arXiv subjects

Zihan Wang

Publications and source records attributed to Zihan Wang.

At least 127 records · Page 7Linked to original sources

arXiv2026

Purifying Multimodal Retrieval: Fragment-Level Evidence Selection for RAG

Multimodal Retrieval-Augmented Generation (MRAG) is widely adopted for Multimodal Large Language Models (MLLMs) with external evidence to reduce hallucinations. Despite its success, most existing MRAG frameworks treat retrieved evidence as indivisible documents, implicitly assuming that all content within a document is equally informative. In practice, however, sometimes only a small fraction of a document is relevant to a given query, while the remaining content introduces substantial noise that may lead to performance degradation. We address this fundamental limitation by reframing MRAG as a fine-grained evidence selection problem. We propose Fragment-level Evidence Selection for RAG (FES-RAG), a framework that selects atomic multimodal fragments rather than entire documents as grounding evidence. FES-RAG decomposes retrieved multimodal documents into sentence-level textual fragments and region-level visual fragments, enabling precise identification of evidence that directly supports generation. To guide fragment selection, we introduce Fragment Information Gain (FIG), a principled metric that measures the marginal contribution of each fragment to the MLLM's generation confidence. Based on FIG, we distill fragment-level utility judgments from a high-capacity MLLM into a lightweight selector, achieving accurate evidence selection with low inference overhead. Experiments on the M2RAG benchmark show that FES-RAG consistently outperforms state-of-the-art document-level MRAG methods, achieving up to 27 percent relative improvement in CIDEr. By selecting fewer yet more informative fragments, our approach substantially reduces context length while improving factual accuracy and generation coherence.

cs.IR↗

arXiv2026

Learning from Medical Entity Trees: An Entity-Centric Medical Data Engineering Framework for MLLMs

Multimodal Large Language Models (MLLMs) have shown transformative potential in medical applications, yet their performance is hindered by conventional data curation strategies that rely on coarse-grained partitioning by modality or department. Such fragmented approaches fail to capture the hierarchical and interconnected nature of clinical medical knowledge, limiting the models' ability to perform fine-grained recognition and complex reasoning. In this paper, we propose a novel Entity-Centric Medical Data Engineering framework. We automatically extract entities from authoritative medical literature to construct a Medical Entity Tree (MET), a hierarchical structure that systematically encodes diseases, anatomical structures, modalities, and symptoms into a unified knowledge repository. Building upon the MET, we propose an advanced data engine that includes: (1) node-guided retrieval to anchor raw data to specific medical concepts, (2) a two-stage hybrid filtering and alignment pipeline to ensure precise visual-semantic correspondence, and (3) knowledge-aware data synthesis to generate enriched captions and targeted reasoning VQA pairs, leveraging structural constraints. Extensive evaluations across six medical benchmarks demonstrate that our approach significantly enhances the medical capabilities of general-purpose MLLMs, improving their ability to handle complex clinical queries and achieve state-of-the-art performance in diverse medical contexts.

cs.CL↗

arXiv2026

Black-Box Skill Stealing Attack from Proprietary LLM Agents: An Empirical Study

Large language model (LLM) agents increasingly rely on skills to package reusable capabilities through instructions, tools, and resources. High-quality skills embed expert knowledge, curated workflows, and execution constraints into agents, fueling a growing skill economy through their value and scalability. Yet this ecosystem also creates a new attack surface, as adversaries can interact with public agent interfaces to extract hidden proprietary skill content. We present the first systematic study of black-box skill stealing against LLM agent systems. Compared with conventional system prompt stealing, skill stealing targets modular and structured capability packages whose leakage is directly actionable for copying, redistribution, and monetization, making the resulting harm potentially greater. To study this threat, we derive an attack taxonomy from prior prompt-stealing methods and build an automated stealing prompt generation agent. Starting from model-generated seed prompts, the framework expands attacks through scenario rationalization and structure injection while enforcing diversity via embedding-based filtering, yielding a reproducible pipeline for evaluating proprietary agent systems. We evaluate these attacks across commercial agent platforms and representative LLMs. Our results show that agent skills can often be extracted easily, posing a serious copyright risk. To mitigate this threat, we design defenses across the agent pipeline, focusing on input, inference, and output phase. Although these defenses substantially reduce leakage, the attack remains inexpensive and repeatable, and a single successful attempt is sufficient to compromise the protected skill. Overall, our findings suggest that these copyright risks remain largely overlooked across proprietary agent ecosystems, motivating stronger protection mechanisms.

cs.CR↗

arXiv2026

MTRouter: Cost-Aware Multi-Turn LLM Routing with History-Model Joint Embeddings

Multi-turn, long-horizon tasks are increasingly common for large language models (LLMs), but solving them typically requires many sequential model invocations, accumulating substantial inference costs. Here, we study cost-aware multi-turn LLM routing: selecting which model to invoke at each turn from a model pool, given a fixed cost budget. We propose MTRouter, which encodes the interaction history and candidate models into joint history-model embeddings, and learns an outcome estimator from logged trajectories to predict turn-level model utility. Experiments show that MTRouter improves the performance-cost trade-off: on ScienceWorld, it surpasses GPT-5 while reducing total cost by 58.7%; on Humanity's Last Exam (HLE), it achieves competitive accuracy while reducing total cost by 43.4% relative to GPT-5, and these gains even carry over to held-out tasks. Further analyses reveal several mechanisms underlying its effectiveness: relative to prior multi-turn routers, MTRouter makes fewer model switches, is more tolerant to transient errors, and exhibits emergent specialization across models. Code: https://github.com/ZhangYiqun018/MTRouter

cs.CL↗

arXiv2026

EMCompress: Video-LLMs with Endomorphic Multimodal Compression

Video-LLMs face a fundamental tension in long-video reasoning: static, sparse frame sampling either dilutes evidence across task-irrelevant segments at significant cost or misses fine-grained temporal semantics altogether. We propose a novel, cognitively-inspired task -- Endomorphic Multimodal Compression (EMC) -- as a structurally-constrained sufficient-statistic problem for VideoQA, and formulate it as an endomorphic transformation F_EMC : (V, Q) -> (v, q) that compresses the multimodal input while preserving answer invariance across reasonable downstream models. The endomorphic form keeps the compressed output in the downstream pipeline's native task space -- a structural mirror of the filter-then-reason mechanism in the cognitive literature motivating EMC -- distinguishing it from latent-code compression (IB / VIB) and making the formulation extensible to other multimodal settings. Under the Markov chain A -> (V, Q) -> (v, q), EMC realizes the classical sufficiency condition I((v, q); A) = I((V, Q); A) in its VideoQA-natural form. As a modular front-end, EMC plugs into both Video Instruction Tuning and Video Question Answering pipelines. We release the first dedicated benchmark and propose ReSimplifyIt, an EMC baseline surpassing prior methods by 0.40 F-1 with competitive query rewriting. Integrating EMC yields relative gains of 7.33% in training and 33.7% in inference for video-language understanding.

cs.CV↗

arXiv2026

Breaking the UV Luminosity Function Degeneracy:Self-Interacting Dark Matter Constraints from Reionization Topology

Self-interacting dark matter (SIDM) is the leading framework resolving small-scale cold dark matter (CDM) crises, yet high-redshift SIDM constraints are fundamentally limited by degeneracies between dark matter microphysics and galaxy formation astrophysics. We demonstrate that the UV luminosity function alone cannot constrain SIDM: star formation suppression from SIDM halo core formation is fully absorbed by modest adjustments to standard astrophysical parameters. We show that 21 cm reionization topology breaks this degeneracy completely, providing a nuisance-immune probe: the SIDM-enhanced duty cycle of ionizing photon escape leaves a morphological signature fully independent of star formation efficiency. Combining JWST UVLF measurements with SKA1-Low forecasts, constant-cross-section SIDM with $σ/m \gtrsim 1$--$2\ \mathrm{cm^2/g}$ is either excluded or detectable across all physically motivated star formation coupling strengths. Our results establish a robust new avenue to probe dark matter microphysics in the early Universe.

astro-ph.CO↗

arXiv2026

VEFX-Bench: A Holistic Benchmark for Generic Video Editing and Visual Effects

As AI-assisted video creation becomes increasingly practical, instruction-guided video editing has become essential for refining generated or captured footage to meet professional requirements. Yet the field still lacks both a large-scale human-annotated dataset with complete editing examples and a standardized evaluator for comparing editing systems. Existing resources are limited by small scale, missing edited outputs, or the absence of human quality labels, while current evaluation often relies on expensive manual inspection or generic vision-language model judges that are not specialized for editing quality. We introduce VEFX-Dataset, a human-annotated dataset containing 5,049 video editing examples across 9 major editing categories and 32 subcategories, each labeled along three decoupled dimensions: Instruction Following, Rendering Quality, and Edit Exclusivity. Building on VEFX-Dataset, we propose VEFX-Reward, a reward model designed specifically for video editing quality assessment. VEFX-Reward jointly processes the source video, the editing instruction, and the edited video, and predicts per-dimension quality scores via ordinal regression. We further release VEFX-Bench, a benchmark of 300 curated video-prompt pairs for standardized comparison of editing systems. Experiments show that VEFX-Reward aligns more strongly with human judgments than generic VLM judges and prior reward models on both standard IQA/VQA metrics and group-wise preference evaluation. Using VEFX-Reward as an evaluator, we benchmark representative commercial and open-source video editing systems, revealing a persistent gap between visual plausibility, instruction following, and edit locality in current models. Our project page is https://xiangbogaobarry.github.io/VEFX-Bench/.

cs.CV↗

arXiv2026

The First Challenge on Mobile Real-World Image Super-Resolution at NTIRE 2026: Benchmark Results and Method Overview

This paper provides a review of the NTIRE 2026 challenge on mobile real-world image super-resolution, highlighting the proposed solutions and the resulting outcomes. The challenge aims to recover high-resolution (HR) images from low-resolution (LR) counterparts generated through unknown degradations with a x4 scaling factor while ensuring the models remain executable on mobile devices. The objective is to develop effective and efficient network designs or solutions that achieve state-of-the-art real-world image super-resolution performance. The track of the challenge evaluates performance using a weighted combination of image quality assessment (IQA) score and speedup ratios. The competition attracted 108 registrants, with 16 teams achieving a valid score in the final ranking. This collaborative effort advances the performance of mobile real-world image super-resolution while offering an in-depth overview of the latest trends in the field.

cs.CV↗

arXiv2026

Reionization Topology as a Probe of Self-Interacting Dark Matter

The topology of cosmic reionization, the sizes, shapes, and connectivity of ionized bubbles is a primary observable of next-generation 21\,cm experiments. We show that this topology is sensitive to the microphysics of dark matter. Self-interacting dark matter (SIDM), with cross-sections $σ/m\sim 1$--$10\;\mathrm{cm^2/g}$ motivated by small-scale structure anomalies, reduces halo gas binding energies and increases the duty cycle of ionizing-photon escape. At fixed global neutral fraction $\bar{x}_{\rm HI}$, this reshapes the source population from rare, very bright emitters to more numerous, moderate emitters, producing qualitatively different ionization morphology. We decompose the effect into two scale-dependent levers: a $2$--$3\%$ emissivity-weighted bias shift at $k\lesssim 0.1\;h/\mathrm{Mpc}$, and a factor $2$--$4$ shot-noise suppression at $k\sim 0.1$--$1\;h/\mathrm{Mpc}$. A halo-by-halo semi-numerical simulation at $128^3$ resolution confirms a $\sim 60$--$70\%$ increase in the Euler characteristic of the ionization field for $σ/m \gtrsim 2\;\mathrm{cm^2/g}$, detected at $3.8σ$ across ten independent realizations. A blowout model connecting the binding-energy reduction to the duty cycle through the ISM column density distribution yields a detection threshold at $σ/m \sim 1$--$2\;\mathrm{cm^2/g}$. The signal exceeds the CDM baryonic uncertainty band and is robust to the functional form of the emissivity parametrization. The signal persists even if gravitational heating offsets $50$--$75\%$ of the blowout enhancement, and is not diluted by unresolved low-mass sources. Velocity-dependent SIDM produces a qualitatively distinct opposite-sign bias shift. These predictions are testable with SKA1-Low, establishing reionization as a new arena for probing dark matter models complementary to dwarf galaxies and galaxy clusters.

astro-ph.CO↗

arXiv2026

ChartNet: A Million-Scale, High-Quality Multimodal Dataset for Robust Chart Understanding

Understanding charts requires models to jointly reason over geometric visual patterns, structured numerical data, and natural language -- a capability where current vision-language models (VLMs) remain limited. We introduce ChartNet, a high-quality, million-scale multimodal dataset designed to advance chart interpretation and reasoning. ChartNet leverages a novel code-guided synthesis pipeline to generate 1.5 million diverse chart samples spanning 24 chart types and 6 plotting libraries. Each sample consists of five aligned components: plotting code, rendered chart image, data table, natural language summary, and question-answering with reasoning, providing fine-grained cross-modal alignment. To capture the full spectrum of chart comprehension, ChartNet additionally includes specialized subsets encompassing human annotated data, real-world data, safety, and grounding. Moreover, a rigorous quality-filtering pipeline ensures visual fidelity, semantic accuracy, and diversity across chart representations. Fine-tuning on ChartNet consistently improves results across benchmarks, demonstrating its utility as large-scale supervision for multimodal models. As the largest open-source dataset of its kind, ChartNet aims to support the development of foundation models with robust and generalizable capabilities for data visualization understanding. The dataset is publicly available at https://huggingface.co/datasets/ibm-granite/ChartNet

cs.CV↗

arXiv2026

Dark QCD Origin of the NANOGrav Signal and Self-Interacting Dark Matter

The NANOGrav 15-year stochastic gravitational wave background (SGWB) amplitude $A_{\rm yr} \approx 2.4 \times 10^{-15}$ lies at the upper edge of population synthesis predictions for supermassive black hole binaries (SMBHBs), motivating exploration of additional cosmological sources. We present a phenomenological framework based on an $\text{SU}(3)_D$ gauge theory that can simultaneously accommodate the gravitational wave signal and resolve small-scale structure anomalies via Self-Interacting Dark Matter (SIDM). The dark matter candidate is a heavy dark baryon $χ= QQQ$ with mass $m_χ\approx 40$~GeV, which self-interacts through a light pseudo-dilaton $d$ $m_d \approx 20$--$50$~MeV as a pseudo-Goldstone boson of approximate scale invariance arising in near-conformal gauge theories with $N_f \sim 6$--$8$ light flavors. A first-order phase transition at the MeV scale, enabled by walking dynamics near the conformal window, produces gravitational waves in the PTA band. For representative parameters $T_n \approx 5$--$6$~MeV, $α\sim 500$--$1000$, $β/H_* \sim 30$--$50$, the model provides a fit to NANOGrav data comparable to SMBHB while naturally connecting the gravitational wave amplitude to the dark matter relic density through entropy dilution $D \approx α^{3/4}$. We present explicit calculations of the bounce action, bubble wall velocity, and $ΔN_{\rm eff}$, demonstrating that the benchmark parameters are theoretically consistent and cosmologically safe ($ΔN_{\rm eff} \lesssim 0.1$ for $m_π> 2m_d$). The distinctive spectral shape ($f^3 \to f^{-4}$) provides a robust prediction testable with future PTAs.

astro-ph.CO↗

arXiv2026

Precise Aggressive Aerial Maneuvers with Sensorimotor Policies

Precise aggressive maneuvers with lightweight onboard sensors remains a key bottleneck in fully exploiting the maneuverability of drones. Such maneuvers are critical for expanding the systems' accessible area by navigating through narrow openings in the environment. Among the most relevant problems, a representative one is aggressive traversal through narrow gaps with quadrotors under SE(3) constraints, which require the quadrotors to leverage a momentary tilted attitude and the asymmetry of the airframe to navigate through gaps. In this paper, we achieve such maneuvers by developing sensorimotor policies directly mapping onboard vision and proprioception into low-level control commands. The policies are trained using reinforcement learning (RL) with end-to-end policy distillation in simulation. We mitigate the fundamental hardness of model-free RL's exploration on the restricted solution space with an initialization strategy leveraging trajectories generated by a model-based planner. Careful sim-to-real design allows the policy to control a quadrotor through narrow gaps with low clearances and high repeatability. For instance, the proposed method enables a quadrotor to navigate a rectangular gap at a 5 cm clearance, tilted at up to 90-degree orientation, without knowledge of the gap's position or orientation. Without training on dynamic gaps, the policy can reactively servo the quadrotor to traverse through a moving gap. The proposed method is also validated by training and deploying policies on challenging tracks of narrow gaps placed closely. The flexibility of the policy learning method is demonstrated by developing policies for geometrically diverse gaps, without relying on manually defined traversal poses and visual features.

cs.RO↗

arXiv2026

Time Series Gaussian Chain Graph Models

Time series graphical models have recently received considerable attention for characterizing (conditional) dependence structures in multivariate time series. In many applications, the multivariate series exhibit variable-partitioned blockwise dependence, with distinct patterns within and across blocks. In this paper, we introduce a new class of time series Gaussian chain graph models that represent contemporaneous and lagged causal relations via directed edges across blocks, while capturing within-block conditional dependencies through undirected edges. In the frequency domain, this formulation induces a cross-frequency shared group sparse plus group low-rank decomposition of the inverse spectral density matrices, which we exploit to establish identifiability of the time series chain graph structure. Building on this, we then propose a three-stage learning procedure for estimating the undirected and directed edge sets, which involves optimizing a regularized Whittle likelihood with a group lasso penalty to encourage group sparsity and a novel tensor-unfolding nuclear norm penalty to enforce group low-rank structure. We investigate the asymptotic properties of the proposed method, ensuring its consistency for exact recovery of the chain graph structure. The superior empirical performance of the proposed method is demonstrated through both extensive simulation studies and an application to U.S. macroeconomic data that highlights key monetary policy transmission mechanisms.

stat.ME↗

arXiv2026

RAGEN-2: Reasoning Collapse in Agentic RL

RL training of multi-turn LLM agents is inherently unstable, and reasoning quality directly determines task performance. Entropy is widely used to track reasoning stability. However, entropy only measures diversity within the same input, and cannot tell whether reasoning actually responds to different inputs. In RAGEN-2, we find that even with stable entropy, models can rely on fixed templates that look diverse but are input-agnostic. We call this template collapse, a failure mode invisible to entropy and all existing metrics. To diagnose this failure, we decompose reasoning quality into within-input diversity (Entropy) and cross-input distinguishability (Mutual Information, MI), and introduce a family of mutual information proxies for online diagnosis. Across diverse tasks, mutual information correlates with final performance much more strongly than entropy, making it a more reliable proxy for reasoning quality. We further explain template collapse with a signal-to-noise ratio (SNR) mechanism. Low reward variance weakens task gradients, letting regularization terms dominate and erase cross-input reasoning differences. To address this, we propose SNR-Aware Filtering to select high-signal prompts per iteration using reward variance as a lightweight proxy. Across planning, math reasoning, web navigation, and code execution, the method consistently improves both input dependence and task performance.

cs.LG↗

arXiv2026

LensAgent: A Self Evolving Agent for Autonomous Physical Inference of Sub-galactic Structure

Probing dark matter distribution on sub-galactic scales is essential for testing the Cold Dark Matter ($Λ$CDM) paradigm. Strong gravitational lensing, as one of the most powerful approach by far, provides a direct, purely gravitational probe of these substructures. However, extracting cosmological constraints is severely bottlenecked by the mass-sheet degeneracy (MSD) and the unscalable nature of manual and neural-network modeling. Here, we introduce LensAgent, a pioneering training-free, large language model (LLM)-driven agentic framework for the autonomous physical inference of mass distributions. Operating as an autonomous scientific agent, LensAgent couples high-level logical reasoning with deterministic physical modeling tools, demonstarting successful reconstruction of mass distribution in SLACS Grade A strong lensing systems. This self-evolving architecture enables the robust extraction of sub-galactic substructures at scale, unlocking the cosmological potential of upcoming wide-field surveys such as the Rubin Observatory (LSST) and Euclid.

astro-ph.GA↗

arXiv2026

WFR-FM: Simulation-Free Dynamic Unbalanced Optimal Transport

The Wasserstein-Fisher-Rao (WFR) metric extends dynamic optimal transport (OT) by coupling displacement with change of mass, providing a principled geometry for modeling unbalanced snapshot dynamics. Existing WFR solvers, however, are often unstable, computationally expensive, and difficult to scale. Here we introduce WFR Flow Matching (WFR-FM), a simulation-free training algorithm that unifies flow matching with dynamic unbalanced OT. Unlike classical flow matching which regresses only a transport vector field, WFR-FM simultaneously regresses a vector field for displacement and a scalar growth rate function for birth-death dynamics, yielding continuous flows under the WFR geometry. Theoretically, we show that minimizing the WFR-FM loss exactly recovers WFR geodesics. Empirically, WFR-FM yields more accurate and robust trajectory inference in single-cell biology, reconstructing consistent dynamics with proliferation and apoptosis, estimating time-varying growth fields, and applying to generative dynamics under imbalanced data. It outperforms state-of-the-art baselines in efficiency, stability, and reconstruction accuracy. Overall, WFR-FM establishes a unified and efficient paradigm for learning dynamical systems from unbalanced snapshots, where not only states but also mass evolve over time. The Python code is available at https://github.com/QiangweiPeng/WFR-FM.

cs.LG↗

arXiv2026

MindCube: Spatial Mental Modeling from Limited Views

Can Vision-Language Models (VLMs) imagine the full scene from just a few views, like humans do? Humans form spatial mental models naturally, internal representations of unseen space, to reason about layout, perspective, and motion. Our MindCube benchmark with 21,154 questions across 3,268 images exposes this critical gap, where existing VLMs exhibit near-random performance. Using MindCube, we systematically evaluate how well VLMs build robust spatial mental models through representing positions (cognitive mapping), orientations (perspective-taking), and dynamics (mental simulation for "what-if" movements). We then explore three approaches to help approximate spatial mental models in VLMs, focusing on incorporating unseen intermediate views, natural language reasoning chains, and cognitive maps. The significant improvement comes from a synergistic approach, "map-then-reason", that jointly trains the model to first generate a cognitive map and then reason upon it. By training models to reason over these internal maps, we boosted accuracy from 37.8% to 57.8% (+20.0%). Adding reinforcement learning pushed performance even further to 61.3% (+23.5%). Our key insight is that such scaffolding of spatial mental models, actively constructing and utilizing internal structured spatial representations with flexible reasoning processes, significantly improves understanding of unobservable space.

cs.AI↗

arXiv2026

On world-volume supersymmetry of supermembrane action in static gauge

We review and elaborate on the issue of 3d world volume supersymmetry that appears as a residual part of global target space supersymmetry in the BST supermembrane action. While there is no direct ``spinning membrane'' analog of the world-volume supersymmetric spinning string action that could be obtained by coupling $D$ copies of 3d scalar multiplet to 3d supergravity, we discuss how one may construct an $N=1$ 3d supersymmetric analog of the derivative expansion of the bosonic membrane action in static gauge. We compare the resulting $N=1$ supersymmetric action for eight 3d scalar multiplets to the $N=8$ 3d supersymmetric action describing the $D=11$ supermembrane in the static gauge. The two actions are not equivalent which is related to the fact that the full $N=8$ supersymmetry of the static-gauge $D=11$ supermembrane action can be realised only if the fermions are described by an $SO(8)$ spinor rather than vector. The two actions are still directly related in special dimensions $D=4$ and 5. We also compute the one-loop world-volume scattering amplitudes for the two theories, finding that they indeed agree for $D=4,5$ but disagree for $D=11$.

hep-th↗