Papers

Force and torque (F/T) sensors enable contact-aware control by providing reactive feedback, but they are often fragile and expensive. To overcome these limitations, sensorless methods estimate F/T or wrench solely from robot proprioception, and have shown success in slow interaction tasks such as grasping. However, their low-pass characteristics limit the estimation of high-frequency signals, which are critical in rapid-contact tasks such as grinding. Communication delays can also make their estimates outdated during deployment, but few methods address this directly. To bridge these gaps, we propose a Frequency-aware Decomposition Network (FDN) to estimate vibration-rich wrench in a sensorless, multi-step-ahead manner. Considering higher-frequency stochasticity, FDN spectrally decomposes the wrench horizon into a low-frequency trend and a high-frequency residual, and estimates each by pointwise regression and a learned conditional distribution, respectively. The frequency-aware layers impose band decomposition priors on the outputs and adaptively enhance frequency amplitudes of the inputs. FDN requires neither an identified robot model nor an F/T sensor during estimation. On real-world grinding data from our 6-DoF hydraulic manipulator, FDN reduces high-frequency amplitude error by up to 47% over the baselines under assumed time delays and maintains competitive low-frequency pointwise accuracy, while the baselines fail to balance these two. We also find multi-step-ahead estimation feasible, with FDN estimating a 1,000 ms horizon within 11 ms on a single CPU thread. Ablation studies further support our design choices. In an exploratory study, transferring wrench dynamics learned from an open-source everyday manipulation dataset reduces low-frequency error by 8%, while high-frequency dynamics appear domain-specific.

Multimodal language models (MLLMs) are increasingly paired with vision tools (e.g., depth, flow, correspondence) to enhance visual reasoning. However, despite access to these tool-generated visual cues, MLLMs often fail to benefit from them. Existing approaches typically feed raw tool outputs into the model, but these dense, pixel-level representations are misaligned with the language-native reasoning strengths of LLMs, leading to weak perception and reliance on language priors. We argue that, in problems where vision tools can provide the necessary visual cues, the bottleneck is not more tool calls or larger MLLMs, it is how tool outputs are represented. We introduce Perception Programs (P$^2$), a training-free, model-agnostic method that rewrites tool outputs into compact, structured, language-native summaries that MLLMs can directly parse and reason over. Across six perception-centric tasks in BLINK, P$^2$ consistently yields large improvements over base models and raw tool-augmented baselines. With GPT-5 Mini as the base model, P$^2$ raises its accuracy from 41.35% to 86.47% on multi-view reasoning, from 52.42% to 81.45% on relative depth, and achieves a 22% average gain across tasks, setting new state-of-the-art results. Even on smaller MLLMs, e.g., InternVL3.5-4B and Qwen3VL-4B, we observe 15-40% absolute gains from P$^2$, surpassing prior agentic, supervised, and RL-based tool-use methods-without any training or model modifications.

Deep learning (DL) compilers rely on cost models and auto-tuning to optimize tensor programs for target hardware. However, existing approaches depend on large offline datasets, incurring high collection costs and offering suboptimal transferability across platforms. In this paper, we introduce TCL, a novel efficient and transferable compiler framework for fast tensor program optimization across diverse hardware platforms to address these challenges. Specifically, TCL is built on three core enablers: (1) the RDU Sampler, a data-efficient active learning strategy that selects only 10% of tensor programs by jointly optimizing Representativeness, Diversity, and Uncertainty, substantially reducing data collection costs while maintaining near-original model accuracy; (2) a new Mamba-based cost model that efficiently captures long-range schedule dependencies while achieving a favorable trade-off between prediction accuracy and computational cost through reduced parameterization and lightweight sequence modeling; and (3) a continuous knowledge distillation framework that effectively and progressively transfers knowledge across multiple hardware platforms while avoiding the parameter explosion and data dependency issues typically caused by traditional multi-task learning. Extensive experiments validate the effectiveness of each individual enabler and the holistic TCL framework. When optimizing a range of mainstream DL models on both CPU and GPU platforms, TCL achieves, on average, 16.8x and 12.48x faster tuning time, and 1.20x and 1.13x lower inference latency, respectively, compared to Tenset-MLP.

Visual tokenizers map high-dimensional raw pixels into a compressed representation for downstream modeling. Beyond compression, tokenizers dictate what information is preserved and how it is organized. A de facto standard approach to video tokenization is to represent a video as a spatiotemporal 3D grid of tokens, each capturing the corresponding local information in the original signal. This requires the downstream model that consumes the tokens, e.g., a text-to-video model, to learn to predict all low-level details "pixel-by-pixel" irrespective of the video's inherent complexity, leading to high learning complexity. We present VideoFlexTok, which represents videos with a variable-length sequence of tokens structured in a coarse-to-fine manner – where the first tokens (emergently) capture abstract information, such as semantics and motion, and later tokens add fine-grained details. The generative flow decoder enables realistic video reconstructions from any token count. This representation structure allows adapting the token count according to downstream needs and encoding videos longer than the baselines with the same budget. We evaluate VideoFlexTok on class- and text-to-video generative tasks and show that it leads to more efficient training compared to 3D grid tokens, e.g., achieving comparable generation quality (gFVD and ViCLIP Score) with a 5x smaller model (1.1B vs 5.2B). Finally, we demonstrate how VideoFlexTok can enable long video generation without prohibitive computational cost by training a text-to-video model on 10-second 81-frame videos with only 672 tokens, 8x fewer than a comparable 3D grid tokenizer.

Radio frequency fingerprints (RFFs) enable secure wireless authentication but struggle in open-set scenarios with unknown devices and varying channels. Existing methods face challenges in generalization and incur high computational costs. We propose a lightweight, self-adaptive RFF extraction framework using Low-Rank Adaptation (LoRA). By pretraining LoRA modules per environment, our method enables fast adaptation to unseen channel conditions without full retraining. During inference, a weighted combination of LoRAs dynamically enhances feature extraction. Experimental results demonstrate a 15% reduction in equal error rate (EER) compared to non-finetuned baselines and an 83% decrease in training time relative to full fine-tuning, using the same training dataset. This approach provides a scalable and efficient solution for open-set RFF authentication in dynamic wireless vehicular networks.

We study fixed-design random feature ridge regression beyond the mean-kernel approximation. The expectation is taken over the frozen-feature ensemble, conditional on the training sample. Because the predictor is a nonlinear function of the empirical kernel, its mean training error, test error, and conditional generalization gap depend on centered kernel covariances as well as on mean kernel objects. We derive the covariance-level, or one-loop, corrections from a finite resolvent identity. This avoids an almost-sure Neumann-series assumption and yields an explicit remainder bound for the training error. The test correction additionally requires mixed train–test covariance tensors and therefore cannot, in general, be recovered from the train-restricted kernel law alone. Numerical experiments show that the covariance term removes the leading inverse-width discrepancy in a controlled regularization regime, while an ablation that discards the mixed train–test covariance fails for the test error even though the training correction is unchanged. They also identify a width–regularization boundary, measured by a resolvent-fluctuation diagnostic, beyond which the second-order truncation loses accuracy. All inverse-width statements are for fixed sample size, depth, design, and positive kernel-level regularization. The analysis concerns frozen features and does not model feature learning or neural-tangent-kernel evolution.

Adversarial training (AT) is an effective defense for large language models (LLMs) against jailbreak attacks, but performing AT on LLMs is costly. To improve the efficiency of AT for LLMs, recent studies propose continuous AT (CAT) that searches for adversarial inputs within the continuous embedding space of LLMs during AT. While CAT has achieved empirical success, its underlying mechanism, i.e., why adversarial perturbations in the embedding space can help LLMs defend against jailbreak prompts synthesized in the input token space, remains unknown. This paper presents the first theoretical analysis of CAT on LLMs based on in-context learning (ICL) theory. For linear transformers trained with adversarial examples from the embedding space on in-context linear regression tasks, we prove a robust generalization bound that has a negative correlation with the perturbation radius in the embedding space. This clearly explains why CAT can defend against jailbreak prompts from the LLM's token space. Further, the robust bound shows that the robustness of an adversarially trained LLM is closely related to the singular values of its embedding matrix. Based on this, we propose to improve LLM CAT by introducing an additional regularization term, which depends on singular values of the LLM's embedding matrix, into the objective function of CAT. Experiments on real-world LLMs demonstrate that our method can help LLMs achieve a better jailbreak robustness-utility tradeoff. The code is available at https://github.com/fshp971/continuous-adv-icl.

Dense Associative Memory (DAM) generalizes Hopfield networks through higher-order interactions and achieves storage capacity that scales as $O(N^{n-1})$ under suitable pattern separation conditions. Existing dynamical analyses primarily study the thermodynamic limit $N\to\infty$ with randomly sampled patterns and therefore do not provide finite-size guarantees or explicit convergence rates. We develop an algorithmic analysis of DAM retrieval dynamics that yields finite-$N$ guarantees under explicit, verifiable pattern conditions. Under a separation assumption and a bounded-interference condition at high loading, we prove geometric convergence of asynchronous retrieval dynamics, which implies $O(\log N)$ convergence time once the trajectory enters the basin of attraction. We further establish adversarial robustness bounds expressed through an explicit margin condition that quantifies the number of corrupted bits tolerable per sweep, and derive capacity guarantees that scale as $\Theta(N^{n-1})$ up to polylogarithmic factors in the worst case, while recovering the classical $\Theta(N^{n-1})$ scaling for random pattern ensembles. Finally, we show that DAM retrieval dynamics admit a potential-game interpretation that ensures convergence to pure Nash equilibria under asynchronous updates. Complete proofs are provided in the appendices, together with preliminary experiments that illustrate the predicted convergence, robustness, and capacity scaling behavior.

Inferring latent interaction structures from observed dynamics is a fundamental inverse problem in many-body interacting systems. Most neural approaches rely on black-box surrogates over trainable graphs, achieving accuracy at the expense of mechanistic interpretability. Symbolic regression offers explicit dynamical equations and stronger inductive biases, but typically assumes known topology and a fixed function library. We propose COSINE (Co-Optimization of Symbolic Interactions and Network Edges), a differentiable framework that jointly discovers interaction graphs and sparse symbolic dynamics. To overcome the limitations of fixed symbolic libraries, COSINE further incorporates an outer-loop large language model that adaptively prunes and expands the hypothesis space using feedback from the inner optimization loop. Experiments on synthetic systems and large-scale real-world epidemic data demonstrate robust structural recovery and compact, mechanism-aligned dynamical expressions. Code: https://anonymous.4open.science/r/COSINE-6D43.

Neuromorphic vision sensors offer low latency and high dynamic range, but their deployment in public spaces raises severe data protection concerns. Recent Event-to-Video (E2V) models can reconstruct high-fidelity intensity images from sparse event streams, inadvertently exposing human identities. Current obfuscation methods, such as masking or scrambling, corrupt the spatio-temporal structure, severely degrading data utility for downstream perception tasks. In this paper, to the best of our knowledge, we present the first generative anonymization framework for event streams to resolve this utility-privacy trade-off. By bridging the modality gap between asynchronous events and standard spatial generative models, our pipeline projects events into an intermediate intensity representation, leverages pretrained models to synthesize realistic, non-existent identities, and re-encodes the features back into the neuromorphic domain. Experiments demonstrate that our method reliably prevents identity recovery from E2V reconstructions while preserving the structural data integrity required for downstream vision tasks. Finally, to facilitate rigorous evaluation, we introduce a novel, synchronized real-world event and RGB dataset captured via precise robotic trajectories, providing a robust benchmark for future research in privacy-preserving neuromorphic vision.

FlashAttention-style online softmax enables exact attention computation with linear memory by streaming score tiles through on-chip memory and maintaining a running maximum and normalizer. However, as attention kernels approach peak tensor-core/cube-core throughput on modern accelerators, non-matmul components of online softmax – especially per-tile rowmax and rowsum reductions and rescale chains – can become vector or SIMD limited and dominate latency. This paper revisits FlashAttention and proposes Vector Relieved Flash Attention (VFA), a hardware-friendly method that reduces rowmax-driven updates of the running maximum while retaining the online-softmax structure. VFA initializes the running maximum via a cheap approximation from key-block representations, reorders key-block traversal to prioritize high-impact sink and local blocks, and freezes the maximum for remaining blocks to avoid repeated reductions and rescaling. We further integrate VFA with block-sparse skipping methods such as BLASST to form Vector Relieved Sparse Attention (VSA), which reduces both block count and per-block overhead. Notably, VFA and VSA completely avoid the conditional rescale operation in the update stage used in FA4.0. Extensive evaluations on benchmarks including MMLU and MATH500, together with attention statistics, verify our design: (i) sink and local reordering stabilizes the running maximum early; (ii) simple Q and K block summaries fail due to intra-block heterogeneity; (iii) m-initialization is required when maxima appear in middle blocks. Overall, VFA and VSA efficiently alleviate online-softmax reduction bottlenecks without performance loss. Compared to the C16V32 baseline, C8V32, C4V32 and C4V16 achieve nearly two times speedup on modern hardware while hitting the vector bottleneck. With upcoming architecture improvements, C4V16 will deliver six times speedup by enhancing exponent capacity.

While 4-bit quantization is essential for high-throughput deployment of Large Language Models, activation outliers often lead to significant accuracy degradation due to the restricted dynamic range of low-bit formats. In this paper, we systematically investigate the spatial distribution of outliers and demonstrate a token-persistent structural clustering effect, where high-magnitude outliers consistently occupy fixed channels across tokens. Building on this insight, we propose OSC, a hardware-efficient framework for outlier suppression. During inference, OSC executes a dual-path computation consisting of a low-precision 4-bit General Matrix Multiplication (GEMM) path and a high-precision 16-bit branch GEMM path. Specifically, OSC uses an offline group-wise strategy to identify the channels where outliers are located and then performs structured sub-tensor extraction to coalesce these scattered activation channels into a compact dense tensor online. This mechanism implements outlier protection through regularized and high-throughput GEMM operations, achieving a seamless fit with modern 4-bit micro-scaling hardware. Furthermore, for the inputs of W2 where outlier clustering is less pronounced, we integrate a fallback strategy to FP8. Evaluation on Qwen3-8B and Qwen3-30B restricts the average accuracy drop to 2.19 and 1.12 points, respectively. Notably, OSC is highly hardware-friendly, achieving a peak speedup of 1.78x over the W8A8 GEMM baseline on a modern AI accelerator.

Federated learning (FL) is a distributed paradigm that coordinates massive local clients to collaboratively train a global model via stage-wise local training processes on the heterogeneous dataset. Previous works have implicitly studied that FL suffers from the “client-drift” problem, which is caused by the inconsistent optimum across local clients. However, till now it still lacks solid theoretical analysis to explain the impact of this local inconsistency. To alleviate the negative impact of “client drift” and explore its substance in FL, in this paper, we first propose an efficient FL algorithm FedInit, which allows employing the personalized relaxed initialization state at the beginning of each local training stage. Specifically, FedInit initializes the local state by moving away from the current global state towards the reverse direction of the latest local state. Moreover, to further understand how inconsistency disrupts performance in FL, we introduce the excess risk analysis and study the divergence term to investigate the test error in FL. Our studies show that optimization error is not sensitive to this local inconsistency, while it mainly affects the generalization error bound. Extensive experiments are conducted to validate its efficiency. The proposed FedInit method could achieve comparable results compared to several advanced benchmarks without any additional training or communication costs. Meanwhile, the stage-wise personalized relaxed initialization could also be incorporated into several current advanced algorithms to achieve higher generalization performance in the FL paradigm.

Adversarial robustness is essential for deploying neural networks in safety-critical applications, yet standard evaluation methods either require expensive adversarial attacks or report only a single aggregate score that obscures how robustness is distributed across classes. We introduce the GF-Score (GREAT-Fairness Score), a framework that decomposes the certified GREAT Score into per-class robustness profiles and quantifies their disparity through four metrics grounded in welfare economics: the Robustness Disparity Index (RDI), the Normalized Robustness Gini Coefficient (NRGC), Worst-Case Class Robustness (WCR), and a Fairness-Penalized GREAT Score (FP-GREAT). The framework further eliminates the original method's dependence on adversarial attacks through a self-calibration procedure that tunes the temperature parameter using only clean accuracy correlations. Evaluating 22 models from RobustBench across CIFAR-10 and ImageNet, we find that the decomposition is exact, that per-class scores reveal consistent vulnerability patterns (e.g., “cat” is the weakest class in 76% of CIFAR-10 models), and that more robust models tend to exhibit greater class-level disparity. These results establish a practical, attack-free auditing pipeline for diagnosing where certified robustness guarantees fail to protect all classes equally. We release our code on \href{https://github.com/aryashah2k/gf-score}{GitHub}.

Stress remains a significant social problem for individuals in modern societies. This paper presents a machine learning approach for the automatic detection of stress of people in a social situation by combining two sensor systems that capture physiological and social responses. We compare the performance using different classifiers including support vector machine, AdaBoost, and k-nearest neighbor. Our experimental results show that by combining the measurements from both sensor systems, we could accurately discriminate between stressful and neutral situations during a controlled Trier social stress test (TSST). Moreover, this paper assesses the discriminative ability of each sensor modality individually and considers their suitability for real-time stress detection. Finally, we present an study of the most discriminative features for stress detection.

While Federated Learning (FL) mitigates direct data exposure, the resulting trained models remain susceptible to membership inference attacks (MIAs). This paper presents an empirical evaluation of Differential Privacy (DP) as a defense mechanism against MIAs in FL, leveraging the environment of the 2025 NIST Genomics Privacy-Preserving Federated Learning (PPFL) Red Teaming Event. To improve inference accuracy, we propose a stacking attack strategy that ensembles seven black-box estimators to train a meta-classifier on prediction probabilities and cross-entropy losses. We evaluate this methodology against target models under three privacy configurations: an unprotected convolutional neural network (CNN, $\epsilon=\infty$), a low-privacy DP model ($\epsilon=200$), and a high-privacy DP model ($\epsilon=10$). The attack outperforms all baselines in the No DP and Low Privacy settings and, critically, maintains measurable membership leakage at $\epsilon=200$ where a single-signal LiRA baseline collapses. Evaluated on an independent third-party benchmark, these results provide an empirical characterisation of how stacking-based inference degrades across calibrated DP tiers in FL.

In recent years, Sound AI is being increasingly used to predict machine failures. By attaching a microphone to the machine of interest, one can get real time data on machine behavior from the field. Traditionally, Convolutional Neural Net (CNN) architectures have been used to analyze spectrogram images generated from the sounds captured and predict if the machine is functioning as expected. CNN architectures seem to work well empirically even though they have biases like locality and parameter-sharing which may not be completely relevant for spectrogram analysis. With the successful application of transformer-based models in the field of image processing starting with Vision Transformer (ViT) in 2020, there has been significant interest in leveraging these in the field of Sound AI. Since transformer-based architectures have significantly lower inductive biases, they are expected to perform better than CNNs at spectrogram analysis given enough data. This paper demonstrates the effectiveness of transformer-driven architectures in analyzing Sound data and compares the embeddings they generate with CNNs on the specific task of machine fault detection.

Classical Fisher-information asymptotics describe the covariance of regular efficient estimators through the local quadratic approximation of the log-likelihood, and thus capture first-order geometry only. In curved models, including mixtures, curved exponential families, latent-variable models, and manifold-constrained parameter spaces, finite-sample behavior can deviate systematically from these predictions. We develop a coordinate-invariant, curvature-aware refinement by viewing a regular parametric family as a Riemannian manifold \((\Theta,g)\) with Fisher–Rao metric, immersed in \(L^2(\mu)\) through the square-root density map. Under suitable regularity and moment assumptions, we derive an \(n^{-2}\) correction to the leading \(n^{-1}I(\theta)^{-1}\) covariance term for score-root, first-order efficient estimators. The correction is governed by a tensor \(P_{ij}\) that decomposes canonically into three parts, an intrinsic Ricci-type contraction of the Fisher–Rao curvature tensor, an extrinsic Gram-type contraction of the second fundamental form, and a Hellinger discrepancy tensor encoding higher-order probabilistic information not determined by immersion geometry alone. The extrinsic term is positive semidefinite, the full correction is invariant under smooth reparameterization, and it vanishes identically for full exponential families. We then extend the picture to singular models, where Fisher information degenerates. Using resolution of singularities under an additive normal crossing assumption, we describe the resolved metric, the role of the real log canonical threshold in learning rates and posterior mean-squared error, and a curvature-based covariance expansion on the resolved space that recovers the regular theory as a special case. This framework also suggests geometric diagnostics of weak identifiability and curvature-aware principles for regularization and optimization.

The deployment of deep neural networks in safety-critical systems necessitates reliable and efficient uncertainty quantification (UQ). A practical and widespread strategy for UQ is repurposing stochastic regularizers as scalable approximate Bayesian inference methods, such as Monte Carlo Dropout (MCD) and MC-DropBlock (MCDB). However, this paradigm remains under-explored for Stochastic Depth (SD), a regularizer integral to the residual-based backbones of most modern architectures. While prior work demonstrated its empirical promise for segmentation, a formal theoretical connection to Bayesian variational inference and a benchmark on complex, multi-task problems like object detection are missing. In this paper, we first provide theoretical insights connecting Monte Carlo Stochastic Depth (MCSD) to principled approximate variational inference. We then present the first comprehensive empirical benchmark of MCSD against MCD and MCDB on state-of-the-art detectors (YOLO, RT-DETR) using the COCO and COCO-O datasets. Our results position MCSD as a robust and computationally efficient method that achieves highly competitive predictive accuracy (mAP), notably yielding slight improvements in calibration (ECE) and uncertainty ranking (AUARC) compared to MCD. We thus establish MCSD as a theoretically-grounded and empirically-validated tool for efficient Bayesian approximation in modern deep learning.

Task-adapted compressed sensing magnetic resonance imaging (CS-MRI) is emerging to address the specific demands of downstream clinical tasks with significantly fewer k-space measurements than required by Nyquist sampling. However, existing task-adapted CS-MRI methods suffer from the uncertainty problem for medical diagnosis and cannot achieve adaptive sampling in end-to-end optimization with reconstruction or clinical tasks. To address these limitations, we propose the first task-adapted CS-MRI from the information-theoretic perspective to simultaneously achieve probabilistic inference for uncertainty prediction and adapt to arbitrary sampling ratios and versatile clinical applications. Specifically, we formalize the task-adapted CS-MRI optimization problem by maximizing the mutual information between undersampled k-space measurements and clinical tasks to enable probabilistic inference for addressing the uncertainty problem. We leverage amortized optimization and construct tractable variational bounds for mutual information to jointly optimize sampling, reconstruction, and task-inference models, which enables flexible sampling ratio control using a single end-to-end trained model. Furthermore, the proposed framework addresses two kinds of distinct clinical scenarios within a unified approach, i.e., i) joint task and reconstruction, where reconstruction serves as an auxiliary process to enhance task performance; and ii) task implementation with suppressed reconstruction, applicable for privacy protection. Extensive experiments on large-scale MRI datasets demonstrate that the proposed framework achieves highly competitive performance on standard metrics like Dice compared to deterministic counterpart but provides better distribution matching to the ground-truth posterior distribution as measured by the generalized energy distance (GED).

Recent advances in deep learning underscore the need for systems that can not only acquire new knowledge through Continual Learning (CL) but also remove outdated, sensitive, or private information through Machine Unlearning (MU). However, while CL methods are well-developed, MU techniques remain in early stages, creating a critical gap for unified frameworks that depend on both capabilities. We find that naively combining existing CL and MU approaches results in knowledge leakage a gradual degradation of foundational knowledge across repeated adaptation cycles. To address this, we formalize Continual Learning Unlearning (CLU) as a unified paradigm with three key goals: (i) precise deletion of unwanted knowledge, (ii) efficient integration of new knowledge while preserving prior information, and (iii) minimizing knowledge leakage across cycles. We propose Bi-Directional Low-Rank Adaptation (BID-LoRA), a novel framework featuring three dedicated adapter pathways-retain, new, and unlearn applied to attention layers, combined with escape unlearning that pushes forget-class embeddings to positions maximally distant from retained knowledge, updating only 5% of parameters. Experiments on CIFAR-100 show that BID-LoRA outperforms CLU baselines across multiple adaptation cycles. We further evaluate on CASIA-Face100, a curated face recognition subset, demonstrating practical applicability to real-world identity management systems where new users must be enrolled and withdrawn users removed.

Text-based web agents offer computational efficiency for autonomous web navigation, yet developing robust agents remains challenging due to the noisy and heterogeneous nature of real-world HTML. Standard Supervised Fine-Tuning (SFT) approaches fail in two critical dimensions: they lack discrimination capabilities to reject plausible but incorrect elements in densely populated pages, and exhibit limited generalization to unseen website layouts. To address these challenges, we introduce the Triton dataset (590k instances) and a progressive training curriculum. Triton is constructed via Structural-Semantic Hard Negative Mining, which explicitly mines topologically similar distractors, and a Dual-Agent Consensus pipeline that synthesizes diverse cross-domain tasks with strict verification. Building upon this foundation, our progressive curriculum produces three models: Triton-SFT-32B for basic imitation, Triton-ORPO-32B for robust discrimination via Odds Ratio Preference Optimization, and Triton-GRPO-32B for long-horizon consistency through Group Relative Policy Optimization. Empirical evaluation on Mind2Web demonstrates that Triton-GRPO-32B achieves state-of-the-art performance among open-source models with 58.7% Step Success Rate, surpassing GPT-4.5 (42.4%) and Claude-4.5 (41.4%) by over 16%, validating that specialized data curriculum outweighs raw parameter scale for web navigation.

Vision-language models(VLMs) are increasingly applied to visual stock price forecasting, yet existing benchmarks inadequately evaluate their understanding of stock price in candlestick charts. First, prior studies fail to isolate VLMs' comprehension of visual inputs genuinely improves predictive performance and whether VLMs truly comprehend candlestick patterns. Further, most existing datasets and evaluation setups are designed around single-period or tabular inputs. However, human analysts strongly rely on multi-scale candlestick charts, where longer-term horizons capture trend direction and shorter-term horizons provide cues for inflection points, making it difficult to systematically assess VLMs' ability to integrate short-term and long-term visual market dynamics. To bridge this gap, we construct a multi-scale candlestick charts dataset and a standardized evaluation framework to assess VLMs' ability to utilize multi-scale visual market signals. Evaluation combines confusion-matrix-based diagnostics with information coefficient(IC) time series metrics and includes XGBoost as a feature-based temporal baseline. Using this dataset, we benchmark representative VLMs and analyze their ability to leverage multi-scale stock price data. Experimental results show that most VLMs perform well only under persistent uptrend or downtrend conditions, while exhibiting weak predictive capability in more common market scenarios. We also identify significant prediction biases and limited sensitivity to explicitly specified forecast horizons in prompts, indicating inherent limitations in precise temporal reasoning.

End-to-end diffusion planning has shown strong potential for autonomous driving, but the physical feasibility of generated trajectories remains insufficiently addressed. In particular, generated trajectories may exhibit local geometric irregularities, violate trajectory-level kinematic constraints, or deviate from the drivable area, indicating that the commonly used noise-centric formulation in diffusion planning is not yet well aligned with the trajectory space where feasibility is more naturally characterized. To address this issue, we propose FeaXDrive, a feasibility-aware trajectory-centric diffusion planning method for end-to-end autonomous driving. The core idea is to treat the clean trajectory as the unified object for feasibility-aware modeling throughout the diffusion process. Built on this trajectory-centric formulation, FeaXDrive integrates adaptive curvature-constrained training to improve intrinsic geometric and kinematic feasibility, drivable-area guidance within reverse diffusion sampling to enhance consistency with the drivable area, and feasibility-aware GRPO post-training to further improve planning performance while balancing trajectory-space feasibility. Experiments on the NAVSIM benchmark show that FeaXDrive achieves strong closed-loop planning performance while substantially improving trajectory-space feasibility. These findings highlight the importance of explicitly modeling trajectory-space feasibility in end-to-end diffusion planning and provide a step toward more reliable and physically grounded autonomous driving planners.

Cloud networks increasingly rely on machine learning based Network Intrusion Detection Systems to defend against evolving cyber threats. However, real-world deployments are challenged by limited labeled data, non-stationary traffic, and adaptive adversaries. While semi-supervised learning can alleviate label scarcity, most existing approaches implicitly assume benign and stationary unlabeled traffic, leading to degraded performance in adversarial cloud environments. This paper proposes a robust semi-supervised temporal learning framework for cloud intrusion detection that explicitly addresses adversarial contamination and temporal drift in unlabeled network traffic. Operating on flow-level data, this framework combines supervised learning with consistency regularization, confidence-aware pseudo-labeling, and selective temporal invariance to conservatively exploit unlabeled traffic while suppressing unreliable samples. By leveraging the temporal structure of network flows, the proposed method improves robustness and generalization across heterogeneous cloud environments. Extensive evaluations on publicly available datasets (CIC-IDS2017, CSE-CIC-IDS2018, and UNSW-NB15) under limited-label conditions demonstrate that the proposed framework consistently outperforms state-of-the-art supervised and semi-supervised network intrusion detection systems in detection performance, label efficiency, and resilience to adversarial and non-stationary traffic.

We study finite horizon reachable set estimation for unknown discrete-time dynamical systems using only sampled state trajectories. Rather than treating scenario optimization as a black-box tool, we show how it can be tailored to reachable set estimation, where one must learn a family of sets based on whole trajectories, while preserving probabilistic guarantees on future trajectory inclusion for the entire horizon. To this end, we formulate a relaxed scenario program with slack variables that yields a tunable trade-off between reachable set size and out-of-sample trajectory inclusion over the horizon, thereby reducing sensitivity to outliers. Leveraging the recent results in adversarially robust scenario optimization, we then extend this formulation to account for bounded adversarial perturbations of the observed trajectories and derive a posteriori probabilistic guarantees on future trajectory inclusion. When probability distribution shifts in the Wasserstein distance occur, we obtain an explicit bound on how gracefully the theoretical probabilistic guarantees degrade. For different geometries, i.e., $p$-norm balls, ellipsoids, and zonotopes, we derive tractable convex reformulations and corroborate our theoretical results in simulation.

Despite the recent success of large language models (LLMs) in time-series forecasting, most existing methods still adopt a Deep Synchronous Fusion strategy, where dense interactions between textual and temporal features are enforced at every layer of the network. This design overlooks the inherent granularity mismatch between modalities and leads to what we term semantic perceptual dissonance: high-level abstract semantics provided by the LLM become inappropriately entangled with the low-level, fine-grained numerical dynamics of time series, making it difficult for semantic priors to effectively guide forecasting. To address this issue, we propose TimeSAF, a new framework based on hierarchical asynchronous fusion. Unlike synchronous approaches, TimeSAF explicitly decouples unimodal feature learning from cross-modal interaction. It introduces an independent cross-modal semantic fusion trunk, which uses learnable queries to aggregate global semantics from the temporal and prompt backbones in a bottom-up manner, and a stage-wise semantic refinement decoder that asynchronously injects these high-level signals back into the temporal backbone. This mechanism provides stable and efficient semantic guidance while avoiding interference with low-level temporal dynamics. Extensive experiments on standard long-term forecasting benchmarks show that TimeSAF significantly outperforms state-of-the-art baselines, and further exhibits strong generalization in both few-shot and zero-shot transfer settings.

Large language models (LLMs) face a fundamental trade-off between computational efficiency (e.g., number of parameters) and output quality, especially when deployed on computationally limited devices such as phones or laptops. One way to address this challenge is by following the example of humans and have models ask for help when they believe they are incapable of solving a problem on their own; we can overcome this trade-off by allowing smaller models to respond to queries when they believe they can provide good responses, and deferring to larger models when they do not believe they can. To this end, in this paper, we investigate the viability of Predict-Answer/Act (PA) and Reason-Predict-Reason-Answer/Act (RPRA) paradigms where models predict – prior to responding – how an LLM judge would score their output. We evaluate three approaches: zero-shot prediction, prediction using an in-context report card, and supervised fine-tuning. Our results show that larger models (particularly reasoning models) perform well when predicting generic LLM judges zero-shot, while smaller models can reliably predict such judges well after being fine-tuned or provided with an in-context report card. Altogether, both approaches can substantially improve the prediction accuracy of smaller models, with report cards and fine-tuning achieving mean improvements of up to 55% and 52% across datasets, respectively. These findings suggest that models can learn to predict their own performance limitations, paving the way for more efficient and self-aware AI systems.

Group Relative Policy Optimization (GRPO) enhances LLM reasoning but often induces overconfidence, where incorrect responses yield lower perplexity than correct ones, degrading relative calibration as described by the Area Under the Curve (AUC). Existing approaches either yield limited improvements in calibration or sacrifice gains in reasoning accuracy. We first prove that this degradation in GRPO-style algorithms stems from their uncertainty-agnostic advantage estimation, which inevitably misaligns optimization gradients with calibration. This leads to improved accuracy at the expense of degraded calibration. We then propose Calibration-Aware Policy Optimization (CAPO). It adopts a logistic AUC surrogate loss that is theoretically consistent and admits regret bound, enabling uncertainty-aware advantage estimation. By further incorporating a noise masking mechanism, CAPO achieves stable learning dynamics that jointly optimize calibration and accuracy. Experiments on multiple mathematical reasoning benchmarks show that CAPO-1.5B significantly improves calibration by up to 15% while achieving accuracy comparable to or better than GRPO, and further boosts accuracy on downstream inference-time scaling tasks by up to 5%. Moreover, when allowed to abstain under low-confidence conditions, CAPO achieves a Pareto-optimal precision-coverage trade-off, highlighting its practical value for hallucination mitigation.

The post-training pipeline for diffusion models currently has two stages: supervised fine-tuning (SFT) on curated data and reinforcement learning (RL) with reward models. A fundamental gap separates them. SFT optimizes the denoiser only on ground-truth states sampled from the forward noising process; once inference deviates from these ideal states, subsequent denoising relies on out-of-distribution generalization rather than learned correction, exhibiting the same exposure bias that afflicts autoregressive models, but accumulated along the denoising trajectory instead of the token sequence. RL can in principle address this mismatch, yet its terminal reward signal is sparse, suffers from credit-assignment difficulty, and risks reward hacking. We propose SOAR (Self-Correction for Optimal Alignment and Refinement), a bias-correction post-training method that fills this gap. Starting from a real sample, SOAR performs a single stop-gradient rollout with the current model, re-noises the resulting off-trajectory state, and supervises the model to steer back toward the original clean target. The method is on-policy, reward-free, and provides dense per-timestep supervision with no credit-assignment problem. On SD3.5-Medium, SOAR improves GenEval from 0.70 to 0.78 and OCR from 0.64 to 0.67 over SFT, while simultaneously raising all model-based preference scores. In controlled reward-specific experiments, SOAR surpasses Flow-GRPO in final metric value on both aesthetic and text-image alignment tasks, despite having no access to a reward model. Since SOAR's base loss subsumes the standard SFT objective, it can directly replace SFT as a stronger first post-training stage after pretraining, while remaining fully compatible with subsequent RL alignment.