arXiv subjects

Jun Zhang

Publications and source records attributed to Jun Zhang.

At least 109 records · Page 6Linked to original sources

arXiv2026

OmniDelta: Skill-Driven Budget Allocation for Token Compression in OmniLLMs

Emerging Omni-modal Large Language Models (OmniLLMs) enable unified understanding of text, audio, and video, but their long audio-video token sequences introduce substantial memory and inference costs. Existing compression methods mainly focus on selecting important tokens under fixed budgets, leaving the preceding budget-allocation problem underexplored. We show that direct query-to-audio/video similarity is unreliable for inter-modal budget allocation, and that uniform intra-modal budgets can miss key evidence while retaining redundant content. To address these limitations, we propose OmniDelta, a training-free, skill-driven framework that couples intent-aware inter-modal allocation with content-aware intra-modal allocation. OmniDelta first constructs audio and video skill pools to shift the fixed retained-token budget according to query demand, then reallocates modality budgets over audio segments and video frames using local complexity and temporal redundancy. The resulting local budgets can be combined with existing pruning strategies, preserving the total retained-token ratio while changing where the budget is spent. Experiments on four audio-video benchmarks with two Qwen2.5-Omni models show that OmniDelta establishes a new accuracy-efficiency Pareto frontier across pruning ratios. At 25% token retention on Qwen2.5-Omni-7B, OmniDelta reduces GPU memory by 22.0% and achieves a 1.64x end-to-end speedup over full-token inference.

cs.AI↗

arXiv2026

Inferring Hidden Motives: A Utility Bayesian Model of learning the values of others

Cooperation depends on judging whom to trust: identifying the strengths of other people's morally relevant social motives, called social preferences, and refining those beliefs after observing their behavior over time. We model this process with a Utility Bayesian Model (UBM): a continuous probability distribution over the parameters of another person's utility function, reweighted after each observed choice. In repeated binary dictator games with randomized payoffs, the UBM captured participants' predictions of preprogrammed agents and of one another better than non-Bayesian alternatives and over 15,000 models that represent others as discrete types: people track motives as graded positions in a continuous space, not as categories. Because participants alternated between choosing and predicting, the framework separately estimates preferences and beliefs about preferences. Choosers' own utilities, fit with a seven-parameter function selected from 526 candidate forms, showed that most people place reliably positive weight on a stranger's payoff but weight their own roughly nine times more; that aversion to being ahead exceeds aversion to being behind, reversing the canonical ordering; that sensitivity to others' payoffs and to inequalities grows more than proportionally with their size; and that antisocial preferences are more common than prior estimates suggest. Predictors' initial beliefs captured the dominance of self-interest but expected aversion to being behind to outweigh aversion to being ahead. Together, these parameter distributions locate any individual's social preferences, and beliefs about them, in a common seven-dimensional space, enabling commensurable cross-study comparisons, meta-analyses, and empirically informed priors for future Bayesian models of social cognition.

q-bio.NC↗

arXiv2026

BeyondFusion: Self-Aligned Latent Diffusion for Calibration-Free Infrared Super-Resolution and Infrared-Visible Fusion

Mobile infrared-visible imaging typically pairs a compact infrared sensor with a high-resolution visible camera for complementary perception. While cross-sensor misalignment caused by different optics, viewpoints, fields of view, and exposure timings hinders practical deployment. In this paper, we propose BeyondFusion, a unified latent diffusion framework for calibration-free visible-guided infrared super-resolution and infrared-visible fusion tasks. The proposed framework supports both task-specific training and joint training where two tasks are optimized and executed as two readouts of the same generative process. Instead of relying on explicit registration or geometric warping, BeyondFusion introduces a cross-modal self-aligning (CMSA) module into the denoising U-Net. CMSA reorganizes infrared and visible latent tokens into a shared attention space to learn content-adaptive cross-modal correspondence during the denoising process. Together with misalignment augmentation module, the model is facilitated to exploit visible structural and semantic cues while preserving thermal consistency, enabling high-frequency infrared reconstruction and informative fused-image generation under uncalibrated conditions. Extensive experiments on public benchmarks and a mobile infrared-visible imaging system show strong performance across aligned inputs, low-resolution infrared observations, synthetic misalignments, and real mobile captures with unsynchronized sensors. Ablation studies, unified training analysis, and downstream pedestrian detection further validate the effectiveness of BeyondFusion for calibration-free multimodal imaging.

cs.CV↗

arXiv2026

FA-LAM: Focus-Aware Large Avatar Model for One-Shot 4D Animatable Gaussian Head

We propose FA-LAM, a Focus-Aware Large Avatar Model for one-shot animatable Gaussian head creation, while simultaneously enabling static 3D and dynamic 4D full-head recovery. The core of our method lies in a thorough analysis of the attention mechanisms and the entangled reconstruction and animation training pipeline adopted by prior state-of-the-art approaches. Our analysis identifies two main factors that compromise the quality of 3D full-head generation: (1) incorrect and noisy attention activations, and (2) conflicts between the tasks of reconstruction and animation. To address the first issue, we introduce a symmetric and semantic attention regularization strategy that leverages the inherent semantics and structural symmetry of human heads. To disentangle the objectives of reconstruction and animation, we develop a novel dual-phase training pipeline that separates the model's capabilities for large-view hallucination and animation into distinct modules. Moreover, we enhance our model to support multi-view and streaming 4D reconstruction in an efficient and memory-friendly manner through a core autoregressive modification with tailored visibility-aware token fusion. Collectively, these innovations enable FA-LAM to reconstruct animatable Gaussian full heads with superior quality, particularly in fine facial regions and large viewing angles.

cs.CV↗

arXiv2026

On-chip Radio Frequency Maser

Room-temperature solid-state masers offer exceptional frequency selectivity and ultra-low noise for weak-signal detection. However, their reliance on bulky metallic resonators has significantly hindered integration, miniaturization, and extension to lower frequencies. Here, we demonstrate the first on-chip radio-frequency maser operating at room temperature, exploiting optically pumped triplet states of pentacene. The device produces stimulated emission at 106.62 MHz and enables ultra-sensitive microwave magnetic-field detection with a sensitivity of ($\sim 10\,\rm{fT/\sqrt{Hz}}$), functioning simultaneously as a local oscillator and a sensor. By actively controlling microwave dissipation, we achieve efficient regulation of the maser output, revealing a key mechanism for tuning emission in open cavity-free systems. This work extends pentacene-based masers into the radio-frequency regime and establishes a highly integrated on-chip architecture for room-temperature masers, offering a new pathway toward portable quantum devices.

quant-ph↗

arXiv2026

The Second LoViF 2026 Challenge on Real-World All-in-One Image Restoration: Methods and Results

This paper presents a review of the second LoViF Challenge on Real-World All-in-One Image Restoration. The challenge aims to advance unified image restoration under diverse real-world degradation conditions, including blur, low-light, haze, rain, and snow. It provides a common benchmark for evaluating the restoration accuracy, robustness, and generalization capability of models across multiple degradation categories within a unified framework. The competition attracted 158 registered participants, and 20 teams were included in the final ranking after their submitted results were successfully reproduced and verified. This report provides a comprehensive analysis of the submitted solutions and corresponding results, highlighting recent advances in real-world all-in-one image restoration. The summarized methods and empirical findings reveal effective design strategies and establish an updated benchmark for future research in real-world low-level vision.

cs.CV↗

arXiv2026

Duality and Reverse Self-Dual Constructions for Hyperderivative Reed-Solomon Codes

Hyperderivative Reed-Solomon (HRS) codes form a class of maximum-distance-separable codes under the Niederreiter-Rosenbloom-Tsfasman metric and may be viewed as a derivative-evaluation extension of classical Reed-Solomon codes. For generalized Reed-Solomon codes, the Euclidean dual is again a generalized Reed-Solomon code. In this paper, we investigate the corresponding duality problem for HRS codes. Using a residue-theoretic argument, we derive an explicit component-wise representation for the Euclidean dual of an HRS code. The formula shows that, in general, the Euclidean dual is not an HRS code. Instead, it is blockwise upper-triangularly equivalent to a reverse-order HRS evaluation code, where the reversal occurs in the hyperderivative orders within each evaluation block. In particular, for full-domain HRS codes with low multiplicity, the triangular transformations reduce to diagonal scalings, and the Euclidean dual is obtained as the row reversal of an HRS code. Based on this reverse-order dual structure, we further study reverse self-dual HRS codes. We establish explicit criteria for reverse self-duality and construct several families from additive and multiplicative coset structures.

cs.IT↗

arXiv2026

MeanFlowNFT: Bringing Forward-Process RL to Average-Velocity Generators

MeanFlow generators achieve fast few-step sampling by predicting average velocities over time intervals, making them attractive for efficient generation. Reinforcement learning (RL) has become a powerful way to align diffusion and flow models with human preferences and task-specific objectives. In particular, DiffusionNFT offers an efficient forward-process RL framework that does not require reverse-process trajectories or likelihood estimation. However, applying such RL methods to MeanFlow remains underexplored. DiffusionNFT optimizes instantaneous velocities, whereas MeanFlow samples with average velocities. To bridge this gap, we introduce MeanFlowNFT. Inspired by the MeanFlow identity, which bridges average and instantaneous velocities, we construct an induced instantaneous-velocity predictor. We apply the DiffusionNFT objective to this predictor, making reward optimization well-defined for MeanFlow. Sampling remains based on the average velocity, preserving MeanFlow's fast few-step generation. We further prove that MeanFlowNFT inherits DiffusionNFT's strict policy-improvement guarantee. Experiments on image and video generation show that MeanFlowNFT consistently improves baselines. Moreover, it outperforms prior state-of-the-art RL-tuned few-step generators on most metrics ($6$ of $8$ on SD3.5-M), and can even surpass multi-step RL-tuned diffusion while using only a few sampling steps. For instance, on Wan 2.1, $4$-step MeanFlowNFT reaches a VBench score of $84.33$, surpassing $50$-step LongCat-Video RL ($82.57$).

cs.CV↗

arXiv2026

Non-Square UPA-Enabled XL-MIMO Systems: Anisotropic Near-Field Characterization, Fundamental Limits, and Channel Estimation

Extremely large-scale multiple-input multiple-output (XL-MIMO) is crucial for next-generation communication systems. In practice, the deployment of non-square uniform planar arrays (UPAs) fundamentally alters wavefront characteristics and induces anisotropic beamfocusing capability along different axes due to the aperture disparity originating from the non-square array geometry. To fully uncover the performance impact of such non-square geometry and thus unleash the potential of the non-square UPAs, we investigate the anisotropic near-field characteristics, fundamental limits, and channel estimation for non-square UPA-enabled XL-MIMO systems. First, we derive the effective beamfocusing distances for the long and short axes of the array. Interestingly, the radiation space of a non-square UPA can be partitioned into three regions, i.e., the fully near-field, the anisotropic near-field, and the far-field regions, and the anisotropic region asymptotically dominates the overall near-field space as the array aspect ratio increases. Then, the asymptotic effective degree of freedom for non-square UPA-enabled XL-MIMO systems is provided, which reveals that distance-domain multiplexing is governed by the long-axis aperture in the large array aspect ratio regime. Furthermore, the closed-form Cramer-Rao bound for distance estimation and the three-dimensional (3D) position error bound (PEB) are derived to reveal the geometry-induced performance trade-offs among distance, azimuth, and elevation estimation, based on which the optimal array aspect ratio that minimizes the 3D PEB is determined. Finally, by exploiting the anisotropic wavefront properties, we design a 3D anisotropic near-field codebook to facilitate low-complexity channel estimation for non-square UPAs. Numerical results validate that the proposed codebook achieves comparable accuracy to the 3D polar-domain codebook at reduced complexity.

eess.SP↗

arXiv2026

Task-Oriented Precoding for Edge Inference over Large-Scale MIMO Systems

Future wireless networks are expected to support networked artificial intelligence (AI) services, where multiple devices transmit learned features to an edge server for distributed inference. This setting calls for task-oriented physical-layer optimization, where wireless transmission should preserve useful information for inference rather than only maximize the rate or reconstruct the transmitted signals. A key physical-layer control variable is the multiple-input multiple-output precoder, which determines how device features are shaped and combined over wireless channels. Existing task-oriented precoding methods typically adapt the precoder to instantaneous channel state information at the transmitter (CSIT). However, in multi-device MIMO systems, acquiring the aggregate channel, feeding back CSI or optimized precoders, and reoptimizing across coherence blocks introduce substantial overhead. This paper develops a random-matrix-theoretic framework based on statistical CSIT that designs a slow-timescale precoder from channel covariance statistics and training-set feature statistics, without requiring instantaneous CSIT. We adopt maximal coding rate reduction (MCR${^2}$) to measure the class separability of the received features, yielding a task-aware utility for MIMO precoder design. Since this utility still depends on random small-scale fading, we derive a deterministic approximation that converts it into a fixed-point objective depending only on long-term statistics and large-system dimension ratios via random matrix theory. A projected block-coordinate ascent and successive convex approximation algorithm is developed to optimize this deterministic objective under per-device power constraints. Experiments on ModelNet10 verify the approximation and show that the proposed statistical precoder improves task-aware mode allocation and inference performance over competitive benchmarks.

cs.IT↗

arXiv2026

TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs

Video multimodal large language models (MLLMs) can describe what happens in a video, but rarely identify when the supporting evidence occurs. We study generalist video temporal grounding, in which one model predicts a variable-cardinality set of evidence intervals across video lengths, domains, query forms, and viewpoints. Existing training strategies are misaligned with this set-valued task: long-video labels often rely on brittle one-pass annotation, while reinforcement-learning rewards either fail to distinguish non-overlapping predictions or require fragile segment matching. TimeLens2 treats temporal evidence as an interval set throughout supervision and optimization. TimeLens2-93K constructs reliable multi-span supervision through caption-derived proposals, independent localization, cross-agent consensus, semantic verification, and boundary refinement. Our temporal Wasserstein reward computes exact one-dimensional $W_1$ between uniform distributions over merged interval supports, providing dense, matching-free feedback under unequal cardinalities and equivalent fragmentation; temporal IoU complements it with precise-overlap feedback. Across seven benchmarks, TimeLens2-2B outperforms all size-matched baselines on every benchmark, while the 4B and 8B variants achieve state-of-the-art performance, surpassing open-source models with up to 397B parameters. The 2B, 4B, and 8B variants improve over their Qwen3-VL backbones by 14.2, 13.0, and 18.1 mIoU points, respectively.

cs.CV↗

arXiv2026

Task-Oriented Communication with Hybrid-Precision Models

Edge inference has emerged as a promising solution for the proliferation of artificial intelligence (AI) services by deploying models at the network edge to circumvent cloud-routing latency. Existing edge inference approaches mainly focused on either cooperative inference to reduce latency or lightweight model design to fit resource-constrained devices. These solutions often address the communication and computation challenges separately, and thus struggle to achieve a balanced trade-off among transmission efficiency, on-device processing cost, and inference accuracy. To bridge this gap, this paper proposes a hybrid-precision task-oriented communication framework for edge inference to holistically balance communication, on-device computation, and utility. In this framework, a binarized front-end is deployed on the edge device to extract and transmit binary features via orthogonal frequency-division multiplexing (OFDM) signals, while a full-precision back-end on the edge server performs the final inference. To ensure model consistency, we introduce an on-device binarization method tailored for split inference and develop an integrated channel-aware transmission scheme featuring subcarrier-based feature calibration. Furthermore, a knowledge distillation (KD)-based training strategy, supported by specialized gradient estimators, is developed to optimize the end-to-end system and inherit semantic knowledge from a full-precision teacher model. Extensive experiments on the large-scale ImageNet dataset demonstrate the superiority of the proposed hybrid system. Our analysis confirms that this design achieves an optimal trade-off among communication efficiency, on-device computational cost, and inference accuracy, outperforming existing edge inference solutions.

eess.SP↗

arXiv2026

AgentSociety 2: An Integrated Research Environment for Executable Social Science

AI scientist systems are beginning to automate parts of scientific research, but social science poses a distinct challenge: its objects of inquiry are not merely datasets or laboratory protocols, but integrated social processes involving situated participants, interaction contexts, interventions, and outcomes. Yet a critical link is missing: existing systems either assist isolated research tasks or simulate agents as experimental subjects, leaving the research workflow and simulated society decoupled. Here we introduce AgentSociety 2, an Integrated Research Environment for executable social science. It couples two roles of LLM agents in the same runtime: AI social scientists that coordinate literature grounding, hypothesis generation, experiment design, simulation execution, result interpretation, and manuscript drafting; and silicon participants that generate behavioral responses within configurable social environments. This dual-role design turns hypotheses into auditable agent behaviors, environment rules, interventions, and measurements, thereby supporting an end-to-end workflow. Across seven illustrative studies spanning micro-level social-science laboratory experiments, meso-level dynamics in social media, and macro-level urban scenarios, we demonstrate its capacity to support diverse disciplinary questions, reproduce major qualitative patterns from prior studies, identify informative deviations, and enable large-scale simulations through optimized agent-environment interactions. By preserving human researchers' high-level agency while delegating procedural orchestration to agentic systems, it provides a human-in-the-loop and controllable infrastructure for next-generation computational social science, with broader applications in scalable computational social experimentation and AI-enabled social governance platforms.

cs.CY↗

arXiv2026

Prospect of Measuring the Cosmic Dipole by Associating Strongly Lensed Gravitational Waves with Galaxy Surveys

The cosmic dipole observed in the cosmic microwave background (CMB) is traditionally interpreted as being caused by the observer's motion relative to the background. However, tensions with dipole measurements from radio galaxy counts motivate the need for independent probes. This work investigates the feasibility of using strongly lensed gravitational wave (GW) events to measure the cosmic dipole. Strongly lensed GWs produce multiple time-delayed images, which can be used to infer the distances to both the lens and the source. These distances, associated with the observed redshifts of the lens and the source from galaxy catalogues, encode information about the background cosmology and cosmic dipole effects. By reconstructing a statistical sample of doubly lensed GW events based on the singular isothermal sphere lens model, the cosmic dipole can be estimated jointly with background cosmological parameters. Using realistic simulations for Einstein Telescope and Cosmic Explorer, we forecast that a dipole magnitude $g$ consistent with both the CMB and number count measurement could be detected with 10 years of observation. Furthermore, constraints on $g$ are greatly improved by combining constraints from doubly lensed events with those from triply or quadruply lensed events. In the most optimistic scenario, where we measure the number count dipole magnitude with 10 years of observation, we obtain $g = (2.45^{+1.53}_{-1.28}) \times 10^{-3}$ from the combined constraint, provided that systematic uncertainties are mitigated. Although challenging, strongly lensed GWs offer a novel approach to measuring the cosmic dipole, providing an independent consistency test with different systematics from electromagnetic probes.

astro-ph.CO↗

arXiv2026

Observation of node-dependent Rydberg molecular bound states

Ultralong-range Rydberg molecules, formed by the interaction between a highly excited Rydberg atom and a ground-state atom, provide a unique platform for exploring quantum phenomena spanning nanometer-to-micrometer distances as well as exotic few-body interactions. The formation mechanisms and resultant physical properties differ markedly between s-wave and p-wave scattering channels. Here we report the experimental observation of node-dependent p-wave molecular signals in Rb(nS)-Rb(5S) Rydberg molecular spectra, where variations in the principal quantum number n directly reveal the shift of molecular binding energies induced by the moving nodal structure of the Rydberg electron wavefunction. This node-dependence is attributed to a cooperative effect between the local gradient of the nS-electron wavefunction and the energydependent p-wave scattering length. In addition, resolving two p-wave bound states associated with adjacent nodes highlights the remarkable sub-nanometer spatial resolution achieved in our experiment. Our findings reveal a more profound quantum control mechanism, wherein the principal quantum number acts as a switch for nodal-selective molecular bound states, and the reported method provides a sensitive spectroscopic probe of electron-atom scattering.

physics.atom-ph↗

arXiv2026

RoadBench: Benchmarking MLLMs on Fine-Grained Spatial Understanding and Reasoning under Urban Road Scenarios

Multimodal large language models (MLLMs) have demonstrated powerful capabilities in general spatial understanding and reasoning. However, their fine-grained spatial understanding and reasoning capabilities in complex urban scenarios have not received significant attention in the fields of both research and industry. To fill this gap, we focus primarily on road markings as a typical example of fine-grained spatial elements under urban scenarios, given the essential role of the integrated road traffic network they form within cities. Around road markings and urban traffic systems, we propose \textbf{RoadBench}, a systematic benchmark that comprehensively evaluates MLLMs' fine-grained spatial understanding and reasoning capabilities using Bird's-Eye View (BEV) and First-Person View (FPV) image inputs. This benchmark comprises eight tasks consisting of 3,040 strictly manually verified test cases, constructed from 2,137 unique BEV images and 721 unique FPV images collected from five Chinese cities with relatively consistent traffic conventions. These tasks form a systematic evaluation framework that bridges understanding at local spatial scopes to global reasoning. They not only test MLLMs' capabilities in recognition, joint understanding, and reasoning but also assess their ability to integrate image information with domain knowledge. After evaluating 20 mainstream MLLMs, we confirm that RoadBench is a challenging benchmark for MLLMs while revealing significant shortcomings in existing MLLMs' fine-grained spatial understanding and reasoning capabilities within urban scenarios. In certain tasks, their performance even falls short of simple rule-based or random selection baselines. These findings, along with RoadBench itself, will contribute to the comprehensive advancement of spatial understanding capabilities for MLLMs.

cs.CV↗

arXiv2026

Salt: Self-Consistent Distribution Matching with Cache-Aware Training for Fast Video Generation

Distilling video generation models to extremely low inference budgets (e.g., 2--4 NFEs) is crucial for real-time deployment, yet remains challenging. Trajectory-style consistency distillation often becomes conservative under complex video dynamics, yielding an over-smoothed appearance and weak motion. Distribution matching distillation (DMD) can recover sharp, mode-seeking samples, but its local training signals do not explicitly regularize how denoising updates compose across timesteps, making composed rollouts prone to drift. To overcome this challenge, we propose Self-Consistent Distribution Matching Distillation (SC-DMD), which explicitly regularizes the endpoint-consistent composition of consecutive denoising updates. For real-time autoregressive video generation, we further treat the KV cache as a quality parameterized condition and propose Cache-Distribution-Aware training. This training scheme applies SC-DMD over multi-step rollouts and introduces a cache-conditioned feature alignment objective that steers low-quality outputs toward high-quality references. Across extensive experiments on both non-autoregressive backbones (e.g., Wan~2.1) and autoregressive real-time paradigms (e.g., Self Forcing), our method, dubbed \textbf{Salt}, consistently improves low-NFE video generation quality while remaining compatible with diverse KV-cache memory mechanisms. Project page: https://xingtongge.github.io/Salt

cs.CV↗

arXiv2026

Experience-Evolving Multi-Turn Tool-Use Agent with Hybrid Episodic-Procedural Memory

As intents unfold and environments change, multi-turn agents face continuously shifting decision contexts. Although reusing past experience is intuitively appealing, existing approaches remain limited: full trajectories are often too context-specific to transfer, while tool-level reuse ignores the surrounding context and environment. In this paper, we introduce a hybrid episodic-procedural memory strategy (H-EPM) that enables experience-induced self-evolution of multi-turn tool-use policies by adaptively reusing partially overlapping successful experiences during both inference and training. Inspired by human episodic-procedural integration, we construct a tool graph from accumulated trajectories, where recurring tool-to-tool dependencies capture procedural routines and each edge is augmented with compact episodic summaries of relevant context. At inference time, the agent dynamically balances episodic recall for contextual reasoning with procedural execution for routine steps. Beyond inference, H-EPM introduces a memory-guided reinforcement learning paradigm that directly addresses a core challenge in multi-turn agent reinforcement learning, namely ineffective exploration over long trajectories. By biasing exploration toward historically successful tool transitions, H-EPM learns a stronger policy that generalizes at inference time without relying on domain-specific experience collection. Experiments show that H-EPM consistently delivers substantial inference-time gains over strong baselines across multi-turn tool-use benchmarks, reaching improvements of up to fifty percent. It also improves reinforcement learning policy performance, achieving gains of up to forty percent on out-of-distribution tasks.

cs.LG↗