work should prioritize stronger general-purpose visual representations and more effective mechanisms for integrating them into WAMs, with feature composition serving as a supplementary route to improvement.

cs.CV↗
arXiv2026

Electron- and hole-doped superconductivity in two-dimensional electride ZrX$_2$ (X = Cl, Br, I)

Two-dimensional (2D) superconducting systems have recently emerged as a frontier in condensed matter physics, driven by their unique physical properties and potential applications. Using first-principles calculations, we show that monolayer ZrX$_2$ (X = Cl, Br, I), which is a family of 2D semiconducting electrides, hosts phonon-mediated superconductivity under both electron and hole dopings. Electron doping in ZrX$_{2}$ can achieve high superconducting transition temperatures at very low concentrations, with ZrBr$_2$ reaching $T_c = 30.0$ K at 0.01 \textit{e}/f.u., followed by ZrCl$_2$ (22.5 K at 0.05 \textit{e}/f.u.) and ZrI$_2$ (18.1 K at 0.01 \textit{e}/f.u.), for which the high $T_c$ originates from the flat conduction band minimum, with Van Hove singularities nearby enhancing the density of states. In contrast, hole doping requires higher concentrations (0.2--0.25 \textit{h}/f.u.) to activate superconductivity, yielding $T_c \sim 2$ K, and approaches a magnetic instability at elevated doping levels. By comparing with InSe and LaBr$_2$, we identify the spatial location and orbital character of excess electrons as key factors determining the strength of hole-doped superconductivity and the nature of competing orders. These findings establish monolayer ZrX$_2$ as a platform for bipolar superconductivity in a 2D electride and provide insights into flat-band engineering for high-$T_c$ superconductors.

cond-mat.supr-con↗