We propose a diffusion-based framework for prompt optimization that leverages Diffusion Language Models (DLMs) to iteratively refine system prompts through masked denoising. By conditioning on interaction traces, including user queries, model responses, and optional feedback, our method enables flexible, span-level prompt updates without requiring gradient access or modifying the downstream language model. Across diverse benchmarks (e.g., $τ$-bench, SST-2, SST-5), DLM-optimized prompts consistently improve the performance of a frozen target LLM (e.g., GPT-4o-mini). We further show that moderate diffusion step counts provide the best balance between refinement quality and stability. These results highlight diffusion-based prompt optimization as a general, model-agnostic, and scalable approach for enhancing LLM performance through iterative prompt refinement.

Chain-of-Thought reasoning significantly improves the performance of large language models on complex tasks, but incurs high inference latency due to long generation traces. Step-level speculative reasoning aims to mitigate this cost, yet existing approaches face a long-standing trade-off among accuracy, inference speed, and resource efficiency. We propose ConfSpec, a confidence-gated cascaded verification framework that resolves this trade-off. Our key insight is an asymmetry between generation and verification: while generating a correct reasoning step requires substantial model capacity, step-level verification is a constrained discriminative task for which small draft models are well-calibrated within their competence range, enabling high-confidence draft decisions to be accepted directly while selectively escalating uncertain cases to the large target model. Evaluation across diverse workloads shows that ConfSpec achieves up to 2.24$\times$ end-to-end speedups while matching target-model accuracy. Our method requires no external judge models and is orthogonal to token-level speculative decoding, enabling further multiplicative acceleration.

Users increasingly rely on Large Language Models (LLMs) for Deep Research, using them to synthesize diverse sources into structured reports that support understanding and action. In this context, the practical reliability of such reports hinges on logical quality: whether the report's claims and arguments are explicitly supported and can be trusted as a basis for downstream use, rather than merely appearing fluent or informative. However, current evaluation frameworks largely overlook this requirement. To bridge this gap, we introduce ReportLogic, a benchmark that quantifies report-level logical quality through a reader-centric lens of auditability. Specifically, ReportLogic adopts a hierarchical taxonomy that evaluates whether readers can (1) trace an on-topic report structure with a unified analytical arc (Macro-Logic), (2) understand the progression with necessary context (Expositional-Logic), and (3) verify conclusions via explicit claim--support (Structural-Logic). Based on this taxonomy, we construct a human-annotated rubric-guided dataset and train an open-source LogicJudge for scalable evaluation. We further evaluate judge robustness via adversarial attacks, showing that off-the-shelf LLM judges are frequently influenced by superficial cues (e.g., verbosity), and reasoning modes can mask broken support relations. Overall, our results provide actionable guidance for building more robust logic evaluators and improving the logical reliability of LLM-generated reports.

Zeroth-order (ZO) optimization provides a gradient-free alternative to first-order (FO) methods by estimating gradients via finite differences of function evaluations, and has recently emerged as a memory-efficient paradigm for fine-tuning large-scale models by avoiding backpropagation. However, ZO optimization has a fundamental tension between accuracy and query efficiency. In this work, we show that ZO optimization can be substantially improved by unifying two complementary principles: (i) a projection-based subspace view that reduces gradient estimation variance by exploiting the intrinsic low-rank structure of model updates, and (ii) Muon-style spectral optimization that applies gradient orthogonalization to extract informative spectral structure from noisy ZO gradients. These findings form a unified framework of subspace gradient orthogonalization, which we instantiate in a new method, ZO-Muon, admitting a natural interpretation as a low-rank Muon optimizer in the ZO setting. Extensive experiments on large language models (LLMs) and vision transformers (ViTs) demonstrate that ZO-Muon significantly accelerates convergence and achieves a win-win improvement in accuracy and query/runtime efficiency. Notably, compared to the popular MeZO baseline, ZO-Muon requires only 24.7% of the queries to reach the same SST-2 performance for LLM fine-tuning, and improves accuracy by 25.1% on ViT-B fine-tuning on CIFAR-100.

Max pooling is the de facto standard for converting anomaly score maps into image-level decisions in memory-bank-based unsupervised anomaly detection (UAD). However, because it relies on a single extreme response, it discards most information about how anomaly evidence is distributed and structured across the image, often causing normal and anomalous scores to overlap. We propose StructCore, a training-free, structure-aware image-level scoring method that goes beyond max pooling. Given an anomaly score map, StructCore computes a low-dimensional structural descriptor phi(S) that captures distributional and spatial characteristics, and refines image-level scoring via a diagonal Mahalanobis calibration estimated from train-good samples, without modifying pixel-level localization. StructCore achieves image-level AUROC scores of 99.6% on MVTec AD and 98.4% on VisA, demonstrating robust image-level anomaly detection by exploiting structural signatures missed by max pooling.

Coordinating actions is the most fundamental form of cooperation in multi-agent reinforcement learning (MARL). Successful decentralized decision-making often depends not only on good individual actions, but on selecting compatible actions across agents to synchronize behavior, avoid conflicts, and satisfy global constraints. In this paper, we propose Action Graph Policies (AGP), that model dependencies among agents' available action choices. It constructs, what we call, \textit{coordination contexts}, that enable agents to condition their decisions on global action dependencies. Theoretically, we show that AGPs induce a strictly more expressive joint policy compared to fully independent policies and can realize coordinated joint actions that are provably more optimal than greedy execution even from centralized value-decomposition methods. Empirically, we show that AGP achieves 80-95\% success on canonical coordination tasks with partial observability and anti-coordination penalties, where other MARL methods reach only 10-25\%. We further demonstrate that AGP consistently outperforms these baselines in diverse multi-agent environments.

While multimodal large language models (MLLMs) have shown remarkable success across a wide range of tasks, long-form video understanding remains a significant challenge. In this study, we focus on video understanding by MLLMs. This task is challenging because processing a full stream of RGB frames is computationally intractable and highly redundant, as self-attention have quadratic complexity with sequence length. In this paper, we propose ReMoRa, a video MLLM that processes videos by operating directly on their compressed representations. A sparse set of RGB keyframes is retained for appearance, while temporal dynamics are encoded as a motion representation, removing the need for sequential RGB frames. These motion representations act as a compact proxy for optical flow, capturing temporal dynamics without full frame decoding. To refine the noise and low fidelity of block-based motions, we introduce a module to denoise and generate a fine-grained motion representation. Furthermore, our model compresses these features in a way that scales linearly with sequence length. We demonstrate the effectiveness of ReMoRa through extensive experiments across a comprehensive suite of long-video understanding benchmarks. ReMoRa outperformed baseline methods on multiple challenging benchmarks, including LongVideoBench, NExT-QA, and MLVU.

This work presents HAWX, a hardware-aware scalable exploration framework that employs multi-level sensitivity scoring at different DNN abstraction levels (operator, filter, layer, and model) to guide selective integration of heterogeneous AxC blocks. Supported by predictive models for accuracy, power, and area, HAWX accelerates the evaluation of candidate configurations, achieving over 23* speedup in a layer-level search with two candidate approximate blocks and more than (3*106)* speedup at the filter-level search only for LeNet-5, while maintaining accuracy comparable to exhaustive search. Experiments across state-of-the-art DNN benchmarks such as VGG-11, ResNet-18, and EfficientNetLite demonstrate that the efficiency benefits of HAWX scale exponentially with network size. The HAWX hardware-aware search algorithm supports both spatial and temporal accelerator architectures, leveraging either off-the-shelf approximate components or customized designs.

We present a simple experiment that exposes a fundamental limitation in vision-language models (VLMs): the inability to accurately localize filled cells in binary grids when those cells lack textual identity. We generate fifteen 15x15 grids with varying density (10.7%-41.8% filled cells) and render each as two image types -- text symbols (. and #) and filled squares without gridlines -- then ask three frontier VLMs (Claude Opus, ChatGPT 5.2, and Gemini 3 Thinking) to transcribe them. In the text-symbol condition, Claude and ChatGPT achieve approximately 91% cell accuracy and 84% F1, while Gemini achieves 84% accuracy and 63% F1. In the filled-squares condition, all three models collapse to 60-73% accuracy and 29-39% F1. Critically, all conditions pass through the same visual encoder -- the text symbols are images, not tokenized text. The text-vs-squares F1 gap ranges from 34 to 54 points across models, demonstrating that VLMs behave as if they possess a high-fidelity text-recognition pathway for spatial reasoning that dramatically outperforms their native visual pathway. Each model exhibits a distinct failure mode in the squares condition -- systematic under-counting (Claude), massive over-counting (ChatGPT), and template hallucination (Gemini) -- but all share the same underlying deficit: severely degraded spatial localization for non-textual visual elements.

The growing importance of culturally-aware natural language processing systems has led to an increasing demand for resources that capture sociopragmatic phenomena across diverse languages. Nevertheless, Arabic-language resources for politeness detection remain under-explored, despite the rich and complex politeness expressions embedded in Arabic communication. In this paper, we introduce ADAB (Arabic Politeness Dataset), a new annotated Arabic dataset collected from four online platforms, including social media, e-commerce, and customer service domains, covering Modern Standard Arabic and multiple dialects (Gulf, Egyptian, Levantine, and Maghrebi). The dataset was annotated based on Arabic linguistic traditions and pragmatic theory, resulting in three classes: polite, impolite, and neutral. It contains 10,000 samples with linguistic feature annotations across 16 politeness categories and achieves substantial inter-annotator agreement (kappa = 0.703). We benchmark 40 model configurations, including traditional machine learning, transformer-based models, and large language models. The dataset aims to support research on politeness-aware Arabic NLP.

AI agents are increasingly used to solve real-world tasks by reasoning over multi-turn user interactions and invoking external tools. However, applying reinforcement learning to such settings remains difficult: realistic objectives often lack verifiable rewards and instead emphasize open-ended behaviors; moreover, RL for multi-turn, multi-step agentic tool use is still underexplored; and building and maintaining executable tool environments is costly, limiting scale and coverage. We propose CM2, an RL framework that replaces verifiable outcome rewards with checklist rewards. CM2 decomposes each turn's intended behavior into fine-grained binary criteria with explicit evidence grounding and structured metadata, turning open-ended judging into more stable classification-style decisions. To balance stability and informativeness, our method adopts a strategy of sparse reward assignment but dense evaluation criteria. Training is performed in a scalable LLM-simulated tool environment, avoiding heavy engineering for large tool sets. Experiments show that CM2 consistently improves over supervised fine-tuning. Starting from an 8B Base model and training on an 8k-example RL dataset, CM2 improves over the SFT counterpart by 8 points on tau^-Bench, by 10 points on BFCL-V4, and by 12 points on ToolSandbox. The results match or even outperform similarly sized open-source baselines, including the judging model. CM2 thus provides a scalable recipe for optimizing multi-turn, multi-step tool-using agents without relying on verifiable rewards. Code provided by the open-source community: https://github.com/namezhenzhang/CM2-RLCR-Tool-Agent.

Real-world data collection for embodied agents remains costly and unsafe, calling for scalable, realistic, and simulator-ready 3D environments. However, existing scene-generation systems often rely on rule-based or task-specific pipelines, yielding artifacts and physically invalid scenes. We present SAGE, an agentic framework that, given a user-specified embodied task (e.g., "pick up a bowl and place it on the table"), understands the intent and automatically generates simulation-ready environments at scale. The agent couples multiple generators for layout and object composition with critics that evaluate semantic plausibility, visual realism, and physical stability. Through iterative reasoning and adaptive tool selection, it self-refines the scenes until meeting user intent and physical validity. The resulting environments are realistic, diverse, and directly deployable in modern simulators for policy training. Policies trained purely on this data exhibit clear scaling trends and generalize to unseen objects and layouts, demonstrating the promise of simulation-driven scaling for embodied AI. Code, demos, and the SAGE-10k dataset can be found on the project page here: https://research.nvidia.com/labs/dir/sage/.

This paper presents an Internet of Things (IoT) application that utilizes an AI classifier for fast-object detection using the frame difference method. This method, with its shorter duration, is the most efficient and suitable for fast-object detection in IoT systems, which require energy-efficient applications compared to end-to-end methods. We have implemented this technique on three edge devices: AMD AlveoT M U50, Jetson Orin Nano, and Hailo-8T M AI Accelerator, and four models with artificial neural networks and transformer models. We examined various classes, including birds, cars, trains, and airplanes. Using the frame difference method, the MobileNet model consistently has high accuracy, low latency, and is highly energy-efficient. YOLOX consistently shows the lowest accuracy, lowest latency, and lowest efficiency. The experimental results show that the proposed algorithm has improved the average accuracy gain by 28.314%, the average efficiency gain by 3.6 times, and the average latency reduction by 39.305% compared to the end-to-end method. Of all these classes, the faster objects are trains and airplanes. Experiments show that the accuracy percentage for trains and airplanes is lower than other categories. So, in tasks that require fast detection and accurate results, end-to-end methods can be a disaster because they cannot handle fast object detection. To improve computational efficiency, we designed our proposed method as a lightweight detection algorithm. It is well suited for applications in IoT systems, especially those that require fast-moving object detection and higher accuracy.

We study instruction-based image editing under professional workflows and identify three persistent challenges: (i) editors often over-edit, modifying content beyond the user's intent; (ii) existing models are largely single-turn, while multi-turn edits can alter object faithfulness; and (iii) evaluation at around 1K resolution is misaligned with real workflows that often operate on ultra high-definition images (e.g., 4K). We propose Agent Banana, a hierarchical agentic planner-executor framework for high-fidelity, object-aware, deliberative editing. Agent Banana introduces two key mechanisms: (1) Context Folding, which compresses long interaction histories into structured memory for stable long-horizon control; and (2) Image Layer Decomposition, which performs localized layer-based edits to preserve non-target regions while enabling native-resolution outputs. To support rigorous evaluation, we build HDD-Bench, a high-definition, dialogue-based benchmark featuring verifiable stepwise targets and native 4K images (11.8M pixels) for diagnosing long-horizon failures. On HDD-Bench, Agent Banana achieves the best multi-turn consistency and background fidelity (e.g., IC 0.871, SSIM-OM 0.84, LPIPS-OM 0.12) while remaining competitive on instruction following, and also attains strong performance on standard single-turn editing benchmarks. We hope this work advances reliable, professional-grade agentic image editing and its integration into real workflows.

Predictive world models that simulate future observations under explicit camera control are fundamental to interactive AI. Despite rapid advances, current systems lack spatial persistence: they fail to maintain stable scene structures over long trajectories, frequently hallucinating details when cameras revisit previously observed locations. We identify that this geometric drift stems from reliance on screen-space positional embeddings, which conflict with the projective geometry required for 3D consistency. We introduce \textbf{ViewRope}, a geometry-aware encoding that injects camera-ray directions directly into video transformer self-attention layers. By parameterizing attention with relative ray geometry rather than pixel locality, ViewRope provides a model-native inductive bias for retrieving 3D-consistent content across temporal gaps. We further propose \textbf{Geometry-Aware Frame-Sparse Attention}, which exploits these geometric cues to selectively attend to relevant historical frames, improving efficiency without sacrificing memory consistency. We also present \textbf{ViewBench}, a diagnostic suite measuring loop-closure fidelity and geometric drift. Our results demonstrate that ViewRope substantially improves long-term consistency while reducing computational costs.

Loss landscapes are a powerful tool for understanding neural network optimization and generalization, yet traditional low-dimensional analyses often miss complex topological features. We present Landscaper, an open-source Python package for arbitrary-dimensional loss landscape analysis. Landscaper combines Hessian-based subspace construction with topological data analysis to reveal geometric structures such as basin hierarchy and connectivity. A key component is the Saddle-Minimum Average Distance (SMAD) for quantifying landscape smoothness. We demonstrate Landscaper's effectiveness across various architectures and tasks, including those involving pre-trained language models, showing that SMAD captures training transitions, such as landscape simplification, that conventional metrics miss. We also illustrate Landscaper's performance in challenging chemical property prediction tasks, where SMAD can serve as a metric for out-of-distribution generalization, offering valuable insights for model diagnostics and architecture design in data-scarce scientific machine learning scenarios.

Large language models (LLMs) achieve strong capabilities by scaling model capacity and training data, yet many real-world deployments rely on smaller models trained or adapted from low-resource data. This gap motivates the need for mechanisms to transfer knowledge from large, high-resource models to smaller, low-resource targets. While model merging provides an effective transfer mechanism, most existing approaches assume architecture-compatible models and therefore cannot directly transfer knowledge from large high-resource LLMs to heterogeneous low-resource targets. In this work, we propose a cross-architecture merging framework based on optimal transport (OT) that aligns activations to infer cross-neuron correspondences between heterogeneous models. The resulting transport plans are then used to guide direct weight-space fusion, enabling effective high-resource to low-resource transfer using only a small set of inputs. Extensive experiments across low-resource languages and specialized domains demonstrate consistent improvements over target models.

Social choice has become a foundational component of modern machine learning systems. From auctions and resource allocation to the alignment of large generative models, machine learning pipelines increasingly aggregate heterogeneous preferences and incentives into collective decisions. In effect, many contemporary machine learning systems already implement social choice mechanisms, often implicitly and without explicit normative scrutiny. This Review surveys differentiable social choice: an emerging paradigm that formulates voting rules, mechanisms, and aggregation procedures as learnable, differentiable models optimized from data. We synthesize work across auctions, decision aggregation, and preference learning, showing how classical axioms and impossibility results reappear as objectives, constraints, and optimization trade-offs. We conclude by identifying 18 open problems defining a new research agenda at the intersection of machine learning and social choice theory.

Transmission line defect detection remains challenging for automated UAV inspection due to the dominance of small-scale defects, complex backgrounds, and illumination variations. Existing RGB-based detectors, despite recent progress, struggle to distinguish geometrically subtle defects from visually similar background structures under limited chromatic contrast. This paper proposes CMAFNet, a Cross-Modal Alignment and Fusion Network that integrates RGB appearance and depth geometry through a principled purify-then-fuse paradigm. CMAFNet consists of a Semantic Recomposition Module that performs dictionary-based feature purification via a learned codebook to suppress modality-specific noise while preserving defect-discriminative information, and a Contextual Semantic Integration Framework that captures global spatial dependencies using partial-channel attention to enhance structural semantic reasoning. Position-wise normalization within the purification stage enforces explicit reconstruction-driven cross-modal alignment, ensuring statistical compatibility between heterogeneous features prior to fusion. Extensive experiments on the TLRGBD benchmark, where 94.5% of instances are small objects, demonstrate that CMAFNet achieves 32.2% mAP@50 and 12.5% APs, outperforming the strongest baseline by 9.8 and 4.0 percentage points, respectively. A lightweight variant reaches 24.8% mAP50 at 228 FPS with only 4.9M parameters, surpassing all YOLO-based detectors while matching transformer-based methods at substantially lower computational cost.

Reinforcement learning (RL) is widely used for humanoid control, with on-policy methods such as Proximal Policy Optimization (PPO) enabling robust training via large-scale parallel simulation and, in some cases, zero-shot deployment to real robots. However, the low sample efficiency of on-policy algorithms limits safe adaptation to new environments. Although off-policy RL and model-based RL have shown improved sample efficiency, the gap between large-scale pretraining and efficient finetuning on humanoids still exists. In this paper, we find that off-policy Soft Actor-Critic (SAC), with large-batch update and a high Update-To-Data (UTD) ratio, reliably supports large-scale pretraining of humanoid locomotion policies, achieving zero-shot deployment on real robots. For adaptation, we demonstrate that these SAC-pretrained policies can be finetuned in new environments and out-of-distribution tasks using model-based methods. Data collection in the new environment executes a deterministic policy while stochastic exploration is instead confined to a physics-informed world model. This separation mitigates the risks of random exploration during adaptation while preserving exploratory coverage for improvement. Overall, the approach couples the wall-clock efficiency of large-scale simulation during pretraining with the sample efficiency of model-based learning during fine-tuning. For code and videos, see https://lift-humanoid.github.io

Crowd counting remains challenging in variable-density scenes due to scale variations, occlusions, and the high computational cost of existing models. To address these issues, we propose RepSFNet (Reparameterized Single Fusion Network), a lightweight architecture designed for accurate and real-time crowd estimation. RepSFNet leverages a RepLK-ViT backbone with large reparameterized kernels for efficient multi-scale feature extraction. It further integrates a Feature Fusion module combining Atrous Spatial Pyramid Pooling (ASPP) and Context-Aware Network (CAN) to achieve robust, density-adaptive context modeling. A Concatenate Fusion module is employed to preserve spatial resolution and generate high-quality density maps. By avoiding attention mechanisms and multi-branch designs, RepSFNet significantly reduces parameters and computational complexity. The training objective combines Mean Squared Error and Optimal Transport loss to improve both count accuracy and spatial distribution alignment. Experiments conducted on ShanghaiTech, NWPU, and UCF-QNRF datasets demonstrate that RepSFNet achieves competitive accuracy while reducing inference latency by up to 34 percent compared to recent state-of-the-art methods, making it suitable for real-time and low-power edge computing applications.

We study a bi-level online provisioning and scheduling problem motivated by network resource allocation, where provisioning decisions are made at a slow time scale while queue-/state-dependent scheduling is performed at a fast time scale. We model this two-time-scale interaction using an upper-level online convex optimization (OCO) problem and a lower-level constrained Markov decision process (CMDP). Existing OCO typically assumes stateless decisions and thus cannot capture MDP network dynamics such as queue evolution. Meanwhile, CMDP algorithms typically assume a fixed constraint threshold, whereas in provisioning-and-scheduling systems, the threshold varies with online budget decisions. To address these gaps, we study bi-level OCO-CMDP learning under switching costs (budget reprovisioning/system reconfiguration) and cross-level constraints that couple budgets to scheduling decisions. Our new algorithm solves this learning problem via several non-trivial developments, including a carefully designed dual feedback that returns the budget multiplier as sensitivity information for the upper-level update and a lower level that solves a budget-adaptive safe exploration problem via an extended occupancy-measure linear program. We establish near-optimal regret and high-probability satisfaction of the cross-level constraints.

Machine unlearning, which aims to efficiently remove the influence of specific data from trained models, is crucial for upholding data privacy regulations like the ``right to be forgotten". However, existing research predominantly evaluates unlearning methods on relatively balanced forget sets. This overlooks a common real-world scenario where data to be forgotten, such as a user's activity records, follows a long-tailed distribution. Our work is the first to investigate this critical research gap. We find that in such long-tailed settings, existing methods suffer from two key issues: \textit{Heterogeneous Unlearning Deviation} and \textit{Skewed Unlearning Deviation}. To address these challenges, we propose FaLW, a plug-and-play, instance-wise dynamic loss reweighting method. FaLW innovatively assesses the unlearning state of each sample by comparing its predictive probability to the distribution of unseen data from the same class. Based on this, it uses a forgetting-aware reweighting scheme, modulated by a balancing factor, to adaptively adjust the unlearning intensity for each sample. Extensive experiments demonstrate that FaLW achieves superior performance.

We consider the problem of adaptively monitoring a wildfire front using a mobile agent (e.g., a drone), whose trajectory determines where sensor data is collected and thus influences the accuracy of fire propagation estimation. This is a challenging problem, as the stochastic nature of wildfire evolution requires the seamless integration of sensing, estimation, and control, often treated separately in existing methods. State-of-the-art methods either impose linear-Gaussian assumptions to establish optimality or rely on approximations and heuristics, often without providing explicit performance guarantees. To address these limitations, we formulate the fire front monitoring task as a stochastic optimal control problem that integrates sensing, estimation, and control. We derive an optimal recursive Bayesian estimator for a class of stochastic nonlinear elliptical-growth fire front models. Subsequently, we transform the resulting nonlinear stochastic control problem into a finite-horizon Markov decision process and design an information-seeking predictive control law obtained via a lower confidence bound-based adaptive search algorithm with asymptotic convergence to the optimal policy.

Quality of Service (QoS) prediction is one of the most fundamental problems in service computing and personalized recommendation. In the problem, there is a set of users and services, each associated with a set of descriptive features. Interactions between users and services produce feedback values, typically represented as numerical QoS metrics such as response time or availability. Given the observed feedback for a subset of user-service pairs, the goal is to predict the QoS values for the remaining pairs. A key challenge in QoS prediction is the inherent sparsity of user-service interactions, as only a small subset of feedback values is typically observed. To address this, we propose a self-augmented strategy that leverages a model's own predictions for iterative refinement. In particular, we partially mask the predicted values and feed them back into the model to predict again. Building on this idea, we design a self-augmented mixture-of-experts model, where multiple expert networks iteratively and collaboratively estimate QoS values. We find that the iterative augmentation process naturally aligns with the MoE architecture by enabling inter-expert communication: in the second round, each expert receives the first-round predictions and refines its output accordingly. Experiments on benchmark datasets show that our method outperforms existing baselines and achieves competitive results.

Diffusion Language Models (DLMs) enable parallel decoding via iterative denoising, where remasking strategies play a critical role in balancing inference speed and output quality. Existing methods predominantly rely on static confidence thresholds, overlooking the spatial-temporal dynamics of token confidence, causing unnecessary remasking. We propose Spatial-Temporal Token-Dynamics-Aware Responsive Remasking (STaRR), a training-free framework that dynamically adapts remasking decisions based on token confidence evolution. STaRR introduces two metrics, temporal variance and spatial deviance, to guide fine-grained, step-wise dynamic thresholding. We further introduce a step-wise dynamic thresholding strategy, further enhanced with responsiveness optimizations for scalability and robustness. Experiments show that STaRR achieves an average speedup of 4.1 and up to 8.9 while maintaining comparable accuracy.

LLM-based reasoning models have enabled the development of agentic systems that act as co-scientists, assisting in multi-step scientific analysis. However, evaluating these systems is challenging, as it requires realistic, end-to-end research scenarios that integrate data analysis, interpretation, and the generation of new insights from the experimental data. To address this limitation, we introduce HeurekaBench, a framework to create benchmarks with exploratory, open-ended research questions for experimental datasets. Each such question is grounded in a scientific study and its corresponding code repository, and is created using a semi-automated pipeline that leverages multiple LLMs to extract insights and generate candidate workflows, which are then verified against reported findings. We instantiate the framework in single-cell biology to obtain sc-HeurekaBench benchmark and use it to compare state-of-the-art single-cell agents. We further showcase the benefits of our benchmark for quantitatively analyzing current design choices in agentic systems. We find that the addition of a critic module can improve ill-formed responses for open-source LLM-based agents by up to 22% and close the gap with their closed-source counterparts. Overall, HeurekaBench sets a path toward rigorous, end-to-end evaluation of scientific agents, grounding benchmark construction in real scientific workflows.

Reasoning distillation has attracted increasing attention. It typically leverages a large teacher model to generate reasoning paths, which are then used to fine-tune a student model so that it mimics the teacher's behavior in training contexts. However, previous approaches have lacked a detailed analysis of the origins of the distilled model's capabilities. It remains unclear whether the student can maintain consistent behaviors with the teacher in novel test-time contexts, or whether it regresses to its original output patterns, raising concerns about the generalization of distillation models. To analyse this question, we introduce a cross-model Reasoning Distillation Provenance Tracing framework. For each action (e.g., a sentence) produced by the distilled model, we obtain the predictive probabilities assigned by the teacher, the original student, and the distilled model under the same context. By comparing these probabilities, we classify each action into different categories. By systematically disentangling the provenance of each action, we experimentally demonstrate that, in test-time contexts, the distilled model can indeed generate teacher-originated actions, which correlate with and plausibly explain observed performance on distilled model. Building on this analysis, we further propose a teacher-guided data selection method. Unlike prior approach that rely on heuristics, our method directly compares teacher-student divergences on the training data, providing a principled selection criterion. We validate the effectiveness of our approach across multiple representative teacher models and diverse student models. The results highlight the utility of our provenance-tracing framework and underscore its promise for reasoning distillation. We hope to share Reasoning Distillation Provenance Tracing and our insights into reasoning distillation with the community.

Federated learning (FL) supports privacy-preserving, decentralized machine learning (ML) model training by keeping data on client devices. However, non-independent and identically distributed (non-IID) data across clients biases updates and degrades performance. To alleviate these issues, we propose Clust-PSI-PFL, a clustering-based personalized FL framework that uses the Population Stability Index (PSI) to quantify the level of non-IID data. We compute a weighted PSI metric, $WPSI^L$, which we show to be more informative than common non-IID metrics (Hellinger, Jensen-Shannon, and Earth Mover's distance). Using PSI features, we form distributionally homogeneous groups of clients via K-means++; the number of optimal clusters is chosen by a systematic silhouette-based procedure, typically yielding few clusters with modest overhead. Across six datasets (tabular, image, and text modalities), two partition protocols (Dirichlet with parameter $α$ and Similarity with parameter S), and multiple client sizes, Clust-PSI-PFL delivers up to 18% higher global accuracy than state-of-the-art baselines and markedly improves client fairness by a relative improvement of 37% under severe non-IID data. These results establish PSI-guided clustering as a principled, lightweight mechanism for robust PFL under label skew.

We present SiLU network constructions whose approximation efficiency depends critically on proper hyperparameter tuning. For the square function $x^2$, with optimally chosen shift $a$ and scale $β$, we achieve approximation error $\varepsilon$ using a two-layer network of constant width, where weights scale as $β^{\pm k}$ with $k = \mathcal{O}(\ln(1/\varepsilon))$. We then extend this approach through functional composition to Sobolev spaces, we obtain networks with depth $\mathcal{O}(1)$ and $\mathcal{O}(\varepsilon^{-d/n})$ parameters under optimal hyperparameters settings. Our work highlights the trade-off between architectural depth and activation parameter optimization in neural network approximation theory.