Multimodal Sentiment Analysis (MSA) seeks to infer human emotions by integrating textual, acoustic, and visual cues. However, existing approaches often rely on all modalities are completeness, whereas real-world applications frequently encounter noise, hardware failures, or privacy restrictions that result in missing modalities. There exists a significant feature misalignment between incomplete and complete modalities, and directly fusing them may even distort the well-learned representations of the intact modalities. To this end, we propose PRLF, a Progressive Representation Learning Framework designed for MSA under uncertain missing-modality conditions. PRLF introduces an Adaptive Modality Reliability Estimator (AMRE), which dynamically quantifies the reliability of each modality using recognition confidence and Fisher information to determine the dominant modality. In addition, the Progressive Interaction (ProgInteract) module iteratively aligns the other modalities with the dominant one, thereby enhancing cross-modal consistency while suppressing noise. Extensive experiments on CMU-MOSI, CMU-MOSEI, and SIMS verify that PRLF outperforms state-of-the-art methods across both inter- and intra-modality missing scenarios, demonstrating its robustness and generalization capability.

Recent advances in time-series forecasting increasingly rely on pre-trained foundation-style models. While these models often claim broad generalization, existing evaluation protocols provide limited evidence. Indeed, most current benchmarks use static train-test splits that can easily lead to contamination as foundation models can inadvertently train on test data or perform model selection using test scores, which can inflate performance. We introduce Impermanent, a live benchmark that evaluates forecasting models under open-world temporal change by scoring forecasts sequentially over time on continuously updated data streams, enabling the study of temporal robustness, distributional shift, and performance stability rather than one-off accuracy on a frozen test set. Impermanent is instantiated on GitHub open-source activity, providing a naturally live and highly non-stationary dataset shaped by releases, shifting contributor behavior, platform/tooling changes, and external events. We focus on the top 400 repositories by star count and construct time series from issues opened, pull requests opened, push events, and new stargazers, evaluated over a rolling window with daily updates, alongside standardized protocols and leaderboards for reproducible, ongoing comparison. By shifting evaluation from static accuracy to sustained performance, Impermanent takes a concrete step toward assessing when and whether foundation-level generalization in time-series forecasting can be meaningfully claimed. Code and a live dashboard are available at https://github.com/TimeCopilot/impermanent and https://impermanent.timecopilot.dev.

Text-to-motion generation has advanced rapidly, yet two challenges persist. First, existing motion autoencoders compress each frame into a single monolithic latent vector, entangling trajectory and per-joint rotations in an unstructured representation that downstream generators struggle to model faithfully. Second, text-to-motion, pose-conditioned generation, and long-horizon sequential synthesis typically require separate models or task-specific mechanisms, with autoregressive approaches suffering from severe error accumulation over extended rollouts. We present PRISM, addressing each challenge with a dedicated contribution. (1) A joint-factorized motion latent space: each body joint occupies its own token, forming a structured 2D grid (time joints) compressed by a causal VAE with forward-kinematics supervision. This simple change to the latent space -- without modifying the generator -- substantially improves generation quality, revealing that latent space design has been an underestimated bottleneck. (2) Noise-free condition injection: each latent token carries its own timestep embedding, allowing conditioning frames to be injected as clean tokens (timestep0) while the remaining tokens are denoised. This unifies text-to-motion and pose-conditioned generation in a single model, and directly enables autoregressive segment chaining for streaming synthesis. Self-forcing training further suppresses drift in long rollouts. With these two components, we train a single motion generation foundation model that seamlessly handles text-to-motion, pose-conditioned generation, autoregressive sequential generation, and narrative motion composition, achieving state-of-the-art on HumanML3D, MotionHub, BABEL, and a 50-scenario user study.

Although deep neural networks have facilitated significant progress of neural vocoders in recent years, they usually suffer from intrinsic challenges like opaque modeling, inflexible retraining under different input configurations, and parameter-performance trade-off. These inherent hurdles can heavily impede the development of this field. To resolve these problems, in this paper, we propose a novel neural vocoder in the time-frequency (T-F) domain. Specifically, we bridge the connection between the classical range-null decomposition (RND) theory and the vocoder task, where the reconstruction of the target spectrogram is formulated into the superimposition between range-space and null-space. The former aims to project the representation in the original mel-domain into the target linear-scale domain, and the latter can be instantiated via neural networks to further infill the spectral details. To fully leverage the spectrum prior, an elaborate dual-path framework is devised, where the spectrum is hierarchically encoded and decoded, and the cross- and narrow-band modules are leveraged for effectively modeling along sub-band and time dimensions. To enable inference under various configurations, we propose a simple yet effective strategy, which transforms the multi-condition adaption in the inference stage into the data augmentation in the training stage. Comprehensive experiments are conducted on various benchmarks. Quantitative and qualitative results show that while enjoying lightweight network structure and scalable inference paradigm, the proposed framework achieves state-ofthe-art performance among existing advanced methods. Code is available at https://github.com/Andong-Li-speech/RNDVoC.

The surge of highly realistic synthetic videos produced by contemporary generative systems has significantly increased the risk of malicious use, challenging both humans and existing detectors. Against this backdrop, we take a generator-side view and observe that internal cross-attention mechanisms in these models encode fine-grained speech-motion alignment, offering useful correspondence cues for forgery detection. Building on this insight, we propose X-AVDT, a robust and generalizable deepfake detector that probes generator-internal audio-visual signals accessed via DDIM inversion to expose these cues. X-AVDT extracts two complementary signals: (i) a video composite capturing inversion-induced discrepancies, and (ii) an audio-visual cross-attention feature reflecting modality alignment enforced during generation. To enable faithful cross-generator evaluation, we further introduce MMDF, a new multimodal deepfake dataset spanning diverse manipulation types and rapidly evolving synthesis paradigms, including GANs, diffusion, and flow-matching. Extensive experiments demonstrate that X-AVDT achieves leading performance on MMDF and generalizes strongly to external benchmarks and unseen generators, outperforming existing methods with accuracy improved by 13.1%. Our findings highlight the importance of leveraging internal audio-visual consistency cues for robustness to future generators in deepfake detection.

While Physics-Informed Neural Networks (PINNs) offer a mesh-free approach to solving PDEs, standard point-wise residual minimization suffers from convergence pathologies in topologically complex domains like Triply Periodic Minimal Surfaces (TPMS). The locality bias of point-wise constraints fails to propagate global information through tortuous channels, causing unstable gradients and conservation violations. To address this, we propose the Multi-scale Weak-form PINN (MUSA-PINN), which reformulates PDE constraints as integral conservation laws over hierarchical spherical control volumes. We enforce continuity and momentum conservation via flux-balance residuals on control surfaces. Our method utilizes a three-scale subdomain strategy-comprising large volumes for long-range coupling, skeleton-aware meso-scale volumes aligned with transport pathways, and small volumes for local refinement-alongside a two-stage training schedule prioritizing continuity. Experiments on steady incompressible flow in TPMS geometries show MUSA-PINN outperforms state-of-the-art baselines, reducing relative errors by up to 93% and preserving mass conservation.

Recent progress in 3D hand--object interaction (HOI) generation has primarily focused on single--hand grasp synthesis, while bimanual manipulation remains significantly more challenging. Long--horizon planning instability, fine--grained joint articulation, and complex cross--hand coordination make coherent bimanual generation difficult, especially under multimodal conditions. Existing approaches often struggle to simultaneously ensure temporal consistency, physical plausibility, and semantic alignment over extended sequences. We propose StructBiHOI, a Structured articulation modeling framework for long-horizon Bimanual HOI generation. Our key insight is to structurally disentangle temporal joint planning from frame--level manipulation refinement. Specifically, a jointVAE models long-term joint evolution conditioned on object geometry and task semantics, while a maniVAE refines fine-grained hand poses at the single--frame level. To enable stable and efficient long--sequence generation, we incorporate a state--space--inspired diffusion denoiser based on Mamba, which models long--range dependencies with linear complexity. This hierarchical design facilitates coherent dual-hand coordination and articulated object interaction. Extensive experiments on bimanual manipulation and single-hand grasping benchmarks demonstrate that our method achieves superior long--horizon stability, motion realism, and computational efficiency compared to strong baselines.

Federated Learning (FL) suffers significant performance degradation in real-world deployments characterized by moderate to extreme statistical heterogeneity (non-IID client data). While global aggregation strategies promote broad generalization, they often fail to capture the diversity of local data distributions, leading to suboptimal personalization. We address this problem with FedPrism, a framework that uses two main strategies. First, it uses a Prism Decomposition method that builds each client's model from three parts: a global foundation, a shared group part for similar clients, and a private part for unique local data. This allows the system to group similar users together automatically and adapt if their data changes. Second, we include a Dual-Stream design that runs a general model alongside a local specialist. The system routes predictions between the general model and the local specialist based on the specialist's confidence. Through systematic experiments on non-IID data partitions, we demonstrate that FedPrism exceeds static aggregation and hard-clustering baselines, achieving significant accuracy gains under high heterogeneity. These results establish FedPrism as a robust and flexible solution for federated learning in heterogeneous environments, effectively balancing generalizable knowledge with adaptive personalization.

Hundreds of millions of farmers make high-stakes decisions under uncertainty about future weather. Forecasts can inform these decisions, but available choices and their risks and benefits vary between farmers. We introduce a decision-theory framework for designing useful forecasts in settings where the forecaster cannot prescribe optimal actions because farmers' circumstances are heterogeneous. We apply this framework to the case of seasonal onset of monsoon rains, a key date for planting decisions and agricultural investments in many tropical countries. We develop a system for tailoring forecasts to the requirements of this framework by blending systematically benchmarked artificial intelligence (AI) weather prediction models with a new "evolving farmer expectations" statistical model. This statistical model applies Bayesian inference to historical observations to predict time-varying probabilities of first-occurrence events throughout a season. The blended system yields more skillful Indian monsoon forecasts at longer lead times than its components or any multi-model average. In 2025, this system was deployed operationally in a government-led program that delivered subseasonal monsoon onset forecasts to 38 million Indian farmers, skillfully predicting that year's early-summer anomalous dry period. This decision-theory framework and blending system offer a pathway for developing climate adaptation tools for large vulnerable populations around the world.

Table reasoning requires models to jointly perform semantic understanding and precise numerical operations. Most existing methods rely on a single-turn reasoning paradigm over tables which suffers from context overflow and weak numerical sensitivity. To address these limitations, we previously proposed TableMind as a tuning-based autonomous programmatic agent that simulates human-like interaction within a lightweight large language model (LLM). TableMind internalizes planning, action, and reflection through a two-stage training strategy involving supervised fine-tuning (SFT) on filtered high-quality data and reinforcement learning (RL) via a multi-perspective reward and the Rank-Aware Policy Optimization (RAPO) algorithm. While TableMind establishes a solid foundation for programmatic agents, the inherent stochasticity of LLMs remains a critical challenge that leads to hallucinations. In this paper, we extend this foundation to TableMind++ by introducing a novel uncertainty-aware inference framework to mitigate hallucinations. Specifically, we propose memory-guided plan pruning to retrieve historical trajectories for validating and filtering out logically flawed plans to address epistemic uncertainty. To ensure execution precision, we introduce confidence-based action refinement which monitors token-level probabilities to detect and self-correct syntactic noise for aleatoric uncertainty mitigation. Finally, we employ dual-weighted trajectory aggregation to synthesize a robust consensus from multiple reasoning paths. Extensive experiments on diverse benchmarks demonstrate that TableMind++ consistently outperforms previous baselines and proprietary models to validate the effectiveness of integrating autonomous training with uncertainty quantification. Our code is available.

Transit agencies that operate on-demand transportation services have to respond to trip requests from passengers in real time, which involves solving dynamic vehicle routing problems with pick-up and drop-off constraints. Based on discussions with public transit agencies, we observe a real-world problem that is not addressed by prior work: when trips are booked in advance (e.g., trip requests arrive a few hours in advance of their requested pick-up times), the agency needs to promptly confirm whether a request can be accepted or not, and ensure that accepted requests are served as promised. State-of-the-art computational approaches either provide prompt confirmation but lack the ability to continually optimize and improve routes for accepted requests, or they provide continual optimization but cannot guarantee serving all accepted requests. To address this gap, we introduce a novel problem formulation of dynamic vehicle routing with prompt confirmation and continual optimization. We propose a novel computational approach for this vehicle routing problem, which integrates a quick insertion search for prompt confirmation with an anytime algorithm for continual optimization. To maximize the number requests served, we train a non-myopic objective function using reinforcement learning, which guides both the insertion and the anytime algorithms towards optimal, non-myopic solutions. We evaluate our computational approach on a real-world microtransit dataset from a public transit agency in the U.S., demonstrating that our proposed approach provides prompt confirmation while significantly increasing the number of requests served compared to existing approaches.

Generative models have emerged as powerful priors for solving inverse problems. These models typically represent a class of natural signals using a single fixed complexity or dimensionality. This can be limiting: depending on the problem, a fixed complexity may result in high representation error if too small, or overfitting to noise if too large. We develop tunable-complexity priors for diffusion models, normalizing flows, and variational autoencoders, leveraging nested dropout. Across tasks including compressed sensing, inpainting, denoising, and phase retrieval, we show empirically that tunable priors consistently achieve lower reconstruction errors than fixed-complexity baselines. In the linear denoising setting, we provide a theoretical analysis that explicitly characterizes how the optimal tuning parameter depends on noise and model structure. This work demonstrates the potential of tunable-complexity generative priors and motivates both the development of supporting theory and their application across a wide range of inverse problems.

The rapid adoption of transformer-based models in computational pathology has enabled prediction of molecular and clinical biomarkers from H&E whole-slide images, yet interpretability has not kept pace with model complexity. While attribution- and generative-based methods are common, feature visualization approaches such as class visualizations (CVs) and activation atlases (AAs) have not been systematically evaluated for these models. We developed a visualization framework and assessed CVs and AAs for a transformer-based foundation model across tissue and multi-organ cancer classification tasks with increasing label granularity. Four pathologists annotated real and generated images to quantify inter-observer agreement, complemented by attribution and similarity metrics. CVs preserved recognizability for morphologically distinct tissues but showed reduced separability for overlapping cancer subclasses. In tissue classification, agreement decreased from Fleiss k = 0.75 (scans) to k = 0.31 (CVs), with similar trends in cancer subclass tasks. AAs revealed layer-dependent organization: coarse tissue-level concepts formed coherent regions, whereas finer subclasses exhibited dispersion and overlap. Agreement was moderate for tissue classification (k = 0.58), high for coarse cancer groupings (k = 0.82), and low at subclass level (k = 0.11). Atlas separability closely tracked expert agreement on real images, indicating that representational ambiguity reflects intrinsic pathological complexity. Attribution-based metrics approximated expert variability in low-complexity settings, whereas perceptual and distributional metrics showed limited alignment. Overall, concept-level feature visualization reveals structured morphological manifolds in transformer-based pathology models and provides a framework for expert-centered interrogation of learned representations across label granularities.

Recent Vision-Language Models (VLMs) have made remarkable progress in multimodal understanding tasks, yet their evaluation on long video understanding remains unreliable. Due to limited frame inputs, key frames necessary for answering the question may be missing from the model's input. However, models that truthfully refuse to answer under such uncertainty are marked as incorrect, while those that guess may coincidentally produce the correct answer and thus obtain deceptively higher accuracy, leading to misleading evaluation results and encouraging models to guess rather than respond honestly. To address this issue, we introduce VirtueBench, a benchmark explicitly designed to assess model trustworthiness under uncertainty. VirtueBench constructs multiple frame-sampling levels for each video and provides ground truths that distinguish between answerable and unanswerable cases. Evaluations on 25 open-source and commercial VLMs reveal distinct refusal behaviors across different model families, with refusal accuracy ranging from over 70% in the best models to nearly 0% in the worst. Moreover, most models exhibit a substantial drop in refusal when the prompt does not explicitly require them to do so. These findings highlight the need for developing trustworthy VLMs for multimodal understanding, guided by benchmarks and leaderboards that emphasize reliability and trustworthiness.

Alzheimer disease (AD) diagnosis and prognosis increasingly rely on machine learning (ML) models. Although these models provide good results, clinical adoption is limited by the need for technical expertise and the lack of trustworthy and consistent model explanations. SHAP (SHapley Additive exPlanations) is com-monly used to interpret AD models, but existing studies tend to focus on explanations for isolated tasks, providing little evidence about their robustness across disease stages, model architectures, or prediction objectives. This paper proposes a multi-level explainability framework that measures the coherence, stabil-ity and consistency of explanations by integrating: (1) within-model coherence metrics between feature importance and SHAP, (2) SHAP stability across AD boundaries, and (3) SHAP cross-task consistency be-tween diagnosis and prognosis. Using AutoML to optimize classifiers on the NACC dataset, we trained four diagnostic and four prognostic models covering the standard AD progression stages. Stability was then evaluated using correlation metrics, top-k feature overlap, SHAP sign consistency, and domain-level contribution ratios. Results show that cognitive and functional markers dominate SHAP explanations in both diagnosis and prognosis. SHAP-SHAP consistency between diagnostic and prognostic models was high across all classifiers, with 100% sign stability and minimal shifts in explanatory magnitude. Domain-level contributions also remained stable, with only minimal increases in genetic features for prognosis. These results demonstrate that SHAP explanations can be quantitatively vali-dated for robustness and transferability, providing clinicians with more reliable interpretations of ML pre-dictions.

Protein sequence design must balance designability, defined as the ability to recover a target backbone, with multiple, often competing, developability properties such as solubility, thermostability, and expression. Existing approaches address these properties through post hoc mutation, inference-time biasing, or retraining on property-specific subsets, yet they are target dependent and demand substantial domain expertise or careful hyperparameter tuning. In this paper, we introduce ProtAlign, a multi-objective preference alignment framework that fine-tunes pretrained inverse folding models to satisfy diverse developability objectives while preserving structural fidelity. ProtAlign employs a semi-online Direct Preference Optimization strategy with a flexible preference margin to mitigate conflicts among competing objectives and constructs preference pairs using in silico property predictors. Applied to the widely used ProteinMPNN backbone, the resulting model MoMPNN enhances developability without compromising designability across tasks including sequence design for CATH 4.3 crystal structures, de novo generated backbones, and real-world binder design scenarios, making it an appealing framework for practical protein sequence design.

Knowledge distillation between asymmetric architectures often induces severe geometric constraints on the learned representation space. In this work, we investigate the Dimensional Collapse phenomenon when distilling global Vision Transformers (CLIP and DINOv2) into capacity-constrained CNNs. By employing strictly centered SVD and Effective Rank, we first demonstrate a capacity-agnostic phase transition on CIFAR-10 where standard cosine distillation collapses representations to an intrinsic Effective Rank of ~16. To reverse this, we integrate an auxiliary contrastive objective (InfoNCE), expanding the student's manifold by 2.4x (to ~38 effective dimensions). We further demonstrate that while DINOv2's uniform geometry partially prevents collapse, contrastive expansion remains a universal requirement to reach the CNN's topological capacity limit (~82 dimensions). Finally, we reveal a critical capacity-density trade-off: overparameterization within fixed manifolds induces brittleness, while capacity-constrained models act as optimal low-pass semantic filters, successfully recovering inherent noise immunity.

Human gameplay is a visually grounded interaction loop in which players act, reflect on failures, and watch tutorials to refine strategies. Can Vision-Language Models (VLMs) also learn from video-based reflection? We present GameVerse, a comprehensive video game benchmark that enables a reflective visual interaction loop. Moving beyond traditional fire-and-forget evaluations, it uses a novel reflect-and-retry paradigm to assess how VLMs internalize visual experience and improve policies. To facilitate systematic and scalable evaluation, we also introduce a cognitive hierarchical taxonomy spanning 15 globally popular games, dual action space for both semantic and GUI control, and milestone evaluation using advanced VLMs to quantify progress. Our experiments show that VLMs benefit from video-based reflection in varied settings, and perform best by combining failure trajectories and expert tutorials-a training-free analogue to reinforcement learning (RL) plus supervised fine-tuning (SFT).Our project page is available at https://gameverse-bench.github.io/ . Our code is available at https://github.com/THUSI-Lab/GameVerse .

Many scientific problems in the context of machine learning can be reduced to the search of polynomial answers in appropriate variables. The Hevisidization of arbitrary polynomial is actually provided by one-and-the same two-layer expression. What prevents the use of this simple idea is the fatal degeneracy of the Heaviside and sigmoid expansions, which traps the steepest-descent evolution at the bottom of canyons, close to the starting point, but far from the desired true minimum. This problem is unavoidable and can be formulated as a peculiar uncertainty principle -- the sharper the minimum, the smoother the canyons. It is a direct analogue of the usual one, which is the pertinent property of the more familiar Fourier expansion. Standard machine learning software fights with this problem empirically, for example, by testing evolutions, originated at randomly distributed starting points and then selecting the best one. Surprisingly or not, phenomena and problems, encountered in ML application to science are pure scientific and belong to physics, not to computer science. On the other hand, they sound slightly different and shed new light on the well-known phenomena -- for example, extend the uncertainty principle from Fourier and, later, wavelet analysis to a new peculiar class of nearly singular sigmoid functions.

As datasets continue to grow in size and complexity, finding succinct yet accurate data summaries poses a key challenge. Centroid-based clustering, a widely adopted approach to address this challenge, finds informative summaries of datasets in terms of few prototypes, each representing a cluster in the data. Despite their wide adoption, the resulting data summaries often contain redundancies, limiting their effectiveness particularly in datasets characterized by a large number of underlying clusters. To overcome this limitation, we introduce the Khatri-Rao clustering paradigm that extends traditional centroid-based clustering to produce more succinct but equally accurate data summaries by postulating that centroids arise from the interaction of two or more succinct sets of protocentroids. We study two central approaches to centroid-based clustering, namely the well-established k-Means algorithm and the increasingly popular topic of deep clustering, under the lens of the Khatri-Rao paradigm. To this end, we introduce the Khatri-Rao k-Means algorithm and the Khatri-Rao deep clustering framework. Extensive experiments show that Khatri-Rao k-Means can strike a more favorable trade-off between succinctness and accuracy in data summarization than standard k-Means. Leveraging representation learning, the Khatri-Rao deep clustering framework offers even greater benefits, reducing even more the size of data summaries given by deep clustering while preserving their accuracy.

Current methods for textual analysis rely on data annotated within predefined ontologies, often embedding human bias within black-box models. Despite achieving near-perfect performance, these approaches exploit unstructured, linear pattern recognition rather than modeling the structured interactions between entities that naturally emerge in discourse. In this work, we propose a graph-based framework for the detection, analysis, and classification of oppositional narratives and their underlying entities by representing narratives as entity-interaction graphs. Moreover, by incorporating causal estimation at the node level, our approach derives a causal representation of each contribution to the final classification by distilling the constructed sentence graph into a minimal causal subgraph. Building upon this representation, we introduce a classification pipeline that outperforms existing approaches to oppositional thinking classification task.

Memory-efficient optimization methods have recently gained increasing attention for scaling full-parameter training of large language models under the GPU-memory bottleneck. Existing approaches either lack clear convergence guarantees, or only achieve the standard ${\mathcal{O}}(ε^{-4})$ iteration complexity in the nonconvex settings. We propose Omni-Masked Gradient Descent (OMGD), an optimization method based on mask traversal for memory efficient training, and provide a nonconvex convergence analysis that establishes a strictly improved iteration complexity of $\tilde{\mathcal{O}}(ε^{-3})$ for finding an $ε$-approximate stationary point. Empirically, OMGD is a lightweight, plug-and-play approach that integrates seamlessly into most mainstream optimizers, yielding consistent improvements over competitive baselines in both fine-tuning and pre-training tasks.

Large language models sometimes produce false or misleading responses. Two approaches to this problem are honesty elicitation -- modifying prompts or weights so that the model answers truthfully -- and lie detection -- classifying whether a given response is false. Prior work evaluates such methods on models specifically trained to lie or conceal information, but these artificial constructions may not resemble naturally-occurring dishonesty. We instead study open-weights LLMs from Chinese developers, which are trained to censor politically sensitive topics: Qwen3 models frequently produce falsehoods about subjects like Falun Gong or the Tiananmen protests while occasionally answering correctly, indicating they possess knowledge they are trained to suppress. Using this as a testbed, we evaluate a suite of elicitation and lie detection techniques. For honesty elicitation, sampling without a chat template, few-shot prompting, and fine-tuning on generic honesty data most reliably increase truthful responses. For lie detection, prompting the censored model to classify its own responses performs near an uncensored-model upper bound, and linear probes trained on unrelated data offer a cheaper alternative. The strongest honesty elicitation techniques also transfer to frontier open-weights models including DeepSeek R1. Notably, no technique fully eliminates false responses. We release all prompts, code, and transcripts.

Mechanistic interpretability typically relies on post-hoc analysis of trained networks. We instead adopt an interventional approach: testing hypotheses a priori by modifying architectural topology to observe training dynamics. We study grokking - delayed generalization in Transformers trained on cyclic modular addition (Zp) - investigating if specific architectural degrees of freedom prolong the memorization phase. We identify two independent structural factors in standard Transformers: unbounded representational magnitude and data-dependent attention routing. First, we introduce a fully bounded spherical topology enforcing L2 normalization throughout the residual stream and an unembedding matrix with a fixed temperature scale. This removes magnitude-based degrees of freedom, reducing grokking onset time by over 20x without weight decay. Second, a Uniform Attention Ablation overrides data-dependent query-key routing with a uniform distribution, reducing the attention layer to a Continuous Bag-of-Words (CBOW) aggregator. Despite removing adaptive routing, these models achieve 100% generalization across all seeds and bypass the grokking delay entirely. To evaluate whether this acceleration is a task-specific geometric alignment rather than a generic optimization stabilizer, we use non-commutative S5 permutation composition as a negative control. Enforcing spherical constraints on S5 does not accelerate generalization. This suggests eliminating the memorization phase depends strongly on aligning architectural priors with the task's intrinsic symmetries. Together, these findings provide interventional evidence that architectural degrees of freedom substantially influence grokking, suggesting a predictive structural perspective on training dynamics.

Disruptions at critical logistics nodes pose severe risks to global supply chains, yet existing risk prediction systems typically prioritize forecasting accuracy without providing operationally interpretable early warnings. This paper proposes an evidence-grounded framework that jointly performs supply chain bottleneck prediction and faithful natural-language risk explanation by coupling a Temporal Graph Attention Network (TGAT) with a structured large language model (LLM) reasoning module. Using maritime hubs as a representative case study for global supply chain nodes, daily spatial graphs are constructed from Automatic Identification System (AIS) broadcasts, where inter-node interactions are modeled through attention-based message passing. The TGAT predictor captures spatiotemporal risk dynamics, while model-internal evidence -- including feature z-scores and attention-derived neighbor influence -- is transformed into structured prompts that constrain LLM reasoning to verifiable model outputs. To evaluate explanatory reliability, we introduce a directional-consistency validation protocol that quantitatively measures agreement between generated risk narratives and underlying statistical evidence. Experiments on six months of real-world logistics data demonstrate that the proposed framework outperforms baseline models, achieving a test AUC of 0.761, AP of 0.344, and recall of 0.504 under a strict chronological split while producing early warning explanations with 99.6\% directional consistency. Results show that grounding LLM generation in graph-model evidence enables interpretable and auditable risk reporting without sacrificing predictive performance. The framework provides a practical pathway toward operationally deployable explainable AI for supply chain risk early warning and resilience management.

Transformer-based encoder-decoder networks have recently achieved impressive results in handwritten text recognition, partly thanks to their auto-regressive decoder which implicitly learns a language model. However, such networks suffer from a large performance drop when evaluated on a target corpus whose language distribution is shifted from the source text seen during training. To retain recognition accuracy despite this language shift, we propose an external n-gram injection (NGI) for dynamic adaptation of the network's language modeling at inference time. Our method allows switching to an n-gram language model estimated on a corpus close to the target distribution, therefore mitigating bias without any extra training on target image-text pairs. We opt for an early injection of the n-gram into the transformer decoder so that the network learns to fully leverage text-only data at the low additional cost of n-gram inference. Experiments on three handwritten datasets demonstrate that the proposed NGI significantly reduces the performance gap between source and target corpora.

Test-time scaling has shown that allocating more additional computation at inference can improve generation quality, motivating a natural follow-up question: where should this computation be spent? Building on this insight, we introduce PonderLM-3, a pretraining framework for token-wise adaptive pondering that learns to selectively allocate additional computation under purely self-supervised objectives, built on top of the PonderLM-2 backbone. This makes additional inference computation an allocatable per-token resource, so tokens receive more computation only when it is beneficial, rather than paying a uniform extra cost. To make this allocation learnable while maintaining train-inference consistency, PonderLM-3 injects a differentiable attention mask during pretraining and pairs it with a matching hard pruning rule at inference. PonderLM-3 defines a stronger Pareto frontier: compared with existing recursive or adaptive baselines, it achieves lower pretraining perplexity at equal inference FLOPs. On downstream benchmarks, PonderLM-3 attains comparable performance to fixed-step PonderLM-2 under the same maximum number of additional computation steps, while using fewer inference FLOPs in practice. Overall, PonderLM-3 provides an end-to-end differentiable and train-inference consistent framework for token-wise adaptive computation, enabling additional inference compute to be allocated where it is most useful rather than paid uniformly by every token.

We present DOCFORGE-BENCH, the first unified zero-shot benchmark for document forgery detection, evaluating 14 methods across eight datasets spanning text tampering, receipt forgery, and identity document manipulation. Unlike fine-tuning-oriented evaluations such as ForensicHub [Du et al., 2025], DOCFORGE-BENCH applies all methods with their published pretrained weights and no domain adaptation -- a deliberate design choice that reflects the realistic deployment scenario where practitioners lack labeled document training data. Our central finding is a pervasive calibration failure invisible under single-threshold protocols: methods achieve moderate Pixel-AUC (>=0.76) yet near-zero Pixel-F1. This AUC-F1 gap is not a discrimination failure but a score-distribution shift: tampered regions occupy only 0.27-4.17% of pixels in document images -- an order of magnitude less than in natural image benchmarks -- making the standard tau=0.5 threshold catastrophically miscalibrated. Oracle-F1 is 2-10x higher than fixed-threshold Pixel-F1, confirming that calibration, not representation, is the bottleneck. A controlled calibration experiment validates this: adapting a single threshold on N=10 domain images recovers 39-55% of the Oracle-F1 gap, demonstrating that threshold adaptation -- not retraining -- is the key missing step for practical deployment. Overall, no evaluated method works reliably out-of-the-box on diverse document types, underscoring that document forgery detection remains an unsolved problem. We further note that all eight datasets predate the era of generative AI editing; benchmarks covering diffusion- and LLM-based document forgeries represent a critical open gap on the modern attack surface.

Masked diffusion models (MDMs) generate text by iteratively selecting positions to unmask and then predicting tokens at those positions. Yet MDMs lack proper likelihood evaluation: the evidence lower bound (ELBO) is not only a loose bound on log-likelihood, but, as we show, is also computed under the training distribution rather than the test-time distribution. We resolve this within our DUEL framework, which unifies leading MDM sampling strategies that employ $\textit{deterministic}$ position selection. We prove that DUEL samplers admit $\textbf{exact likelihood computation under the test-time distribution}$ -- giving MDMs $\textit{proper}$ likelihood, and hence proper perplexity, for the first time. This proper perplexity is the natural analogue of autoregressive perplexity and lets us revisit key questions about MDMs. $\textbf{MDMs are substantially better than previously thought}$: the MDM-autoregressive perplexity gap shrinks by up to $32\%$ on in-domain data and $82\%$ on zero-shot benchmarks. DUEL enables the first principled comparison of fast,parallel samplers across compute budgets -- an analysis impossible with the ELBO and unreliable with generative perplexity -- identifying a strong default method. Finally, oracle search over position orderings reveals MDMs can far surpass autoregressive models -- achieving $36.47$ vs. $52.11$ perplexity on AG News -- demonstrating the ceiling of MDM performance has not yet been reached.

Real-world tool-using agents operate over long-horizon workflows with recurring structure and diverse demands, where effective behavior requires not only invoking atomic tools but also abstracting, and reusing higher-level tool compositions. However, existing benchmarks mainly measure instance-level success under static tool sets, offering limited insight into agents' ability to acquire such reusable skills. We address this gap by introducing SkillCraft, a benchmark explicitly stress-test agent ability to form and reuse higher-level tool compositions, where we call Skills. SkillCraft features realistic, highly compositional tool-use scenarios with difficulty scaled along both quantitative and structural dimensions, designed to elicit skill abstraction and cross-task reuse. We further propose a lightweight evaluation protocol that enables agents to auto-compose atomic tools into executable Skills, cache and reuse them inside and across tasks, thereby improving efficiency while accumulating a persistent library of reusable skills. Evaluating state-of-the-art agents on SkillCraft, we observe substantial efficiency gains, with token usage reduced by up to 80% by skill saving and reuse. Moreover, success rate strongly correlates with tool composition ability at test time, underscoring compositional skill acquisition as a core capability.