arXiv subjects

Zihao Wang

Publications and source records attributed to Zihao Wang.

At least 19 recordsLinked to original sources

arXiv2026

Latent Visual Contrastive Reasoning with Multi-Branch Exploration

Multimodal large language models have made substantial progress in visual question answering, but complex visual reasoning often requires locating and comparing multiple pieces of question-relevant visual evidence. Existing methods that move intermediate reasoning into the visual latent space usually follow a single linear reasoning trajectory, which can be limiting when multiple visual hypotheses or competing evidence regions are relevant. We propose Latent Visual Contrastive Reasoning, a branch-aware latent reasoning framework that generates multiple question-conditioned latent visual branches as soft hypotheses over image tokens. To train these branches, our method uses bounding-box evidence annotations to build target evidence representations and introduces a soft contrastive branch-evidence alignment objective. This objective encourages evidence-relevant branches to align with annotated visual regions while contrasting them against negative evidence. We instantiate our method on Qwen2.5-VL backbones and train it with standard supervised fine-tuning. Experiments on perception-intensive visual reasoning benchmarks show that the proposed method achieves overall improvements over the corresponding backbone and single-path latent reasoning baselines at both 3B and 7B scales.

cs.CV↗

arXiv2026

CodeCap-Bench: Can Code Serve as a Good Video Caption?

Video captioning aims to preserve visual information through compact descriptions, predominantly using natural language. However, precisely expressing fine-grained spatial structures and temporal dynamics in natural language often leads to lengthy and potentially ambiguous descriptions. In this work, we explore Code as Video Caption, a paradigm where executable programs serve as a complementary representation of visual content through graphical primitives, layout rules, and functions of time. To examine whether code captions effectively preserve video information, we compare them with natural language captions on downstream question answering and video reconstruction tasks. Our results show that code captions achieve superior reasoning accuracy and reconstruction fidelity. Despite these advantages, realizing the potential of code captioning depends on whether current multimodal large language models (MLLMs) can generate faithful code captions. To systematically assess this capability, we introduce CodeCap-Bench, which comprises 323 videos spanning five categories of structured 2D content, along with fine-grained annotations and a multi-level evaluation framework. Extensive experiments reveal that while leading proprietary MLLMs demonstrate promising code captioning capabilities, they still struggle to faithfully capture complex temporal dynamics, with open-source models lagging substantially behind.

cs.CV↗

arXiv2026

LocalAlign: Enabling Generalizable Prompt Injection Defense via Generation of Near-Target Adversarial Examples for Alignment Training

Large language models are increasingly embedded into systems that interact with user data, retrieved web content, and external tools, creating a new attack surface: prompt injection, where malicious commands embedded in untrusted data override the trusted command and induce unintended behavior. Existing defenses mainly rely on fine-tuning the model to preserve an explicit boundary between trusted commands and the untrusted data portion, so that the model learns to prioritize the trusted field and ignore malicious commands in data. However, we observe that while these defenses can block obviously malicious responses caused by injected commands, they generalize poorly to real-world scenarios where the model's response to the injected command is much nearer to the correct response. This is because existing methods typically train against only a fixed set of hand-crafted attack targets, which yields a loose boundary around the correct response and leaves it easier to bypass. To address this challenge, we propose LocalAlign, a more generalizable prompt injection defense inspired by adversarial training. LocalAlign automatically and efficiently generates adversarial examples in which the command embedded in the data portion induces a response that stays near to the correct response while still being wrong. We generate such near-but-wrong adversarial examples using prompting and a single inference step. This design enforces a tighter robustness boundary around the correct response: even small response shifts induced by commands in untrusted data are explicitly penalized. Moreover, the resulting adversarial examples can vary substantially in quality across samples. To address this issue, we further introduce a margin-aware alignment algorithm that quantifies each sample's distance to the correct response and assigns larger training weight to nearer ones.

cs.CR↗

arXiv2026

On Task, Off Track: Benchmarking LLM and Agent Robustness to Near-Target Prompt Injection

Prompt injection benchmarks assess whether large language models (LLMs) and agents resist malicious instructions embedded in external content. However, a system could perform well by rejecting attacks that obviously abandon the user's request, while still accepting changes that retain most of a legitimate answer but make it incorrect or harmful. Testing more instructions for unrelated outcomes would not necessarily reveal this weakness. We call attacks seeking responses designed to resemble legitimate answers while introducing incorrect or harmful changes near-target prompt injection. We introduce NTBench, the Near-Target Benchmark, which builds on legitimate tasks from existing datasets. For each task, a legitimate reference outcome guides the generation of target--instruction pairs, with separate tracks targeting incorrect and harmful outcomes in responses or agent tool invocations. Harmful targets can also accompany otherwise correct task completion. We combine LLM-based checks with human validation of the intended deviations, then assess whether injected instructions induce them using response judgments or tool-call checks, excluding deviations already observed without injection. We evaluate seven LLMs on ten datasets for language tasks and two agent benchmarks. Under a common budget of three attack executions per sample, both tracks achieve higher overall attack success rates than the fixed-target baseline for every evaluated model in each setting. On language tasks, the increases range from 9.3 to 35.8 percentage points for incorrect targets and 1.7 to 17.3 percentage points for harmful targets. These results show that resistance to the tested fixed-target attacks does not imply resistance to related incorrect or harmful outcomes, and that reference-guided targets broaden the failures revealed by robustness evaluation.

cs.CR↗

arXiv2026

An Attribute-Based Measure of Video Complexity

A new framework for the estimation of the complexity posed by video-question pairs to video-LLMs, Video Attribute-Based Complexity (VideoABC), is proposed. Video complexity is defined as the probability of failure of a video-LLM for a given video-question pair. VideoABC is a non-parametric complexity measure, using a reference video dataset and a pre-defined vocabulary of video attributes informative of complexity, e.g. the scene complexity or the speed of the video event relevant to the question. In a training phase, reference videos are projected into the space of these attributes, which is then quantized. The expected ABC of each quantization cell is then computed. Given a new video and its projection into the attribute space, complexity is estimated by the expected ABC of the associated quantization cell. To enable the use of VideoABC with small reference video datasets, two quantizers are combined: a k-means quantizer that enables accurate complexity estimates for samples in the distribution of the reference dataset and a universal lattice quantizer that guarantees generalization to out-of-distribution samples. A synthetic video generation procedure, inspired by target-distractor manipulations of psychophysics studies, is proposed to populate the cells of the lattice quantizer during training, enabling the computation of their expected ABCs. Experimental results show that VideoABC is effective even with very low-dimensional attribute representations, substantially outperforming approaches like 'video-LLM as judge' with much less complexity. Finally, the explainable nature of the VideoABC score, in terms of well-defined attributes, is shown to provide insights on how the attribute composition of benchmarks affects their complexity.

cs.CV↗

arXiv2026

One-Sided Verification Under Limited Expert Review

AI has become increasingly helpful in scientific discovery, yet human expert review remains essential for those problems without reliable automated evaluation, posing a fundamental challenge as candidate generation scales beyond available review capacity. This paper studies verification under limited human effort, where rare valid solutions need to be prioritized while invalid candidates should be filtered out to reduce human workload. We adopt the one-sided verification objective that maximizes the rejection rate of invalid solutions (TNR) under a constraint on the false-rejection rate of valid ones (FNR), making human review cost an explicit term to minimize. Drawing on the intuition that verifying a specific error can be easier than discovering one, we introduce OSVerify, an algorithmic framework that decouples error proposal from error validation. The error proposer searches for errors in solutions across different granularities by hierarchical decomposition, while the validator distinguishes substantial errors from minor or hallucinated ones. Experiments on competition- and research-level benchmarks show that OSVerify reduces the number of human reviews of invalid solutions by 46% on average compared to voting baselines under the same empirical low FNR constraint. Our post-training of the validator further improves the reduction to 57%. As high-impact real-world use cases, we apply OSVerify to identify errors in published research papers that are subsequently confirmed by the original authors and help resolve three Langlands open conjectures that are confirmed by experts, demonstrating its utility to facilitate human-AI research collaboration.

cs.AI↗

arXiv2026

HuMiT: Low-Latency Whole-Body Humanoid Teleoperation via Compact Reference Tracking

Real-time whole-body humanoid teleoperation remains challenging due to the computational latency introduced by optimization-based retargeting pipelines and future-frame conditioning. Existing motion tracking policies typically require dense reference targets including full joint positions, body-part transformations, and even future frames, which tightly couples the controller to expensive upstream modules. In this work, we present HuMiT, a whole-body teleoperation system built on a unified tracking policy that requires only a compact reference target, consisting of root height, root velocity, and the SE(3) poses of four sparse keypoints at the current frame, yet achieves competitive tracking fidelity compared to methods relying on more diverse reference states including complete joint positions, dense body-part transformations, and even future frames. Experiments show that our policy achieves a communication-to-inference latency of $\sim$20 ms measured between headset streaming and policy output and a video-based optical flow latency of 65 ms (median) across different motion types. We validate the system on a Unitree G1 humanoid, showcasing split-second reactive responses such as catching a tossed ball, as well as coordinated loco-manipulation capabilities. Project page: https://xiaohu-art.github.io/Track/.

cs.RO↗

arXiv2026

SCDM: Scalable Causal Discovery in Nonlinear Temporal Systems with Meta-Learning

Causal discovery from nonlinear multivariate time series is challenging in high-dimensional systems, where the number of candidate directed relations grows quadratically with the number of variables. Existing methods are often limited by target-wise model fitting, repeated conditional testing, or expensive graph-search procedures. We propose SCDM, a shared-parameter meta-learning framework for scalable temporal causal discovery. SCDM treats each target variable as a task while learning a shared temporal predictor and a shared causal-strength matrix across tasks. Each meta-episode updates only a subset of target-variable tasks, allowing the model to accumulate structural evidence across the full system without exhaustive target-wise optimization. After training, a post-training graph readout converts the learned predictor into continuous directed causal scores for threshold-independent evaluation. We also provide a finite-error recovery principle showing that, under fully observed delayed-system assumptions and a positive risk-gap condition, thresholding SCDM scores recovers the graph when statistical, meta-optimization, and readout errors are below the population margin. Experiments on controlled synthetic, neuroimaging-inspired, high-dimensional, and real-world temporal benchmarks demonstrate competitive causal recovery and improved scalability in large temporal systems.

cs.LG↗

arXiv2026

Weyl Bounds for $\operatorname{GL}_2\times \operatorname{GL}_2$ Rankin--Selberg $L$-functions via Short Spectral Moments

Let $f$ and $g$ be Hecke--Maass cusp forms of $\operatorname{SL}_2(\mathbb{Z})$, with spectral parameters $t_f$ and $t_g$, respectively. Let $T$ be a large parameter and put $L = |T-t_g|.$ We prove the Lindelöf-on-average estimate \[ \sum_{ |t_f-T|\leq H} \left|L\!\left(\frac12,f\otimes g\right)\right|^2 \ll_{\varepsilon} T^{1+\varepsilon}H, \] uniformly in the range \[ T^{1/3+\varepsilon}\ll H,\qquad HT^\varepsilon\ll L\leq T. \] As an immediate consequence, in the generic range when both $|t_g|$ and $L$ are of size $T$, we obtain the Weyl bound \[ L\!\left(\frac12,f\otimes g\right) \ll_\varepsilon C(f\times g)^{1/6+\varepsilon}. \]

math.NT↗

arXiv2026

F-LIVO: A Feature-Level Geometric Measurement Framework for LiDAR-Inertial-Visual Odometry

Measurement-level LiDAR-Inertial-Visual Odometry (LIVO) can achieve high accuracy, and its practical resource requirements depend on the adopted association and map-maintenance strategy. Persistently attaching dense photometric patches or appearance attributes to mapped LiDAR points can introduce additional storage, association, and update costs. This paper presents F-LIVO, a filter-based LIVO framework that couples a visual-inertial subsystem and a LiDAR-inertial subsystem through feature-level point-to-plane constraints between visual points and LiDAR planes. To preserve cross-modal coupling when visual triangulation is weak, we further recover visual feature depth from the LiDAR map through local plane fitting and ray-plane intersection. In the LiDAR update, the covariance of each point-to-plane residual is adaptively scaled according to local planar reliability, thereby reducing overconfident updates on poorly conditioned surfaces. Because the proposed method relies on sparse geometric features rather than persistent dense photometric map attributes, it retains a compact map representation. Experiments on multiple public datasets show competitive, sequence-dependent accuracy. Across the evaluated degradation cases, F-LIVO maintains continuous estimation with comparatively low accumulated end-to-end drift. Under the evaluated resource configuration, it achieves the lowest mean and peak central processing unit (CPU) usage, the lowest peak resident set size (RSS), and the lowest RSS growth rate among the evaluated implementations, while R2LIVE retains a lower mean RSS. The source code will be made publicly available upon acceptance to facilitate reproducible research.

cs.RO↗

arXiv2026

JockeyBench: Evaluating LLM Agents on Horse-Racing Decision-Making

Repeated event forecasting requires extracting predictive signals from noisy evidence and evaluating their value across successive decisions. We introduce JockeyBench, a benchmark and evaluation protocol that studies this problem through historical Japanese horse racing. The protocol combines time-gated evidence, priced tools, paired frozen-program controls, and two workflows: executing an LLM-written predictor or acquiring evidence for a fresh per-race LLM forecast. Evaluation separates forecast quality, benchmark-specific net utility, and retrospective betting returns and risk under fixed decision rules. Across 28 model--scaffold configurations on 324 races, paired tests identify seven revision gains and two declines after multiple-testing correction. Yet every configuration has negative mean profit under capped-Kelly settlement. In a four-model comparison, the reusable-program workflow has higher mean betting profit and lower research fees, although both workflows lose on average. Further diagnostics reveal that mean probability bias reverses on selected bets and that profitable individual runs can depend on concentrated returns. These findings motivate evaluating predictive competence, adaptation, resource use, and decision value separately when assessing agents for repeated event forecasting.

cs.AI↗

arXiv2026

How Little Red Dots Turn Blue: Evolution of Kelvin-Helmholtz Contracting Black Hole Envelopes and Observational Diagnostics

James Webb Space Telescope (JWST) observations have revealed compact, red-colored, broad-line emitters known as little red dots (LRDs), while a bluer counterpart population has been referred to as little blue dots (LBDs). Their physical connection remains unclear: they may represent different evolutionary stages of the same phenomenon, distinct populations, or different phases of the broader evolution of early galaxies and their black holes (BHs). We propose a model in which a rapidly growing BH is embedded in a geometrically thick gaseous envelope accompanied by concurrent nuclear star formation. The model reproduces the observed spectral energy distributions (SEDs) of LRDs and LBDs, and naturally connects the two phases through Kelvin-Helmholtz contraction of the envelope and exposure of the nuclear accretion disk, as the gas supply declines due to external stellar activity or feedback. We predict the corresponding evolutionary tracks in the UV-optical slope ($β_{\rm UV}$-$β_{\rm opt}$) plane which bridge the apparent gap between LRDs and LBDs. These evolutionary tracks start from the LRD regime and consistently exhibit a red excursion that extends toward $(β_{\rm UV},β_{\rm opt})\sim(2,-1)$ before turning toward the LBD regime. This feature is largely insensitive to variations in the model parameters that govern the envelope structure, providing a robust signature of the transitional population. The transition lasts $\sim5$-$50\,\mathrm{Myr}$, depending primarily on the BH feedback efficiency and the gas-retention efficiency of the envelope. The abundance ratio of transitional systems to LRDs therefore provides a constraint on BH feedback and gas retention during the rapid growth of early supermassive BHs.

astro-ph.GA↗

arXiv2026

Training-Data Extraction from Genomic Language Models

Genomic language models generate long nucleotide continuations from short prefixes, exposing a direct route for studying training-sequence extraction through the generation interface. We present an output-only black-box attack on Evo 2 that selects convergent groups of repeated generations without gradients, token likelihoods, or training-corpus access during candidate selection. Recovered content is then scored by an evaluator combining aligned length, nucleotide identity, match multiplicity, and sequence complexity. Across 304 runs on Evo 2 7B with bacteriophage and plasmid prefixes, the attack yields 888 candidate clusters, of which 85 match training segments of at least 500 bp, giving a cluster-level attack success rate of 9.57%. Final filtering retains 22 uniquely matched extraction events at the 500-bp threshold and 15 at 1,000 bp, covering 14 and 7 distinct training-sequence regions, respectively. Within-category source-homology screening retains 11 of the 22 events at a 99% exclusion threshold and four at 98%, with the latter mapping to two reference regions. The results establish convergence among repeated nucleotide generations as a practical output-only signal for recovering long training-sequence segments and characterize how recovery depends on prefix length, cluster size, perplexity, early stopping, and sequence domains.

cs.CR↗

arXiv2026

From Parameters to Behaviors: A Survey of Model Fusion for Large Language Models

Model fusion integrates the capabilities from source models into a single target model. As of June 2026, Hugging Face hosts more than 2M models. This growing pool provides a rich base for model reuse and capability integration. Yet existing surveys often cover only separate parts of this space, and they do not provide a unified definition or a systematic taxonomy. This survey defines model fusion and organizes prior work into three levels: parameter-level, representation-level, and behavior-level fusion. We also review related metrics, benchmarks, and applications, summarize current challenges, and identify future directions. Our goal is to provide a clear map of this area and support future work on model fusion. A comprehensive list of papers about model fusion is available at https://github.com/Baicaihaochi/Awesome-Model-Fusion-Survey.

cs.CL↗

arXiv2026

FeatCal: Feature Calibration for Post-Merging Models

Model merging combines task experts into one model and avoids joint training, retraining, or deploying many expert models, but the merged model often still underperforms task experts. We study this performance gap through feature drift, the difference between features produced by the merged model and by the expert on the same input. Our theory decomposes this drift into upstream propagation and local mismatch, tracks how it propagates and combines through later layers in forward order, and links final feature drift to output drift. This view motivates FeatCal, which uses a small calibration set to calibrate the merged model weights layer by layer in forward order, reducing feature drift while staying close to merged weights and preserving the benefits of model merging. FeatCal uses an efficient closed-form solution to update model weights, with no gradient descent, iterative optimization, or extra modules. On the main CLIP and GLUE benchmarks, FeatCal beats Surgery and ProbSurgery, the closest post-merging calibration baselines: 85.5% vs. 77.0%/78.8% on CLIP-ViT-B/32 Task Arithmetic (TA) and 85.2% vs. 83.7%/82.2% on FLAN-T5-base GLUE. On CLIP-ViT-B/32, 8 examples per task reach 82.9%, and 256 examples per task take 53 seconds, about 4x faster than both baselines, showing better sample efficiency and lower calibration cost.

cs.LG↗

arXiv2026

SMAT: Simple and Efficient Merge-Aware Training

Model merging integrates the capabilities of multiple experts without joint retraining, but standard expert training optimizes task loss alone and does not guarantee good performance after merging. Merge-aware training (MAT) aims to improve merged performance, but existing methods do not fully account for common merging operations and add training cost. We observe that, from an expert's perspective, common merging methods can be described by three operations: Scale reweights its own update, Mask removes selected coordinates, and Perturb adds updates from other experts. Based on this view, we introduce SMAT (Simple MAT), which jointly optimizes expert loss and expected loss at simulated merged parameters generated by sampling scaling coefficients, masks, and additive noise. We further introduce periodic scheduling, kernel fusion, and parameter storage switching to make SMAT efficient, with one forward and one backward pass per step. Across four language and vision-language backbones, SMAT improves the mean score across five merging methods by 1.07-2.16 points over the strongest baseline for each backbone, with less than 2% training-time overhead over standard fine-tuning.

cs.LG↗

arXiv2026

MoEMB: Scaling Universal Multimodal Embeddings with Efficient Mixture-of-Experts Models

Universal multimodal embedding (UME) increasingly demands encoder's capacity for handling a broad range of tasks and modalities with increased complexity. Prior scaling methods either increase the representation size, retrieval effort, or scales the encoder into a heavy multimodal LLM. Recent works, such as Think-Then-Embed (TTE), explore scaling via reasoning tokens. However, embedding models are hard to scale up: increasing parameters directly tradeoffs for the large training batch size that contrastive learning needs, and retrieval has to be served under tight latency. Moreover, UME tasks are diverse in complexity, where scaling up embedders can bring significant redundant computation. In this work, we propose MOEMB, which instead scales UME along the expert axis through mixture-of-experts (MoE), growing encoder capacity while preserving single-vector, non-autoregressive encoding. Through a systematic study of the design space and training recipes for MoE-based UME, MoEMB sets a new state of the art on both MMEB-V2 and MRMR among models trained on public MMEB-family data: with only 3B active parameters, MoEMB surpasses TTE-based methods with >4x active parameters, using significantly less computes. To further improve the scalability and efficiency, we conduct the first comprehensive study of adaptive computation for MoE-based embedding, spanning diverse strategies across training-based and inference-only methods. Together, these results support expert scaling as an effective and efficient direction for UME, with adaptive computation further improving efficiency for MLLM-based embedding models towards large-scale retrieval and recommendation systems.

cs.LG↗

arXiv2026

Beyond the Flat-Band Paradigm: Open Orbits as a Route to Correlated States in Dispersive Moiré Minibands

The realization of fractional Chern insulators has largely relied on isolated flat bands in complex bilayer or multilayer moiré systems with ideal quantum geometry. Whether such phases can emerge in genuinely dispersive bands, and by what mechanism, remains an open question. Here, we propose a mechanism for magnetic-field-induced strong correlations in the highly dispersive bands of monolayer graphene on hexagonal boron nitride. In the low-field classical-to-quantum crossover regime of the second moiré miniband, we show that the interplay of the moiré potential and weak magnetic fields (B < 1.5 T) profoundly alters electron dynamics. Saddle-point van Hove singularities produce extended open orbits that strongly suppress carrier propagation, potentially favoring a regime in which Coulomb interactions become comparatively more important, while weak magnetic fields lift valley degeneracy and trigonal warping redistributes Berry curvature. High-resolution magnetotransport and temperature-dependent measurements reveal fractional-slope transport features and signatures of interaction-driven incompressibility. Because the Hall conductivity is not fully quantized and the longitudinal resistance remains finite, we describe these observations as incipient FCI behavior: transport signatures consistent with a developing fractional Chern insulating state, without claiming a fully developed FCI phase established by complete Hall quantization and vanishing longitudinal resistance. Our results suggest a possible fermiology-driven route toward correlated topological states in dispersive bands in the classical-to-quantum crossover, distinct from the conventional flat-band scenario.

cond-mat.mes-hall↗