Spatio-temporal epidemic forecasting is critical for public health management, yet existing methods often struggle with insensitivity to weak epidemic signals, over-simplified spatial relations, and unstable parameter estimation. To address these challenges, we propose the Spatio-Temporal priOr-aware Epidemic Predictor (STOEP), a novel hybrid framework that integrates implicit spatio-temporal priors and explicit expert priors. STOEP consists of three key components: (1) Case-aware Adjacency Learning (CAL), which dynamically adjusts mobility-based regional dependencies using historical infection patterns; (2) Space-informed Parameter Estimating (SPE), which employs learnable spatial priors to amplify weak epidemic signals; and (3) Filter-based Mechanistic Forecasting (FMF), which uses an expert-guided adaptive thresholding strategy to regularize epidemic parameters. Extensive experiments on real-world COVID-19 and influenza datasets demonstrate that STOEP outperforms the best baseline by 11.1% in RMSE. The system has been deployed at a provincial CDC in China to facilitate downstream applications.

Convolutional networks require extensive image annotation, which can be costly and time-consuming. Feature Learning from Image Markers (FLIM) tackles this challenge by estimating encoder filters (i.e., kernel weights) from user-drawn markers on discriminative regions of a few representative images without traditional optimization. Such an encoder combined with an adaptive decoder comprises a FLIM network fully trained without backpropagation. Prior research has demonstrated their effectiveness in Salient Object Detection (SOD), being significantly lighter than existing lightweight models. This study revisits FLIM SOD and introduces FLIM-Bag of Feature Points (FLIM-BoFP), a considerably faster filter estimation method. The previous approach, FLIM-Cluster, derives filters through patch clustering at each encoder's block, leading to computational overhead and reduced control over filter locations. FLIM-BoFP streamlines this process by performing a single clustering at the input block, creating a bag of feature points, and defining filters directly from mapped feature points across all blocks. The paper evaluates the benefits in efficiency, effectiveness, and generalization of FLIM-BoFP compared to FLIM-Cluster and other state-of-the-art baselines for parasite detection in optical microscopy images.

Long-range graph tasks are challenging for Graph Neural Networks (GNNs): global mechanisms such as attention or rewiring schemes can be computationally expensive, while deep local propagation is prone to vanishing gradients, oversmoothing, and oversquashing. The introduced mu-ChebNet architecture is a simple spectral GNN that learns a node-wise weight function mu before applying ChebNet-style filters. The learned weighting mu induces a modified graph Laplacian which effectively changes the propagation geometry without altering the graph topology. This task-dependent geometry promotes preferred routes for information propagation, thereby helping long-range signals avoid highly contractive bottlenecks, and obviating the need for repeated layer stacking. In practice, we replace the fixed graph Laplacian L by a learned operator L_mu, keeping the proposed mu-ChebNet architecture lightweight while making propagation task-adaptive. Furthermore, we provide a spectral analysis demonstrating how mu modulates propagation dynamics, and empirically observe improved performance on both synthetic long-range reasoning tasks and real-world graph benchmarks. The learned weight function is not only interpretable, but also offers a lightweight alternative to attention and rewiring for adaptive graph propagation.

Augmented reality (AR) guidance in laparoscopic liver surgery requires accurate registration of preoperative 3D models to intraoperative 2D video, but remains challenging due to partial visibility, specularities, and tissue deformation. Existing methods often rely on contour-based rigid initialization and finite-element (FE) models for deformable registration, increasing modeling and engineering complexity. We present a depth-augmented, FE-free 3D--2D registration pipeline that combines robust rigid initialization with patient-specific non-rigid refinement. For rigid alignment, we adapt the RefineNet module of FoundationPose to laparoscopic liver scenes by using multi-class contour maps and monocular depth for relative pose refinement. For deformable alignment, we construct a patient-specific statistical deformation model from non-rigid ICP (NICP) correspondences and optimize pose and shape parameters using a coarse-to-fine L-BFGS-B strategy. On a public clinical laparoscopic liver dataset, the proposed method achieves a mean target registration error (TRE) of 14.73\,mm under a controlled manual-contour setting designed to isolate registration performance. Ablation studies show that monocular depth improves rigid initialization over contour-only inputs, while tumor-mapping analysis indicates that good surface alignment does not necessarily translate into lower target localization error. On an external dataset without ground truth, the method produces visually plausible overlays for qualitative assessment. These results suggest that depth-augmented pose refinement and FE-free statistical deformation modeling provide a promising alternative to FE-based pipelines for controlled 3D--2D liver registration in surgical AR.

Task Arithmetic yields a modular, scalable way to adapt foundation models. Combining multiple task vectors, however, can lead to cross-task interference, causing representation drift and degraded performance. Representation drift regularization provides a natural remedy to disentangle task vectors; however, existing approaches typically require external task data, conflicting with modularity and data availability constraints (e.g., privacy requirements). We propose a dataless approach by framing regularization against representation drift as a curvature matrix approximation problem. This allows us to leverage well-established techniques; in particular, we adopt Kronecker-Factored Approximate Curvature and obtain a practical regularizer that achieves state-of-the-art results in task addition and negation. Our method has constant complexity in the number of tasks and promotes robustness to task vector rescaling, eliminating the need for held-out tuning.

Evaluating moral alignment in agents navigating conflicting, hierarchically structured human norms is a critical challenge at the intersection of AI safety, moral philosophy, and cognitive science. We introduce Morality Chains, a novel formalism for representing moral norms as ordered deontic constraints, and MoralityGym, a benchmark of 98 ethical-dilemma problems presented as trolley-dilemma-style Gymnasium environments. By decoupling task-solving from moral evaluation and introducing a novel Morality Metric, MoralityGym allows the integration of insights from psychology and philosophy into the evaluation of norm-sensitive reasoning. Baseline results with Safe RL methods reveal key limitations, underscoring the need for more principled approaches to ethical decision-making. This work provides a foundation for developing AI systems that behave more reliably, transparently, and ethically in complex real-world contexts.

Adapting large pre-trained models to downstream tasks often produces task-specific parameter updates that are expensive to relearn for every model variant. While recent work has shown that such updates can be transferred between models with identical architectures, transferring them across models of different widths remains unexplored. In this work, we introduce Theseus, a training-free method for transporting task updates across heterogeneous-width models. Rather than matching parameters, we characterize a task update by the functional effect it induces on intermediate representations. We formalize task-vector transport as a functional matching problem on observed activations and show that, after aligning representation spaces via orthogonal Procrustes analysis, it admits a stable closed-form solution that preserves the geometry of the update. We evaluate Theseus on vision and language models across different widths, showing consistent improvements over baselines without additional training or backpropagation. Our results show that task updates can be meaningfully transferred across architectures when task identity is defined functionally rather than parametrically. Code is available at https://github.com/apanariello4/merge-and-rebase.

Reinforcement learning (RL) finetuning has become a key technique for enhancing large language models (LLMs) on reasoning-intensive tasks, motivating its extension to vision-language models (VLMs). While RL-tuned VLMs improve on visual reasoning benchmarks, they remain vulnerable to weak visual grounding, hallucinations, and over-reliance on textual cues. We show that simple, controlled textual perturbations, including misleading captions or incorrect chain-of-thought (CoT) traces, cause substantial drops in robustness and confidence, and that these effects are more pronounced when CoT consistency is taken into account across open-source multimodal reasoning models. In contrast, closed models exhibit similar failure modes but maintain markedly greater robustness and reasoning consistency, suggesting that the gap reflects a shortcoming in current open-source RL finetuning rather than an inherent limitation of the task. To better understand these vulnerabilities, we further analyze RL finetuning dynamics and uncover an accuracy-faithfulness trade-off: finetuning raises benchmark accuracy, but can simultaneously erode the reliability of the accompanying CoT and its robustness to contextual shifts. Although adversarial augmentation improves robustness, it does not by itself prevent faithfulness drift. Incorporating a faithfulness-aware reward can restore alignment between answers and reasoning, but when paired with augmentation, training risks collapsing onto shortcut strategies and robustness remains elusive. Together, these findings highlight the limitations of accuracy-only evaluations and motivate training and assessment protocols that jointly emphasize correctness, robustness, and the faithfulness of visually grounded reasoning.

Two-player games such as board games have long been used as traditional benchmarks for reinforcement learning. This work revisits a policy optimization method with reverse Kullback-Leibler regularization and entropy regularization and analyzes this combination in two-player zero-sum settings from theoretical and empirical perspectives. From a theoretical perspective, we investigate the stability of the policy update rule in two theoretical settings: game-theoretic normal-form games and finite-length games. We provide novel convergence guarantees and verify our theoretical results through numerical experiments on synthetic games. From an empirical perspective, we derive a practical model-free reinforcement learning algorithm based on the regularized policy optimization. We validate the training efficiency of our algorithm through comprehensive experiments on five board games: Animal Shogi, Gardner Chess, Go, Hex, and Othello. Experimental results show that our agent learns more efficiently than existing methods across environments.

A core quality of general intelligence is the ability to open-endedly expand and evolve its set of mastered skills autonomously. While recent Foundation Model (FM) driven approaches have shown promising results towards this goal, they typically rely on significant human-in-the-loop engineering, limiting their transferability to novel environments. To address this, we introduce Continuous Open-ended Discovery and Evolution of Skills as Hierarchical Reward Programs (CODE-SHARP), a framework that leverages FMs to open-endedly grow and evolve an archive of Python programs encoding skills to train a generalist agent policy entirely from scratch via reinforcement learning, directly from source code. These programs, termed Skills as Hierarchical Reward Programs (SHARPs), each encode a local success condition and a set of prerequisites delegated to previously discovered SHARPs. At runtime, SHARPs dynamically route the agent through their prerequisite chain based on the current state, rewarding each completion along the way, requiring the agent to learn only the marginal behaviour each new SHARP introduces, enabling efficient learning of long-horizon skills without any pre-defined rewards. On Craftax-Classic and XLand, agents trained fully autonomously by CODE-SHARP outperform previous works by 6x and 2.6x in median performance and are the only agents capable of crafting iron tools and mining diamonds. Scaled to Craftax-Extended, CODE-SHARP trains a generalist agent on over 90 discovered SHARPs, enabling the agent to solve challenging long-horizon tasks zero-shot, matching agents trained on ground-truth rewards.

Large Language Models (LLMs) are unable to reliably reason about specific physical systems. Attempts to imbue LLMs with knowledge of the necessary physics concepts have shown great promise, but explainability and validation remain open challenges. An emerging alternative is tooling, where LLMs can query physical simulators and use the resulting simulation traces as context for validation. This approach suffers from poor scalability since simulation traces contain large volumes of fine-grained numerical and semantic data. We show that translating simulation traces to a sparse representation of "high-level" structural patterns leads to more effective interpretation by LLMs. We propose an unsupervised learning scheme to perform this translation, or annotation, via program synthesis. Our learning results in a library of programs that act as pattern detectors which can translate simulation traces to sparse, annotated pattern sequences. The detected patterns may optionally be guided by human experts via string labels (rigid collision, stretching spring, etc.). We show, using a recent physics benchmark, that such annotated representations are more amenable to natural language reasoning about specific physical systems. The synthesized programs serve as transparent, explainable functions that map system states to a sparse and efficient annotation space. As an example application, we show how goals within physical systems that are specified in natural language may be converted to reward programs which are maximized to find solutions.

Feature Engineering (FE) is pivotal in automated machine learning (AutoML) but remains a bottleneck for traditional methods, which operate within rigid search spaces and lack domain awareness. While Large Language Models (LLMs) offer a promising alternative to generate unbounded operators with semantic reasoning, existing methods focus on isolated subtasks such as feature generation, falling short of free-form FE pipelines. Moreover, they are rarely coupled with hyperparameter optimization (HPO) of the downstream ML model, leading to greedy "FE-then-HPO" workflows that cannot capture strong FE-HPO interactions. In this paper, we present CoFEH, a collaborative framework that interleaves LLM-based FE and Bayesian HPO for robust end-to-end AutoML. CoFEH uses an LLM-driven FE optimizer powered by Tree of Thought (TOT) to explore flexible FE pipelines, a Bayesian optimization (BO) module to solve HPO, and a dynamic optimizer selector that adaptively interleaves FE and HPO steps. Crucially, we introduce a mutual conditioning mechanism that shares context between LLM and BO, enabling mutually informed decisions. Experiments show that CoFEH outperforms both traditional and LLM-based baselines in both standalone FE and joint FE+HPO settings.

The long-standing tension between Pre- and Post-Norm remains an open problem in Transformer architecture, reflecting a fundamental trade-off between training stability and representational capacity. Prior attempts to combine their strengths have made progress, but often show limited robustness across training settings, restricting their broader applicability. We revisit this dilemma, showing that single-stream architectures struggle to reconcile Pre-Norm's stable identity-gradient propagation with Post-Norm's normalization of the main residual path. To address this structural tension, we propose SiameseNorm, a simple yet effective two-stream architecture that remains compatible with Pre-Norm training recipes. SiameseNorm couples Pre-Norm-like and Post-Norm-like streams through shared residual blocks, allowing each residual block to receive optimization signals from both pathways with negligible overhead. Extensive experiments on 400M and 1.3B dense language models, 15B MoE models, Vision Transformers, and Diffusion Transformers show that SiameseNorm consistently improves performance while maintaining strong training stability across architectures and modalities. Code is available at https://github.com/Qwen-Applications/SiameseNorm.

Safety alignment of large language models remains brittle under domain shift and noisy preference supervision. Most existing robust alignment methods focus on uncertainty in alignment data, while overlooking optimization-induced fragility in preference-based objectives. In this work, we revisit robustness for LLM safety alignment from an optimization geometry perspective, and argue that robustness failures cannot be addressed by data-centric methods alone. We propose \textit{ShaPO}, a geometry-aware preference optimization framework that enforces worst-case alignment objectives via selective geometry control over alignment-critical parameter subspace. By avoiding uniform geometry constraints, ShaPO mitigates the over-regularization that can harm robustness under distribution shift. We instantiate ShaPO at two levels: token-level ShaPO stabilizes likelihood-based surrogate optimization, while reward-level ShaPO enforces reward-consistent optimization under noisy supervision. Across diverse safety benchmarks and noisy preference settings, ShaPO consistently improves safety robustness over popular preference optimization methods. Moreover, ShaPO composes cleanly with data-robust objectives, yielding additional gains and empirically supporting the proposed optimization-geometry perspective. The code is available at https://github.com/liujilong0116/ShaPO.

This paper surveys the state-of-the-art in the nexus of SLAM and Wireless Communications, attributing the bidirectional impact of each with a focus on visual SLAM (V-SLAM) integration. We provide an overview of key concepts related to wireless signal propagation, geometric channel modeling, and radio frequency (RF)-based localization and sensing. In addition to this, we show image processing techniques that can detect landmarks, proactively predicting optimal paths for wireless channels. Several dimensions are considered, including the prerequisites, techniques, background, and future directions and challenges of the intersection between SLAM and wireless communications. We analyze estimation and control approaches such as Bayesian filters, feature-based pose estimation, perception-aware motion control, spatial methods for signal processing such as vector fields, and key technological aspects. We expose techniques and items towards enabling a highly effective retrieval of the autonomous robot state. Among other interesting findings, we observe that monocular V-SLAM would benefit from RF relevant information, as the latter can serve as a proxy for the scale ambiguity resolution. Conversely, we find that wireless communications in the context of 5G and beyond can potentially benefit from visual odometry that is central in SLAM. Moreover, we examine other sources besides the camera for SLAM and describe the twofold relation with wireless communications. Finally, integrated solutions performing joint communications and SLAM appear to be in their infancy: theoretical and practical advancements are required to add higher-level localization and semantic perception capabilities to RF and multi-antenna technologies.

Fake Image Detection (FID), aiming at unified detection across four image forensic subdomains, is critical in real-world forensic scenarios. Compared with ensemble approaches, monolithic FID models are theoretically more promising, but to date, consistently yield inferior performance in practice. In this work, we identify the intrinsic distinctness of artifacts across subdomains, a critical barrier we term the ``Ji-Zhe phenomenon". Driven by this phenomenon, we diagnose the cause of this underperformance for the first time: the collapse of the artifact feature space. The core challenge for developing a practical monolithic FID model thus boils down to the ``unified-yet-discriminative" reconstruction of the artifact feature space. To address this paradoxical challenge, we hypothesize that high-level semantics can serve as a structural prior for the reconstruction, and further propose Semantic-Induced Constrained Adaptation (SICA), the first monolithic FID paradigm. Extensive experiments on our OpenMMSec dataset demonstrate that SICA outperforms 15 state-of-the-art methods and reconstructs the target unified-yet-discriminative artifact feature space in a near-orthogonal manner, thus firmly validating our hypothesis. The code and dataset are available at: https://github.com/venus-guangjian/SICA_OpenMMSec.

Bayesian (deep) neural networks (BNN) are often more attractive than the vanilla point-estimate deep learning in various aspects including uncertainty quantification, robustness to noise, resistance to overfitting, and more. The variational inference (VI) is one of the most widely adopted approximate inference methods. Whereas the ELBO-based variational free energy method is a dominant choice in the literature, in this paper we introduce a score-based alternative for BNN variational inference. Score-based VI can address the known issue of mode collapsing in ELBO-based VI. Although several score-based VI methods have been proposed in the community, most are not adequate for large-scale BNNs for various computational and technical reasons. We propose a novel scalable VI method where the learning objective combines the score matching loss and the proximal penalty term in iterations, which helps our method avoid the reparametrized sampling, and allows for noisy unbiased mini-batch scores through stochastic gradients. This in turn makes our method scalable to large-scale neural networks including Vision Transformers. On several benchmarks including visual recognition and time-series forecasting with large-scale deep networks, we empirically show the effectiveness of our approach.

Model merging combines multiple fine-tuned models into a single model by adding their weight updates, providing a lightweight alternative to retraining. Existing methods primarily target resolving conflicts between task updates, leaving the failure mode of over-counting shared knowledge unaddressed. We show that when tasks share aligned spectral directions (i.e., overlapping singular vectors), a simple linear combination repeatedly accumulates these directions, inflating the singular values and biasing the merged model toward shared subspaces. To mitigate this issue, we propose Singular Value Calibration (SVC), a training-free and data-free post-processing method that quantifies subspace overlap and rescales inflated singular values to restore a balanced spectrum. Across vision and language benchmarks, SVC consistently improves strong merging baselines and achieves state-of-the-art performance. Furthermore, by modifying only the singular values, SVC improves the performance of Task Arithmetic by 13.0%. Code is available at https://github.com/lyymuwu/SVC.

Stochastic variance-reduced algorithms such as Stochastic Average Gradient (SAG) and SAGA, and their deterministic counterparts like the Incremental Aggregated Gradient (IAG) method, have been extensively studied in large-scale machine learning. Despite their popularity, existing analyses for these algorithms are disparate, relying on different proof techniques tailored to each method. Furthermore, the original proof of SAG is known to be notoriously involved, requiring computer-aided analysis. Focusing on finite-sum optimization with smooth and strongly convex objective functions, our main contribution is to develop a single unified convergence analysis that applies to all three algorithms: SAG, SAGA, and IAG. Our analysis features two key steps: (i) establishing a bound on delays due to stochastic sub-sampling using simple concentration tools, and (ii) carefully designing a novel Lyapunov function that accounts for such delays. The resulting proof is short and modular, providing the first high-probability bounds for SAG and SAGA that can be seamlessly extended to non-convex objectives and Markov sampling. As an immediate byproduct of our new analysis technique, we obtain the best known rates for the IAG algorithm, significantly improving upon prior bounds.

Graph-structured data underpins many critical applications. While foundation models have transformed language and vision via large-scale pretraining and lightweight adaptation, extending this paradigm to general, real-world graphs is challenging. In this work, we present Graph Billion-Foundation-Fusion (GraphBFF): an end-to-end recipe for building billion-parameter Graph Foundation Models (GFMs) for large-scale heterogeneous graphs. Central to the recipe is the GraphBFF Transformer, a flexible and scalable architecture designed for practical billion-scale GFMs. Using the GraphBFF, we present neural scaling laws for heterogeneous graphs and show that loss decreases predictably as either model capacity or training data scales, depending on which factor is the bottleneck. The GraphBFF framework provides concrete methodologies for data batching, pretraining, and fine-tuning for building GFMs at scale. We demonstrate the effectiveness of the framework over a real-world billion-scale graph, with an evaluation of a billion-parameter GraphBFF Transformer following the proposed recipe. Across ten diverse, real-world downstream tasks on graphs unseen during training, spanning node- and link-level classification and regression, GraphBFF consistently outperforms baselines, with large margins of up to 31 PRAUC points, including in few-shot settings. Finally, we discuss key challenges and open opportunities for making GFMs a practical and principled foundation for graph learning at industrial scale.

Long-context LLM agents often struggle with growing token, memory, and latency costs, making efficient context compression essential for practical deployment. Existing LLM-as-a-compressor methods remain noticeably inferior to using the full context. We find that this gap partly stems from their inability to preserve contextual information effectively. In this work, we revisit context compression from a structural perspective and identify two key bottlenecks in standard LLM-based compressors: limited coordination among compression tokens during information aggregation, and layerwise dilution that weakens useful signals from intermediate hidden states. To address these limitations, we propose ComprExIT, a new context compression framework based on explicit information transmission. ComprExIT adaptively selects features across frozen LLM layers, then allocates information from anchors to compression slots through a globally coordinated transport plan. Experiments on 12 datasets show that ComprExIT consistently outperforms strong soft-compression baselines, improving average F1 by up to 18.5%, while adding only ~1% trainable parameters and achieving more than 2x faster compression than the fastest baselines. The code will be released upon acceptance.

Recent multi-LLM agent systems have shown promising capabilities for automated problem-solving, yet they predominantly rely on frozen agents or static fine-tuning pipelines. To address this limitation, our primary contribution is ATLAS (Adaptive Task-distributed Learning for Agentic Self-evolution), a multi-agent framework where specialized meta-agents collaboratively train and refine an active agent toward a domain-specific policy. A core challenge in iterative preference learning within these pipelines is the reliance on fixed reference models, which typically leads to overly conservative updates or training stagnation. To overcome this, the framework's algorithmic engine utilizes Evolving Direct Preference Optimization (EvoDPO). EvoDPO employs an inspection agent to perform adaptive, proxy-KL gated reference policy updates based on continuous training telemetry. We evaluate this full framework across a diverse set of challenging environments-including non-stationary contextual bandits, partial differential equations (PINNs), and combinatorial optimization tasks (TSP, Bin Packing). Through comparison against fixed-reference, adaptive-reference, and external automated-discovery baselines, our results suggest that ATLAS combines supporter-driven exploration with EvoDPO-driven stability to improve long-horizon evaluator-driven self-improvement.

Masked diffusion models (MDMs) are a potential alternative to autoregressive models (ARMs) for language generation, but generation quality depends critically on the generation order. Prior work either hard-codes an ordering (e.g., blockwise left-to-right) or learns an ordering policy for a pretrained MDM, which incurs extra cost and can yield suboptimal solutions due to the two-stage optimization. Motivated by this, we propose order-expressive masked diffusion model (OeMDM) for a broad class of diffusion generative processes with various generation orders, enabling the interpretation of MDM, ARM, and block diffusion in a single framework. Furthermore, building on OeMDM, we introduce learnable-order masked diffusion model (LoMDM), which jointly learns the generation ordering and diffusion backbone through a single objective from scratch, enabling the diffusion model to generate text in context-dependent ordering. Empirically, we confirm that LoMDM outperforms various discrete diffusion models across multiple language modeling benchmarks.

Vision tool-use reinforcement learning (RL) can equip vision language models with visual operators such as crop-and-zoom and achieves strong performance gains, yet it remains unclear whether these gains are driven by improvements in tool use or evolving intrinsic capabilities. We introduce MED (Measure--Explain--Diagnose), a coarse-to-fine framework that disentangles intrinsic capability changes from tool-induced effects, decomposes the tool-induced performance difference into gain and harm terms, and probes the mechanisms driving their evolution. Across checkpoint-level analyses in the crop-and-zoom setting on two VLMs with different tool priors and six benchmarks, we find that improvements are dominated by intrinsic learning, while tool-use RL mainly reduces tool-induced harm (e.g., fewer call-induced errors and weaker tool schema interference) and yields limited progress in tool-based correction of intrinsic failures. Overall, in the crop-and-zoom setting studied here, current vision tool-use RL learns to coexist safely with tools rather than master them.

Fine-tuning large language models (LLMs) on sensitive datasets raises privacy concerns, as training data extraction (TDE) attacks can expose highly confidential information. Existing defenses against such attacks either lack formal privacy guarantees or incur substantial utility degradation. We observe that fine-tuning induces widespread probability shifts, yet preserving only a small subset of influential token-level deviations is sufficient; the remaining shifts can be aggressively smoothed with minimal impact on utility. Motivated by this insight, we propose SCP-$Δ_r$, a Near Access Freeness (NAF)-based algorithm that operates on relative probabilities and explicitly smooths low-impact tokens using a base model. SCP-$Δ_r$ achieves orders-of-magnitude better theoretical bounds than existing NAF based methods and provides strong empirical protection against TDE attacks with minimal performance loss.

Existing multimodal large language models for long-video understanding predominantly rely on uniform sampling and single-turn inference, limiting their ability to identify sparse yet critical evidence amid extensive redundancy. We introduce Video-o3, a novel framework that supports iterative discovery of salient visual clues, fine-grained inspection of key segments, and adaptive termination once sufficient evidence is acquired. Technically, we address two core challenges in interleaved tool invocation. First, to mitigate attention dispersion induced by the heterogeneity of reasoning and tool-calling, we propose Task-Decoupled Attention Masking, which isolates per-step concentration while preserving shared global context. Second, to control context length growth in multi-turn interactions, we introduce a Verifiable Trajectory-Guided Reward that balances exploration coverage with reasoning efficiency. To support training at scale, we further develop a data synthesis pipeline and construct Seeker-173K, comprising 173K high-quality tool-interaction trajectories for effective supervised and reinforcement learning. Extensive experiments show that Video-o3 substantially outperforms state-of-the-art methods, achieving 72.1% accuracy on MLVU and 46.5% on Video-Holmes. These results demonstrate Video-o3's strong multi-hop evidence-seeking and reasoning capabilities, and validate the effectiveness of native tool invocation in long-video scenarios.

Mixture-of-Experts (MoE) layers have emerged as an important tool in scaling up modern neural networks by decoupling total trainable parameters from activated parameters in the forward pass for each token. However, sparse MoEs add complexity to training due to (i) new trainable parameters (router weights) that, like all other parameter groups, require hyperparameter (HP) tuning; (ii) new architecture scale dimensions (number of and size of experts) that must be chosen and potentially taken large. To make HP selection cheap and reliable, we propose a new parameterization for transformer models with MoE layers when scaling model width, depth, number of experts, and expert (hidden) size. Our parameterization is justified by a novel dynamical mean-field theory (DMFT) analysis. When varying different model dimensions trained at a fixed token budget, we find empirically that our parameterization enables reliable HP transfer across models from 51M to over 2B total parameters. We further take HPs identified from sweeping small models on a short token horizon to train larger models on longer horizons and report performant model behaviors.

Recent Vision-Language Models (e.g., ColPali) enable fine-grained Visual Document Retrieval (VDR) but incur prohibitive multi-vector index storage overhead. Existing training-free pruning methods either rely on heuristic layer choices or degrade sharply under aggressive compression, leading prior work to argue that effective high-compression pruning requires query-dependent training. We challenge this view with Structural Anchor Pruning (SAP), a self-calibrating, training-free, and query-agnostic index-time pruning framework with three components: (i) Score Retention (SR), a white-box per-layer compression diagnostic; (ii) SR-guided window selection, a procedure that automatically locates the structural pruning region for any backbone with no per-model hyperparameters; and (iii) a visual in-degree centrality scorer that identifies anchor patches within the selected window. On the ViDoRe v1/v2 benchmarks across three architectures spanning 18, 28, and 36 backbone layers, SAP retains over 90\% of NDCG@5 while pruning more than 90\% of visual tokens, without any per-model parameter tuning. Our layer-resolved SR analysis reveals an Alignment-Aggregation Divergence: the document's visual structure is preserved as a stable ``Structural Plateau'' within the backbone, but the final layers reshape this representation into a sparse, query-aligned form that is no longer suitable for pruning. This is the mechanistic reason SAP succeeds where final-layer methods fail.

We present UIKA, a feed-forward animatable Gaussian head model from an arbitrary number of pose-free inputs, including a single image, multi-view captures, and smartphone-captured videos. Unlike the traditional avatar method, which requires a studio-level multi-view capture system and reconstructs a human-specific model through a long-time optimization process, we rethink the task through the lenses of model representation, network design, and data preparation. First, we introduce a UV-guided avatar modeling strategy, in which each input image is associated with a pixel-wise facial correspondence estimation. Such correspondence estimation allows us to reproject each valid pixel color from screen space to UV space, which is independent of camera pose and character expression. Furthermore, we design learnable UV tokens on which the attention mechanism can be applied at both the screen and UV levels. The learned UV tokens can be decoded into canonical Gaussian attributes using aggregated UV information from all input views. To train our large avatar model, we additionally prepare a large-scale, identity-rich synthetic training dataset. Our method significantly outperforms existing approaches in both monocular and multi-view settings.

Reinforcement learning with verifiable rewards (RLVR) has driven significant performance gains in reasoning-oriented large language models (LLMs), yet its internal training dynamics remain largely a black box. In this work, we perform a comprehensive trajectory-level analysis of RLVR and uncover a striking regularity: across various model families, RL algorithms, and training configurations, RLVR consistently enters a robust linear regime, where both parameter weights and output log-probabilities, measured rigorously via teacher-forced evaluation, evolve in a highly linear manner ($R^2 > 0.7$). Through controlled experiments and theoretical analysis, we demonstrate that this linearity is not a coincidence, but stems from the high-variance, noisy nature of RLVR training signals, which act as a low-pass filter to concentrate optimization along a stable, low-dimensional drift. Moreover, we show that this linear structure is not merely descriptive but powerfully predictive and actionable. Specifically, weight-space extrapolation matches the performance of standard RL optimization while achieving a 6.1x training speedup through periodic re-grounding. Meanwhile, output-space extrapolation serves as a lightweight intervention that effectively bypasses late-stage model collapse, consistently outperforming standard RL across mathematical and coding benchmarks, with an average performance improvement of 4.2%. Our code is available at https://github.com/Miaow-Lab/RLVR-Linearity.