Currently, artificial intelligence-generated content (AIGC) poses a severe threat to facial video deepfakes, and many AIGC companies employ video watermarking to mitigate unauthorized fake video usage, but its effectiveness is limited due to watermark tampering and private-model forgery. Additionally, some schemes actively add watermarks to user videos. However, current approaches focus primarily on the quality of watermark embedding in offline scenarios, and no existing work offers a solution for protecting human facial videos in real-time online applications. To fill this gap, we propose, to the best of our knowledge, a novel biometric video watermarking method, UltraGuard, for the first time. It uses the smartphone's built-in speaker and microphone to emit ultrasonic signals and capture biometric features, which are then fused with spatio-temporal dynamic features derived from the video content. The fused features are encoded as watermarks and embedded into videos in real time, enabling dynamic, content-aware watermarking throughout transmission and propagation. Extensive experimental validation demonstrates that UltraGuard can effectively defend against malicious attacks such as forgery and editing, both at the source and during propagation, proving the system's efficacy. In the future, it can be deployed on users' mobile devices to provide real-time protection for users' facial videos, thereby preventing security and privacy breaches caused by various malicious attacks.
AI agents are increasingly automating data analysis, but responsible adoption requires reliable performance evaluation, which is costly in specialized scientific domains. We introduce a framework that translates scientific publications into auditable benchmarks for AI agents and demonstrate its use on spatial transcriptomics alignment tasks. Across 121 tasks, 16 baseline methods, and six AI agents, the best agents (GPT-6-Astra and Gemini-4-Argon) matched or outperformed all baselines on 80% of tasks, often using surprisingly simple heuristics. Agent performance was robust to prompting and access to pre-installed packages. Benchmark construction also uncovered errors in prior work and shortcuts in current evaluation metrics. Finally, we show that agent-generated solutions can be distilled into a reusable program that ranks in the top 10% of baseline methods on unseen data. Our framework offers a scalable way to evaluate AI agents in specialized scientific domains by leveraging and stress-testing existing papers and metrics.
We introduce ChemInsight, a benchmark designed to evaluate large language models (LLMs) on chemistry tasks that require multi-step reasoning and structured chemical understanding. We propose a two-dimensional scoring system (scope and depth) to assess both task coverage and reasoning completeness. ChemInsight consists of 203 challenging chemistry questions organized into 50 bundles, which support the two-dimensional scoring evaluation scheme. We evaluate ten representative LLMs using ChemInsight and observe consistently low performance across models, with scope scores below 30% and depth scores below 11%, substantially lower than human expert performance. We then proceed to ask (1) What factor limits the advancement of current LLMs in chemistry? (2) To what extent are statistically driven language-model training methods effective at modeling chemistry symbolic language such as SMILES? And (3) Can LLM performance in chemistry be improved through intensive training on large, domain-specific corpora?
Deep learning models for brain tumor MRI classification are usually compared by accuracy alone, on public datasets whose independence is rarely checked. We benchmarked six architectures: two convolutional networks (EfficientNetV2-M and ConvNeXt-V2-Tiny), two vision transformers (Swin-V2-S and ViT-L/16), and two foundation models (DINOv2-L and MedSigLIP-448), for four-class classification of 7,216 public MRI images. We assessed accuracy, calibration, explainability, ensembles, and decision curves, with every fitted quantity estimated on validation data only. Pixel-verified duplicate screening showed that 36% of the test images had a near-duplicate in the training data. Test accuracy ranged from 94.5% to 95.6%, with no pairwise difference after Holm correction, and fell to 90.3% to 91.7% on the 773 test images free of duplicates and shared patients. Ensembles did not outperform the best single model. Attribution maps passed a model-randomization test but localized expert-annotated tumors poorly. On 1,675 external BDNeuro-MRI images that did not duplicate the training data, accuracy ranged from 77.6% (MedSigLIP-448) to 95.0% (ViT-L/16), and 12 of 15 pairwise differences were significant. Internal accuracy on this public collection therefore did not distinguish the architectures, whereas the duplicate-filtered external test did.
Preference optimization has emerged as a promising alternative to reward-based training for neural combinatorial optimization. However, existing methods typically rely on binary pairwise labels, which discard the magnitude of solution-quality differences and provide supervision only at the complete-solution level. We propose Advantage-aware Preference Optimization for Two-level Constructive Search (APO2CS), a preference-based training objective for constructive neural solvers that integrates both macro-level and micro-level supervision. At the macro level, APO2CS enriches pairwise preferences with an advantage term derived from objective-value differences. At the micro level, it compares partial trajectories, thereby providing denser supervision throughout the sequential construction process. The resulting objective can be used either to train a policy from scratch or to fine-tune a pretrained one. Experimental results show that APO2CS improves the expressiveness and search capability of preference learning for combinatorial optimization, while demonstrating strong effectiveness and adaptability across multiple combinatorial optimization tasks.
Multimodal face presentation attack detection (fPAD) improves robustness using complementary RGB, depth, and infrared (IR) cues, but deployment remains challenging when modalities are unavailable or incomplete. Existing flexible-modality methods mainly use zero-filling or modality omission, potentially discarding cross-modal information and introducing modality-dependent bias. We propose the Flexible Modality Adaptation Network (FMA-Net), a unified framework that compensates for missing modalities through instruction-guided latent-space translation. The Face Modality Instruct Translator (FMIT) estimates target modality observations from RGB inputs, while context-aware category enhancement uses learnable text prompts as modality-invariant semantic anchors. A Facial Feature Adapter (F2A) integrates global facial structure and local spoof-relevant cues to enhance discriminative representations. Synthetic-to-Real Face Domain Alignment (SR-FDA) reduces feature discrepancies between translated and real modalities, with benefits depending on the missing-modality pattern and data domain. Experiments on WMCA and PADISI-Face validate the framework. On WMCA with a 70% modality missing ratio, FMA-Net achieves a mean ACER of 3.22% across four intra-domain missing-modality protocols, compared with 4.24% for the strongest baseline by the same criterion. When both depth and IR are entirely unavailable, it achieves 5.23% ACER, compared with 5.99% for the best competing result. Cross-domain experiments also reveal limitations of synthetic-to-real alignment under dataset shifts. These results support FMA-Net as a practical approach to fPAD with flexible and incomplete modalities.
Robots are typically evaluated in episodes that begin from a reset, while industrial work consists of repeated jobs whose outcomes affect subsequent operations. Evaluating such work at scale requires simulation of task-critical physical interactions, construction of industrial environments, and measurements of performance as scene state persists. We present HOLIDAYFACTORY 1.0, a framework for simulating, generating, and evaluating industrial robot work. Given a human-specified workplace and task, the framework defines a world pairing a scenario of scene variants and process logic with an evaluation defining a job, metrics, and a rubric. HOLIDAYFACTORY 1.0 comprises three components. HOLIDAY-SIM extends Newton's SDF-based hydroelastic contact with additional constitutive laws and a semi-implicit primal-dual solver, couples stateful actuator dynamics to articulated motion, and derives tactile and joint observations from simulated states. Agentic world construction generates simulation-ready assets and scene variants using geometric and interaction checks, and proposes task-conditioned evaluations for human confirmation and executable checks. 24/7-BENCH introduces seven industrial tasks in which jobs are evaluated without resetting the scene, measuring completed, failed, and unfinished jobs, throughput, latency, and motor energy. Controlled studies characterize the proposed contact, actuator, and tactile models and examine the effectiveness of verification-guided world construction. Experiments on scene and evaluation generation further reveal the limitations of model critics and the difficulty of ensuring that generated evaluations faithfully capture task requirements. We examine job-level evaluations with scripted trajectories (e.g., code-as-policy) but have not yet compared policies over persistent runs.
Population game dynamics describe how aggregate behavior evolves in response to payoff signals. In many applications, the payoff function may change, and the goal is to predict the population trajectory induced by a new payoff design before trajectory data under that design are available. We study this problem as $\textbf{payoff-aware prediction of population game dynamics}$. Standard dynamics learning methods fit the motion observed under training payoffs, but do not separate the payoff-dependent incentive from the payoff-independent response structure needed for prediction under new payoffs. Building on the existing framework of Riemannian game dynamics, we propose PA-RmD, a payoff-aware learning method that keeps payoff information explicit and learns the payoff-independent response structure from data. Theoretically, we show that the learned model preserves fixed-payoff Nash equilibria and derive finite-time one-step prediction bounds under shared-response and payoff-coverage conditions. Empirically, we evaluate PA-RmD on synthetic zero-sum and potential games, as well as a YouTube Trending multi-country dataset, showing that our method performs well under changing payoff designs.
Electroencephalography (EEG) foundation models have emerged as promising paradigms for decoding brain activity through large-scale self-supervised pretraining. Notably, most existing approaches employ masked patch reconstruction, in which randomly selected EEG patches are masked and recovered by the network. However, purely random masking may limit representation learning as informative patches from highly correlated channels often remain unmasked, causing the network to underutilize weakly correlated channels or serial dependencies. Furthermore, because these models are typically based on transformer architectures, their performance is influenced by the choice of positional encoding. Although positional encoding has been extensively explored in natural language processing and computer vision, few studies have adapted it to the unique characteristics of EEG signals. To address these limitations, this study introduces a conditional positional encoding method based on graph convolutional networks (GCNs) to explicitly model inter-channel correlations, along with an adaptive masking strategy that enhances representation learning by modulating the difficulty of masked patch reconstruction. To the best of our knowledge, this is the first study to integrate graph-based representations into conditional positional encoding and introduce adaptive masking within EEG foundation model pretraining. Experimental results on ten benchmark datasets demonstrate that the proposed GCN-based positional encoding and adaptive masking strategies, both individually and in combination, achieve state-of-the-art or competitive performance, highlighting their effectiveness in EEG foundation modeling.
Sampling-based planners such as MPPI and CEM query a learned world model thousands of times per control step, so low-bit world models are an obvious way to make model-based control cheaper on a robot. With round-to-nearest activation quantization and no calibration, we find that planning can collapse, by up to 100 points of success, while the world model's latent one-step prediction error stays no larger than that of a stochastic rounding scheme that plans well. At the same bit width, what decides the outcome is not the size of the rounding error but whether it persists across planner iterations and is shared across candidates. Round-to-nearest leaves an error that persists, and an iterating planner drifts toward the candidates that this error favors; in TD-MPC2, holding the same dither fixed within a step or an episode brings the drift back. Dither shared by all candidates and redrawn at every iteration removes most of the damage at the same bit width and without training: at least half of it in all 19 affected Meta-World cells (task $\times$ bit width) of TD-MPC2 (median 95%), in 4 of 5 DINO-WM cells, and in all 8 affected tasks of a 317M-parameter multi-task TD-MPC2. Captured as CUDA graphs on an Ampere GPU, a W8A4 integer TD-MPC2 world model runs its planning rollouts 1.9-3.3$\times$ and the whole planning step 1.3-1.5$\times$ faster than FP32, though not faster than BF16 with pre-converted weights; shared dither adds at most 2.3% latency. In TD-MPC2, block-scaled NVFP4 removes most of the damage and MXFP4 part of it, which ties the failure to coarse grids with a shared scale. Hypotheses, thresholds and analysis code were fixed before the data that test them, and the hypotheses that failed are reported.
The rapid progress of generative speech models has made high-fidelity voice cloning and editing cheap, fast, and widely accessible, weakening the security of voiceprint-based authentication in mobile speech interactions. Existing speech watermarking techniques focus on offline protection or watermarking synthetic speech, and thus cannot verify whether a live utterance originates from the genuine speaker at capture time. We present \textit{MoBiMark}, a real-time \textbf{mo}bile \textbf{bi}ologic speech water\textbf{mark}ing system that binds ultrasonic articulatory signatures to speech content and embeds a dynamic watermark during capture. MoBiMark emits a multi-frequency ultrasonic probe on commodity smartphones, extracts echo features from the vocal tract, and fuses them with a content-dependent speech embedding to produce a dual-layer dynamic watermark combining a persistent user identifier (UID) and a per-utterance speech identifier (SID), enabling both source authentication and propagation-time tamper detection. A jointly trained embedder and detector injects the watermark while preserving perceptual quality, and system-level optimization with INT8 quantization enables lightweight on-device deployment. Evaluated on 50 users across four smartphone models, MoBiMark achieves an average 99.79% accuracy in source and propagation attacks, with negligible impact on speech quality and intelligibility.
Distributed artificial intelligence (AI) increasingly relies on wireless networks, where communication and computing resources shape how AI tasks are configured and executed. However, resource availability alone does not determine how effectively these tasks can be supported, since different information selections and transmission strategies may yield different AI performance under the same resource constraints. To address this challenge, this article presents Wireless AI Networks as a unified design perspective for bidirectional adaptation between wireless communication and distributed intelligence. Specifically, an AI-executable resource budget characterizes candidate AI configurations, while task-level information value guides their selection and wireless resource allocation. Building on these principles, a wireless-AI co-adaptation framework coordinates resource abstraction, joint configuration and allocation, and execution feedback. Two complementary experiments illustrate this perspective: update selection and statistical correction improve federated learning under identical model-update payload budgets, while sensitivity-aware power allocation reduces model accuracy degradation under the same average transmit-power constraint. Together, these results highlight the importance of coordinating resource feasibility and task-level information value in distributed AI execution.
Retrieval-Augmented Generation (RAG) systems have demonstrated remarkable effectiveness in mitigating hallucinations by incorporating external knowledge. However, the retrieval process inevitably introduces noise, posing significant challenges to RAG robustness. Fundamentally, noise robustness is a composite competency encompassing noise identification, retrieval filtering, and robust reasoning. Unfortunately, existing methods often fail to capture this holistic nature, typically restricted to single-dimensional enhancements or hampered by noise-induced optimization interference. To address these limitations, we decompose RAG robustness into three sub-tasks: knowledge boundary perception, retrieval verification, and robust reasoning. Accordingly, we propose ARMOR, an Adaptive Curriculum Meta-Learning for Noise Robust RAG Reasoning. Inspired by the teaching process, ARMOR introduces an adaptive task scheduler as a teacher to curate training plans according to the absolute learning progress of RAG generators (Student). Distinct from GRPO, ARMOR utilizes inter-group advantage estimation during learning to effectively mitigate mode collapse caused by limited output spaces. Experiments on three complex reasoning datasets (2WikiMultiHopQA, MusiQue, and HotPotQA) indicate that ARMOR achieves synergistic improvements in retrieval noise robustness, significantly outperforming baselines in output quality.
Global structure-from-motion (SfM) commonly refines relative poses on individual image pairs and then estimates global rotations by rotation averaging. Pairwise refinement does not explicitly combine feature-level constraints with multi-view consistency during rotation initialization. We introduce Global Rotations from Aggregated Matches (GRAM), which estimates global rotations directly from initial relative rotations and inlier matches by minimizing edge-wise rotation-only algebraic residuals. GRAM replaces two-view relative-pose optimization and rotation averaging with a single stage. We further extend rotation- only imaging geometry to multiple views and develop Multi-view Optimization with Rotation-only Reprojection Error (MORE) for nonlinear global rotation refinement. Analytical Rotation-guided Translation (ART) derives locally scaled triplet translations from Linear Global Translation (LiGT) constraints and recovers global translations using weighted least squares. Synthetic and real-world experiments demonstrate the accuracy, robustness, and efficiency of the proposed methods.
Monocular depth estimation has advanced through large-scale pretraining and strong visual backbones, yet metric depth remains sensitive to scene-level scale. Indoor rooms and outdoor driving scenes occupy very different depth ranges, but most depth decoders infer this global structure implicitly from visual features. We ask whether scene-level semantic structure can instead act as an upstream prior that shapes the metric output space before dense prediction. We present LangDepth (SemanticBins), a monocular depth architecture that couples strong visual features with a scene-conditioned output space. A frozen CLIP encoder and a learnable bank of scene- and geometry-oriented concept keys convert image features into a compact scene descriptor, without requiring per-image text or an external captioner. A Scene Prior Module (SPM) maps this descriptor to a scene-conditioned metric depth interval, providing a global range anchor before dense decoding. In parallel, a DINOv2-L backbone with a lightweight FPN extracts dense visual features, and a Per-Pixel Assignment Network (PPAN) assigns pixels to the scene-prescribed bin structure. A confidence gate fuses the dense visual estimate with the SPM prior. We train LangDepth jointly on NYU Depth V2 and KITTI under a modest supervised depth budget, without web-scale depth supervision or pseudo-labelled depth data. A single model achieves delta_1 = 0.960 on NYU, 0.960 on KITTI, 0.887 on DIODE indoor, 0.946 on iBims-1, and 0.861 on SUN RGB-D. Beyond aggregate accuracy, the SPM predicts compact indoor ranges and much wider outdoor KITTI ranges, the gate becomes SPM-dominant on outdoor scenes, and perturbing the routed scene context causes large drops on KITTI and iBims-1. These results support scene-conditioned output spaces as a practical and interpretable mechanism for cross-domain monocular depth estimation.
LLM inference has two phases with different bottlenecks: LLM prefill runs once per request and is compute-intensive. Decoding repeats for every output token and is limited by memory bandwidth. To address the challenges associated with this issue, different solutions have been proposed. All of these solutions use the same model for both prefill and decode phases. We propose BridgeKV, which uses a larger model for prefill and a smaller model of the same family for decoding. A lightweight KV-stitching module translates the larger model's prompt cache into the smaller model's cache layout. It handles differences in cache shape and representation through learned local layer mixing and separate nonlinear maps for keys and values. With both models frozen, we train the module to align their caches and make the smaller model follow the larger model's next-token predictions. A second training stage adds low-rank adapters when further task adaptation is needed. Across six model pairs and five datasets, BridgeKV outperforms fine-tuned small-model baselines with matched trainable parameter budgets in all 30 settings. It closes up to 80% of the quality gap to using the larger model for both phases. On one A40 GPU, Llama-3.1-Instruct-8B to Llama-3.2-Instruct-1B achieves 3.0X faster requests than the larger model alone, with latency close to the smaller model alone. BridgeKV can improve the hardware and system solutions that separate prefill and decode phases across GPUs. A separate serving simulation models request arrivals, queues, and cache transfer. For Llama-3.1-Instruct-70B to Llama-3.2-Instruct-1B, it predicts up to 33.86X faster requests and 27.5X lower energy use than using the larger model for both prefill and decode phases.
Compact object detectors such as YOLOv8n are a common choice for camera-based fire and smoke detection on low-power devices. These detectors are usually evaluated with box-level metrics such as mean average precision, and their outputs are turned into decisions with a fixed confidence threshold such as 0.25 or 0.5. A fire monitoring system, however, makes one decision per image: raise an alarm or not. In this paper we evaluate two publicly available compact detectors on the D-Fire test set at the image level. We compare three simple rules for turning box detections into one alarm score per image, and we use a split conformal procedure to choose the alarm threshold from held-out images so that the expected fraction of missed fire and smoke images stays below a chosen target. We find that the same fixed threshold leads to very different miss rates for the two detectors, 17.3 and 8.4 percent at a threshold of 0.5, while the calibrated threshold meets miss-rate targets of 10, 5 and 1 percent for both detectors and all three rules. At a 5 percent target, the false alarm rate on images without fire or smoke stays between 1.2 and 3.0 percent. We also find that smoke-only images stay below the target even when the overall target is met, and we describe the main failure cases.
Video snapshot compressive imaging (SCI) recovers multiple frames from one spatiotemporally coded measurement. This paper proposes SpinSCI, a mask-conditioned Spatial-Temporal Spin Transformer whose Mask-Conditioned Ising Refinement Block (MC-IRB) combines attention logits and sensing information to form signed spin couplings. Finite mean-field updates gate residual feature refinement. Physics-guided initialization and final data consistency incorporate the measurement model. Compared with HiSViT$_9$, SpinSCI achieves average PSNR gains of 0.51 dB, 0.23 dB, and 0.04 dB on small-scale gray, mid-scale color, and large-scale color benchmarks, respectively, with 16.9% fewer parameters, 21.0-21.6% fewer FLOPs, 17.1-49.9% lower latency, and 62.3-63.3% lower peak memory. Relative to DU-9stg, average PSNR differs by at most 0.13 dB across these benchmarks, with 28.2-28.7% fewer FLOPs, 64.6-71.7% lower inference latency, and 30.1-35.0% lower peak memory. Qualitative comparisons on the real benchmark further show recovery of fine details and textures. Code and reference checkpoints are available at https://github.com/skethz/SpinSCI.
Language-model post-training modifies weights through cumulative updates and successive displacements, yet existing studies report seemingly conflicting spectral properties: nearly full-rank updates and low stable-rank changes. We show these observations are compatible and characterize update directions during training. For singular values following $\sigma_i \propto i^{-\alpha}$ with $0 < \alpha \leq 1/2$, we prove that stable rank becomes negligible relative to fixed-energy rank as dimension increases. Full singular value decompositions of selected OLMo-2-7B updates confirm this separation: three proximal policy optimization (PPO) attention updates have stable ranks of 9.6-18.3 but require 532-1,222 singular components to capture 80% of their energy. We next compute consecutive checkpoint-displacement cosines from full-model inner products. Across nine publicly released reinforcement learning from verifiable rewards (RLVR) stages, all 131 adjacent pairs have negative cosines on their native checkpoint grids, including PPO and Group Relative Policy Optimization (GRPO) trajectories. Supervised fine-tuning comparisons across four model families show configuration-dependent alignment, including negative late-training cosines for Mistral and FP32 Llama. Finally, a controlled OLMo-2-1B study with three paired seeds and a 2,048-token response cap compares group-relative advantages against a learned critic with generalized advantage estimation (GAE). At matched 60-rollout windows, late-training mean cosines are +0.259 and +0.012, respectively; the paired critic-minus-group difference is -0.247 (seed-level 95% interval [-0.358, -0.137]). These findings distinguish spectral concentration, energy coverage, and trajectory direction, and demonstrate that the advantage system influences directional alignment under matched training conditions.
When a mesh reconstructed from aerial remote-sensing imagery disagrees with reference geometry, deciding what to change requires a diagnosis of the reconstruction process. We present GeoPilot, a workflow that turns diagnostics and prior observations into executable program or camera-configuration interventions. A large language model proposes the change, numerical tools reconstruct the mesh, and an external evaluator measures its geometry under a stated coordinate and reference protocol. The main case tests a shared-intrinsics hypothesis proposed before the comparison. On 224 Metashape-raster development images, shared intrinsics improved all six native-coordinate measures and reduced mean surface-to-reference distance by 7.04%. Reference-assisted alignment fitted and evaluated on separate within-scene checkerboard samples reversed the mean-distance ranking, limiting that gain to native-frame delivery. A separate matched comparison on publisher-undistorted images favored fixed published cameras on all six measures, reducing mean distance by 12.11%; the cameras retain upstream image and laser-scanning adjustment information. With one reconstruction per configuration, these comparisons provide conditional evidence for camera handling. Three end-to-end runs each delivered a mesh and six-metric evaluation from all 327 images of a second development scene.