arXiv subjects

Kai Wu

Publications and source records attributed to Kai Wu.

At least 37 records · Page 2Linked to original sources

arXiv2026

EVOM: Agentic Meta-Evolution of Actor-Critic Architectures for Reinforcement Learning

In actor-critic reinforcement learning, network architectures are typically manually designed. Automating this design is challenging because each candidate must be trained before evaluation, and the design space is open-ended. To address these challenges, we introduce EVOM, an agentic meta-evolution framework for discovering high-performance actor-critic architectures. We frame architecture search as a bi-level optimization: an inner loop trains weights via the low-fidelity proximal policy optimization (PPO), while an outer loop drives meta-evolution by iteratively refining architecture programs. Crucially, this outer loop is powered by an LLM-based design agent that operates purely as an architecture designer, completely decoupled from policy execution and environment control. Experiments reveal that EVOM outperforms the manually designed baseline, an LLM-guided random search, and the state-of-the-art LLM-guided programmatic policy search method MLES, delivering superior performance on Ant-v4 and HalfCheetah-v4. Ablation studies validate that both the meta-evolution loop and the LLM Design Agent are indispensable for final performance.

cs.LG↗

arXiv2026

Feature to Dynamics: Feature-space to Autoregression strategy for Zero-shot Time Series Forecasting

Zero-shot time series forecasting aims to predict future values for previously unseen series, requiring models to generalize temporal dynamics beyond the training distribution. While recent foundation models achieve strong in-domain performance through large-scale pretraining, their effectiveness often relies on broad data coverage and implicit pattern memorization, which can limit generalization when data are scarce or source and target domains are disjoint. In this work, we propose FSA, a feature-to-strategy framework for controlled zero-shot univariate forecasting. Instead of directly modeling raw sequences in the observation space, FSA learns a structured mapping from an interpretable feature space to an autoregressive strategy space. This design introduces explicit inductive biases that disentangle global trends, periodic components, and local temporal dynamics, enabling the model to capture transferable time-series structure with fewer data assumptions. Empirical results show that, under identical pretraining data, training protocol, and comparable parameter budgets, FSA outperforms Transformer-based architectures in our controlled zero-shot setting.

cs.LG↗

arXiv2026

ReactBench: A Cause-Driven Benchmark for Multimodal Hallucination via Systematic Evaluation

While multimodal large language models (MLLMs) have achieved rapid progress in vision-language understanding, they remain prone to multimodal hallucinations, producing responses that are inconsistent with the visual input. Existing benchmarks predominantly focus on detecting hallucination outcomes rather than evaluating the underlying causes of these failures. Moreover, many benchmarks rely on simplistic scenarios and limited evaluation formats that no longer challenge state-of-the-art models. To address these limitations, we introduce ReactBench, a cause-driven hallucination benchmark featuring multiple tasks and an exam-style evaluation format. By generating adversarial images and hallucination-inducing queries, ReactBench introduces four targeted tasks: Relational Erasure, Counterfactual Attribute, Alteration Tracing, and Dense Counting. These tasks systematically expose co-occurrence bias, language priors, cross-image comparative perception deficiencies, and fine-grained perceptual bottlenecks. Beyond standard accuracy-based evaluation, we leverage Chain-of-Thought reasoning to identify fine-grained sub-causes of hallucination within each task. Extensive evaluations reveal that current MLLMs remain notably vulnerable to cause-specific hallucination triggers, demonstrating the value of ReactBench as a systematic and interpretable testbed for diagnosing and improving multimodal model robustness. The project page is available at https://reactbench.github.io/.

cs.CV↗

arXiv2026

Posterior-Aware Differential Channel Tracking for Reliable Single-Stream DAB+ Passive Radar

Digital audio broadcasting plus (DAB+) is an attractive illuminator for passive radar because it provides persistent, high-power, and geographically widespread very high frequency (VHF) orthogonal frequency-division multiplexing (OFDM) signals. A channel state information (CSI) sensing approach can convert a single received DAB+ stream into a CSI sequence for radar sensing, avoiding the need for a separately received reference signal in conventional passive radars. However, CSI estimation in DAB+ is challenging due to the differentially encoded communication symbols across time. A wrong symbol transition estimation leads to a persistent multiplicative error in the sequential CSI sequence within a DAB+ frame. This paper formulates single-stream DAB+ passive radar as a posterior-probability-aware differential CSI tracking problem. The proposed method uses the previously tracked CSI as a channel prior, performs prediction-aided maximum a posteriori detection of current symbol, converts posterior transition reliability into observation uncertainty, and applies linear minimum mean squared error fusion to obtain a stable tracking CSI. A reliability-informed CSI fusion strategy is also introduced to preserve weak target information. Theoretical analysis is provided, showing guaranteed performance again in symbol and CSI estimation. Simulation results show that the proposed method can reduce CSI estimation error by over 15~dB compared with prior art. It also improves median target-to-background ratio by more than 11~dB in random fading scenes. Experiments in Sydney, Australia demonstrate improved range-Doppler maps for commercial aircraft sensing.

eess.SP↗

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

Direct N-body simulations of rotating and extremely massive Population III star clusters

Aims. We present eight direct N-body simulations with NBODY6++GPU of extremely massive, initially rotating Population III star clusters with 1.01 x 10^5 stars. Methods. Our models include primordial binaries, a continuous initial mass function, differential rotation, tidal mass loss, updated fitting formulae for extremely massive metal-poor Population III stars, and general-relativistic merger recoil kicks. We assess their impact on cluster dynamics. Results. All runs form black holes below, within, and above the pair-instability gap, with multi-generation growth. Faster-rotating clusters core-collapse earlier; post-collapse clusters host a rotating, axisymmetric subsystem of intermediate-mass black holes (IMBHs) at the centre and an expanding halo of lower-mass objects. Pair-instability supernovae and compact-object formation at ~2-3 Myr sharply reduce total mass and a large fraction of the cluster's angular momentum. All Population III clusters in our simulations have the gravothermal-gravogyro catastrophe phase. Conclusions. We confirm two of the hypothesized formation channels of galactic nucleus seed black holes: gravitational runaway mergers of black holes and of Population III stars, which core-collapse into IMBHs thereafter. A higher initial star cluster bulk rotation correlates with earlier core collapse and, in the event counts reported here, with more coalescences and collisions, as well as lower retained (compact) binary abundances. Initial bulk rotation is a primary control parameter of cluster evolution: faster rotation accelerates early angular-momentum transport, gravothermal collapse, mass segregation, and amplifies post-collapse expansion, which also favours the formation of a compact central IMBH subsystem.

astro-ph.GA↗

arXiv2026

Excitons in WSe2 time-resolved ARPES: particle or oscillation?

The time-resolved angle-resolved photoemission spectra of WSe$_2$, a paradigmatic transition metal dichalcogenide, are dominated by a transient signal that, after being initially observed in the gap at the K valley, scatters, on an ultra-fast time scale of $\sim$ 30 fs, to the $Σ$ valley. In this work we question the common interpretation of the experimental dynamics in terms of a massive bound electron-hole exciton that scatters with phonons and behaves as a quasi-particle. By using a combined theoretical and experimental investigation, we demonstrate that the observed dynamics can be interpreted as the photo-induced transition from direct to indirect excitonic-insulating order. The features that appear in the experimental spectrum correspond to single-particle levels renormalized by the excitonic spontaneous polarization.

cond-mat.mtrl-sci↗

arXiv2026

MedXIAOHE: A Comprehensive Recipe for Building Medical MLLMs

We present MedXIAOHE, a medical vision-language foundation model designed to advance general-purpose medical understanding and reasoning in real-world clinical applications. MedXIAOHE achieves state-of-the-art performance across diverse medical benchmarks and surpasses leading closed-source multimodal systems on multiple capabilities. To achieve this, we propose an entity-aware continual pretraining framework that organizes heterogeneous medical corpora to broaden knowledge coverage and reduce long-tail gaps (e.g., rare diseases). For medical expert-level reasoning and interaction, MedXIAOHE incorporates diverse medical reasoning patterns via reinforcement learning and tool-augmented agentic training, enabling multi-step diagnostic reasoning with verifiable decision traces. To improve reliability in real-world use, MedXIAOHE integrates user-preference rubrics, evidence-grounded reasoning, and low-hallucination long-form report generation, with improved adherence to medical instructions. We release this report to document our practical design choices, scaling insights, and evaluation framework, hoping to inspire further research.

cs.CL↗

arXiv2026

Med-CMR: A Fine-Grained Benchmark Integrating Visual Evidence and Clinical Logic for Medical Complex Multimodal Reasoning

MLLMs MLLMs are beginning to appear in clinical workflows, but their ability to perform complex medical reasoning remains unclear. We present Med-CMR, a fine-grained Medical Complex Multimodal Reasoning benchmark. Med-CMR distinguishes from existing counterparts by three core features: 1) Systematic capability decomposition, splitting medical multimodal reasoning into fine-grained visual understanding and multi-step reasoning to enable targeted evaluation; 2) Challenging task design, with visual understanding across three key dimensions (small-object detection, fine-detail discrimination, spatial understanding) and reasoning covering four clinically relevant scenarios (temporal prediction, causal reasoning, long-tail generalization, multi-source integration); 3) Broad, high-quality data coverage, comprising 20,653 Visual Question Answering (VQA) pairs spanning 11 organ systems and 12 imaging modalities, validated via a rigorous two-stage (human expert + model-assisted) review to ensure clinical authenticity. We evaluate 18 state-of-the-art MLLMs with Med-CMR, revealing GPT-5 as the top-performing commercial model: 57.81 accuracy on multiple-choice questions (MCQs) and a 48.70 open-ended score, outperforming Gemini 2.5 Pro (49.87 MCQ accuracy, 45.98 open-ended score) and leading open-source model Qwen3-VL-235B-A22B (49.34 MCQ accuracy, 42.62 open-ended score). However, specialized medical MLLMs do not reliably outperform strong general models, and long-tail generalization emerges as the dominant failure mode. Med-CMR thus provides a stress test for visual-reasoning integration and rare-case robustness in medical MLLMs, and a rigorous yardstick for future clinical systems.

cs.AI↗

arXiv2026

Can LLMs Fool Graph Learning? Exploring Universal Adversarial Attacks on Text-Attributed Graphs

Text-attributed graphs (TAGs) enhance graph learning by integrating rich textual semantics and topological context for each node. While boosting expressiveness, they also expose new vulnerabilities in graph learning through text-based adversarial surfaces. Recent advances leverage diverse backbones, such as graph neural networks (GNNs) and pre-trained language models (PLMs), to capture both structural and textual information in TAGs. This diversity raises a key question: How can we design universal adversarial attacks that generalize across architectures to assess the security of TAG models? The challenge arises from the stark contrast in how different backbones-GNNs and PLMs-perceive and encode graph patterns, coupled with the fact that many PLMs are only accessible via APIs, limiting attacks to black-box settings. To address this, we propose BadGraph, a novel attack framework that deeply elicits large language models (LLMs) understanding of general graph knowledge to jointly perturb both node topology and textual semantics. Specifically, we design a target influencer retrieval module that leverages graph priors to construct cross-modally aligned attack shortcuts, thereby enabling efficient LLM-based perturbation reasoning. Experiments show that BadGraph achieves universal and effective attacks across GNN- and LLM-based reasoners, with up to a 76.3% performance drop, while theoretical and empirical analyses confirm its stealthy yet interpretable nature.

cs.AI↗

arXiv2026

Uplink Networked Sensing via Multiuser Correlation Exploitation

In this correspondence, we investigate networked sensing in perceptive mobile networks under a bistatic multi-transmitter single-receiver uplink topology, where multiple user equipments (UEs) transmit signals over orthogonal frequency-division multiple access (OFDMA) resources and a single base station performs joint sensing. Uplink clock asynchronism introduces offsets that destroy inter-packet coherence and hinder high-resolution sensing, while multi-user observations exhibit exploitable cross-user correlation. We therefore formulate an asynchronous multi-user uplink OFDMA sensing model and exploit common delay-cluster sparsity across UEs. A line-of-sight (LoS)-referenced calibration first suppresses the offsets, after which a shared-private delay-domain sparse Bayesian learning (SBL) model is used for delay support recovery and user grouping. Doppler and angle of arrival are then estimated from temporal and spatial phase differences. Simulation results show that the proposed scheme outperforms per-user processing, particularly under limited subcarrier budgets and in low signal-to-noise ratio (SNR) regimes.

eess.SP↗

arXiv2026

Boosting Latent Diffusion Models via Disentangled Representation Alignment

Latent Diffusion Models (LDMs) rely heavily on the compressed latent space provided by Variational Autoencoders (VAEs) for high-quality image generation. Recent studies have attempted to obtain generation-friendly VAEs by directly adopting alignment strategies from LDM training, leveraging Vision Foundation Models (VFMs) as representation alignment targets. However, such alignment paradigms overlook the fundamental differences in representational requirements between LDMs and VAEs. Simple feature mapping from local patches to high-dimensional semantics can induce semantic collapse, leading to the loss of fine-grained attributes. In this paper, we reveal a key insight: unlike LDMs that benefit from high-level global semantics, a generation-friendly VAE must possess strong semantic disentanglement capabilities to preserve fine-grained, attribute-level information in a structured manner. To address this discrepancy, we propose the Semantic-Disentangled VAE (Send-VAE). Deviating from previous shallow alignment approaches, Send-VAE introduces a non-linear mapping architecture to effectively bridge the local structures of VAEs and the dense semantics of VFMs, thereby encouraging emergent disentangled properties in the latent space without explicit regularization. Extensive experiments establish a new paradigm for evaluating VAE latent spaces via low-level attribute separability and demonstrate that Send-VAE achieves state-of-the-art generation quality (FID of 1.21) on ImageNet 256x256.

cs.CV↗

arXiv2026

Iterative Closed-Loop Motion Synthesis for Scaling the Capabilities of Humanoid Control

Physics-based humanoid control relies on training with motion datasets that have diverse data distributions. However, the fixed difficulty distribution of datasets limits the performance ceiling of the trained control policies. Additionally, the method of acquiring high-quality data through professional motion capture systems is constrained by costs, making it difficult to achieve large-scale scalability. To address these issues, we propose a closed-loop automated motion data generation and iterative framework. It can generate high-quality motion data with rich action semantics, including martial arts, dance, combat, sports, gymnastics, and more. Furthermore, our framework enables difficulty iteration of policies and data through physical metrics and objective evaluations, allowing the trained tracker to break through its original difficulty limits. On the PHC single-primitive tracker, using only approximately 1/10 of the AMASS dataset size, the average failure rate on the test set (2201 clips) is reduced by 45% compared to the baseline. Finally, we conduct comprehensive ablation and comparative experiments to highlight the rationality and advantages of our framework.

cs.RO↗

arXiv2026

Spherical Latent Motion Prior for Physics-Based Simulated Humanoid Control

Learning motion priors for physics-based humanoid control is an active research topic. Existing approaches mainly include variational autoencoders (VAE) and adversarial motion priors (AMP). VAE introduces information loss, and random latent sampling may sometimes produce invalid behaviors. AMP suffers from mode collapse and struggles to capture diverse motion skills. We present the Spherical Latent Motion Prior (SLMP), a two-stage method for learning motion priors. In the first stage, we train a high-quality motion tracking controller. In the second stage, we distill the tracking controller into a spherical latent space. A combination of distillation, a discriminator, and a discriminator-guided local semantic consistency constraint shapes a structured latent action space, allowing stable random sampling without information loss. To evaluate SLMP, we collect a two-hour human combat motion capture dataset and show that SLMP preserves fine motion detail without information loss, and random sampling yields semantically valid and stable behaviors. When applied to a two-agent physics-based combat task, SLMP produces human-like and physically plausible combat behaviors only using simple rule-based rewards. Furthermore, SLMP generalizes across different humanoid robot morphologies, demonstrating its transferability beyond a single simulated avatar.

cs.RO↗

arXiv2026

Bistatic Passive Sensing via CSI Power

Passive object sensing with communication signals is a key enabler of perceptive mobile networks and integrated sensing and communication. In practical bistatic deployments, transmitter-receiver asynchrony and hardware impairments introduce time-varying random phase offsets in Channel State Information (CSI). Together with limited bandwidth and small antenna arrays, these effects degrade sensing accuracy. This work proposes a lightweight bistatic passive tracking and sensing framework that operates in the CSI-power domain. CSI power suppresses these offsets without explicit phase calibration, while preserving target-induced sensing cues. We show that physically admissible constraints in the spatial-frequency domain induced by transmitter-receiver geometry can resolve the mirror ambiguity inherent to real-valued CSI power. Building on these properties, we develop a real-time 3D Fourier-domain processing pipeline that jointly recovers spectral (delay), spatial (angle), and temporal (Doppler) signatures. The resulting features are integrated into an online framework with adaptive motion detection, outlier suppression, and extended Kalman filter tracking with deterministic initialization, followed by position-refined micro-Doppler feature extraction for micro-motion sensing. Extensive experiments, including simulations, a real-world prototype using 3.1 GHz LTE signals, and an open-source gait recognition dataset, demonstrate the effectiveness of the proposed CSI-power-based framework for bistatic passive tracking and sensing.

eess.SP↗

arXiv2026

Rethinking RSSI for WiFi Sensing

The Received Signal Strength Indicator (RSSI) is ubiquitously available on commodity WiFi devices but is commonly regarded as too coarse for fine-grained sensing. This paper revisits its sensing potential and presents WiRSSI, a bistatic WiFi sensing framework that enables RSSI-only passive human tracking and motion sensing. WiRSSI employs a transmitter and a receiver equipped with a three-antenna array (1Tx-3Rx), and is readily extensible to Multiple-Input Multiple-Output (MIMO) deployments. We first show how Channel State Information (CSI) power implicitly preserves phase-related motion modulation and how this relationship carries over to RSSI, indicating that RSSI can retain exploitable Doppler, Angle-of-Arrival (AoA), and delay cues. WiRSSI extracts Doppler-AoA features via a lightweight 2D Fast Fourier Transform (FFT) pipeline and infers bistatic delay from amplitude-only information in the absence of subcarrier-level phase. The estimated AoA and delay are then mapped to Cartesian coordinates and denoised to recover motion trajectories. Experiments in practical environments show that WiRSSI achieves median XY localization errors of 0.905 m, 0.784 m, and 0.785 m for elliptical, linear, and rectangular trajectories, respectively, compared with 0.574 m, 0.599 m, and 0.514 m from a representative CSI-based method. We further demonstrate RSSI-only gesture recognition on the Widar3.0 dataset, where WiRSSI features provide meaningful discriminative performance. These results suggest that, despite lacking subcarrier-level information compared with CSI, RSSI can support practical WiFi sensing as a complementary and hardware-friendly option when CSI is restricted, unreliable, or privacy-sensitive.

eess.SP↗

arXiv2026

BreathNet: Generalizable Audio Deepfake Detection via Breath-Cue-Guided Feature Refinement

As deepfake audio becomes more realistic and diverse, developing generalizable countermeasure systems has become crucial. Existing detection methods primarily depend on XLS-R front-end features to improve generalization. Nonetheless, their performance remains limited, partly due to insufficient attention to fine-grained information, such as physiological cues or frequency-domain features. In this paper, we propose BreathNet, a novel audio deepfake detection framework that integrates fine-grained breath information to improve generalization. Specifically, we design BreathFiLM, a feature-wise linear modulation mechanism that selectively amplifies temporal representations based on the presence of breathing sounds. BreathFiLM is trained jointly with the XLS-R extractor, in turn encouraging the extractor to learn and encode breath-related cues into the temporal features. Then, we use the frequency front-end to extract spectral features, which are then fused with temporal features to provide complementary information introduced by vocoders or compression artifacts. Additionally, we propose a group of feature losses comprising Positive-only Supervised Contrastive Loss (PSCL), center loss, and contrast loss. These losses jointly enhance the discriminative ability, encouraging the model to separate bona fide and deepfake samples more effectively in the feature space. Extensive experiments on five benchmark datasets demonstrate state-of-the-art (SOTA) performance. Using the ASVspoof 2019 LA training set, our method attains 1.99% average EER across four related eval benchmarks, with particularly strong performance on the In-the-Wild dataset, where it achieves 4.70% EER. Moreover, under the ASVspoof5 evaluation protocol, our method achieves an EER of 4.94% on this latest benchmark.

cs.SD↗

arXiv2026

Modeling Tidal Disruption Events and Compact Object Plunges in Nuclear Star Clusters

We study tidal disruption events (TDEs) and compact object inspirals in nuclear star clusters (NSCs) hosting a central supermassive black hole (SMBH), focusing on their role in SMBH growth. Using the STARDISK version of the direct N-body code NBODY6++GPU, we perform pilot simulations with two improved models: one for mass fallback from TDEs and another for compact object plunges based on orbital decay timescales. Our results show that mass accretion via TDEs peaks within the first 2 Myr and decreases more rapidly for higher initial SMBH masses, with roughly half the disrupted stellar debris being accreted. Compact object accretion is confined mostly to orbits with pericenters between 4 and 27 Schwarzschild radii and is suppressed by an order of magnitude when inspiral criteria are applied.

astro-ph.GA↗