Automatic speech quality assessment has become increasingly important as modern speech generation systems continue to advance, while human listening tests remain costly, time-consuming, and difficult to scale. Most existing learning-based assessment models rely primarily on scarce human-annotated mean opinion score (MOS) data, which limits robustness and generalization, especially when training across heterogeneous datasets. In this work, we propose UrgentMOS, a unified speech quality assessment framework that jointly learns from diverse objective and perceptual quality metrics, while explicitly tolerating the absence of arbitrary subsets of metrics during training. By leveraging complementary quality facets under heterogeneous supervision, UrgentMOS enables effective utilization of partially annotated data and improves robustness when trained on large-scale, multi-source datasets. Beyond absolute score prediction, UrgentMOS explicitly models pairwise quality preferences by directly predicting comparative MOS (CMOS), making it well suited for preference-based evaluation scenarios commonly adopted in system benchmarking. Extensive experiments across a wide range of speech quality datasets, including simulated distortions, speech enhancement, and speech synthesis, demonstrate that UrgentMOS consistently achieves state-of-the-art performance in both absolute and comparative evaluation settings.

This work presents a speech-to-text system "Pisets" for scientists and journalists which is based on a three-component architecture aimed at improving speech recognition accuracy while minimizing errors and hallucinations associated with the Whisper model. The architecture comprises primary recognition using Wav2Vec2, false positive filtering via the Audio Spectrogram Transformer (AST), and final speech recognition through Whisper. The implementation of curriculum learning methods and the utilization of diverse Russian-language speech corpora significantly enhanced the system's effectiveness. Additionally, advanced uncertainty modeling techniques were introduced, contributing to further improvements in transcription quality. The proposed approaches ensure robust transcribing of long audio data across various acoustic conditions compared to WhisperX and the usual Whisper model. The source code of "Pisets" system is publicly available at GitHub: https://github.com/bond005/pisets.

Learning in smooth games fundamentally differs from standard minimization due to rotational dynamics, which invalidate classical hyperparameter tuning strategies. Despite their practical importance, effective methods for tuning in games remain underexplored. A notable example is LookAhead (LA), which achieves strong empirical performance but introduces additional parameters that critically influence performance. We propose a principled approach to hyperparameter selection in games by leveraging frequency estimation of oscillatory dynamics. Specifically, we analyze oscillations both in continuous-time trajectories and through the spectrum of the discrete dynamics in the associated frequency-based space. Building on this analysis, we introduce \emph{Modal LookAhead (MoLA)}, an extension of LA that selects the hyperparameters adaptively to a given problem. We provide convergence guarantees and demonstrate in experiments that MoLA accelerates training in both purely rotational games and mixed regimes, all with minimal computational overhead.

This report addresses larger-than-memory image analysis for petascale datasets such as 1.4 PB electron-microscopy volumes and 150 TB human-organ atlases. We argue that performance is fundamentally I/O-bound. We show that structuring analysis as streaming passes over data is crucial. For 3D volumes, two representations are popular: stacks of 2D slices (e.g., directories or multi-page TIFF) and 3D chunked layouts (e.g., Zarr/HDF5). While for a few algorithms, chunked layout on disk is crucial to keep disk I/O at a minimum, we show how the slice-based streaming architecture can be built on top of either image representation in a manner that minimizes disk I/O. This is in particular advantageous for algorithms relying on neighbouring values, since the slicing streaming architecture is 1D, which implies that there are only 2 possible sweeping orders, both of which are aligned with the order in which images are read from the disk. This is in contrast to 3D chunks, in which any sweep cannot be done without accessing each chunk at least 9 times. We formalize this with sweep-based execution (natural 2D/3D orders), windowed operations, and overlap-aware tiling to minimize redundant access. Building on these principles, we introduce a domain-specific language (DSL) that encodes algorithms with intrinsic knowledge of their optimal streaming and memory use; the DSL performs compile-time and run-time pipeline analyses to automatically select window sizes, fuse stages, tee and zip streams, and schedule passes for limited-RAM machines, yielding near-linear I/O scans and predictable memory footprints. The approach integrates with existing tooling for segmentation and morphology but reframes pre/post-processing as pipelines that privilege sequential read/write patterns, delivering substantial throughput gains for extremely large images without requiring full-volume residency in memory.

In this paper, we propose \textbf{Superlinear attention}, a fully trainable multi-step attention architecture that achieves subquadratic complexity for long sequences while preserving \textbf{random context access} (a.k.a.\ structural non-exclusion): no eligible token position is structurally excluded from being selected for attention. Superlinear attention reformulates standard causal self-attention as a multi-step search problem with $N$ steps, yielding an overall complexity of $O(L^{1+\frac{1}{N}})$. To illustrate the architecture, we present a baseline $N=2$ implementation, which is algorithmically analogous to standard jump search. In this $O(L^{3/2})$ instantiation, the first step performs $O(L^{3/2})$ span-search to select relevant spans of the sequence, and the second step applies $O(L^{3/2})$ span-attention (standard attention restricted to the selected spans). In an upscaled $O(L^{1.54})$ configuration for robustness, we achieve an average decoding throughput of 114 tokens/sec at 1M context length and 80 tokens/sec at 10M context in our implementation on a modified 30B hybrid MoE model on a single B200 GPU. With limited training, we also obtain strong performance on the NIAH (Needle In A Haystack) task up to 256K context length, demonstrating that the routed span selection is learnable end-to-end. This paper emphasizes architectural formulation, scaling analysis, and systems feasibility, and presents initial validation; comprehensive quality evaluations across diverse long-context tasks are left to future work.

Visual information, such as subtitles in a movie, often helps automatic speech recognition. In this paper, we propose Donut-Whisper, an audio-visual ASR model with dual encoder to leverage visual information to improve speech recognition performance in both English and Chinese. Donut-Whisper combines the advantage of the linear and the Q-Former-based modality alignment structures via a cross-attention module, generating more powerful audio-visual features. Meanwhile, we propose a lightweight knowledge distillation scheme showcasing the potential of using audio-visual models to teach audio-only models to achieve better performance. Moreover, we propose a new multilingual audio-visual speech recognition dataset based on movie clips containing both Chinese and English partitions. As a result, Donut-Whisper achieved significantly better performance on both English and Chinese partition of the dataset compared to both Donut and Whisper large V3 baselines. In particular, an absolute 5.75% WER reduction and a 16.5% absolute CER reduction were achieved on the English and Chinese sets respectively compared to the Whisper ASR baseline.

Deep learning applications in Magnetic Resonance Imaging (MRI) predominantly operate on reconstructed magnitude images, a process that discards phase information and requires computationally expensive transforms. Standard neural network architectures rely on local operations (convolutions or grid-patches) that are ill-suited for the global, non-local nature of raw frequency-domain (k-Space) data. In this work, we propose a novel complex-valued Vision Transformer (kViT) designed to perform classification directly on k-Space data. To bridge the geometric disconnect between current architectures and MRI physics, we introduce a radial k-Space patching strategy that respects the spectral energy distribution of the frequency-domain. Extensive experiments on the fastMRI and in-house datasets demonstrate that our approach achieves classification performance competitive with state-of-the-art image-domain baselines (ResNet, EfficientNet, ViT). Crucially, kViT exhibits superior robustness to high acceleration factors and offers a paradigm shift in computational efficiency, reducing VRAM consumption during training by up to 68$\times$ compared to standard methods. This establishes a pathway for resource-efficient, direct-from-scanner AI analysis.

Existing automated attack suites operate as static ensembles with fixed sequences, lacking strategic adaptation and semantic awareness. This paper introduces the Agentic Reasoning for Methods Orchestration and Reparameterization (ARMOR) framework to address these limitations. ARMOR orchestrates three canonical adversarial primitives, Carlini-Wagner (CW), Jacobian-based Saliency Map Attack (JSMA), and Spatially Transformed Attacks (STA) via Vision Language Models (VLM)-guided agents that collaboratively generate and synthesize perturbations through a shared ``Mixing Desk". Large Language Models (LLMs) adaptively tune and reparameterize parallel attack agents in a real-time, closed-loop system that exploits image-specific semantic vulnerabilities. On standard benchmarks, ARMOR achieves improved cross-architecture transfer and reliably fools both settings, delivering a blended output for blind targets and selecting the best attack or blended attacks for white-box targets using a confidence-and-SSIM score.

Digital watermarking techniques are essential to prevent unauthorized use of images. Since pirated images are often geometrically distorted by operations such as scaling and cropping, accurate synchronization - detecting the embedding position of the watermark - is critical for proper extraction. In particular, cropping changes the origin of the image, making synchronization difficult. However, few existing methods are robust against cropping. To address this issue, we propose a watermarking method that estimates geometric transformations applied to a stego image using a pilot signal, allowing synchronization even after cropping. A grid-shaped pilot signal with distinct horizontal and vertical values is embedded in the image. When the image is transformed, the grid is also distorted. By analyzing this distortion, the transformation matrix can be estimated. Applying the Radon transform to the distorted image allows estimation of the grid angles and intervals. In addition, since the horizontal and vertical grid lines are encoded differently, the grid orientation can be determined, which reduces ambiguity. To validate our method, we performed simulations with anisotropic scaling, rotation, shearing, and cropping. The results show that the proposed method accurately estimates transformation matrices with low error under both single and composite attacks.

With natural language processing (NLP), researchers aim to enable computers to identify and understand patterns in human languages. This is often difficult because a language embeds many dynamic and varied properties in its syntax, pragmatics and phonology, which need to be captured and processed. The capacity of computers to process natural languages is increasing because NLP researchers are pushing its boundaries. But these research works focus more on well-resourced languages such as English, Japanese, German, French, Russian, Mandarin Chinese, etc. Over 95% of the world's 7000 languages are low-resourced for NLP, i.e. they have little or no data, tools, and techniques for NLP work. In this thesis, we present an overview of diacritic ambiguity and a review of previous diacritic disambiguation approaches on other languages. Focusing on the Igbo language, we report the steps taken to develop a flexible framework for generating datasets for diacritic restoration. Three main approaches, the standard n-gram model, the classification models and the embedding models were proposed. The standard n-gram models use a sequence of previous words to the target stripped word as key predictors of the correct variants. For the classification models, a window of words on both sides of the target stripped word was used. The embedding models compare the similarity scores of the combined context word embeddings and the embeddings of each of the candidate variant vectors.

Hierarchical text classification (HTC) depends on taxonomies that organize labels into structured hierarchies. However, many real-world taxonomies introduce ambiguities, such as identical leaf names under similar parent nodes, which prevent language models (LMs) from learning clear decision boundaries. In this paper, we present TaxMorph, a framework that uses large language models (LLMs) to transform entire taxonomies through operations such as renaming, merging, splitting, and reordering. Unlike prior work, our method revises the full hierarchy to better match the semantics encoded by LMs. Experiments across three HTC benchmarks show that LLM-refined taxonomies consistently outperform human-curated ones in various settings up to +2.9pp. in F1. To better understand these improvements, we compare how well LMs can assign leaf nodes to parent nodes and vice versa across human-curated and LLM-refined taxonomies. We find that human-curated taxonomies lead to more easily separable clusters in embedding space. However, the LLM-refined taxonomies align more closely with the model's actual confusion patterns during classification. In other words, even though they are harder to separate, they better reflect the model's inductive biases. These findings suggest that LLM-guided refinement creates taxonomies that are more compatible with how models learn, improving HTC performance.

Gaze prediction plays a critical role in Virtual Reality (VR) applications by reducing sensor-induced latency and enabling computationally demanding techniques such as foveated rendering, which rely on anticipating user attention. However, direct eye tracking is often unavailable due to hardware limitations or privacy concerns. To address this, we present a novel gaze prediction framework that combines Head-Mounted Display (HMD) motion signals with visual saliency cues derived from video frames. Our method employs UniSal, a lightweight saliency encoder, to extract visual features, which are then fused with HMD motion data and processed through a time-series prediction module. We evaluate two lightweight architectures, TSMixer and LSTM, for forecasting future gaze directions. Experiments on the EHTask dataset, along with deployment on commercial VR hardware, show that our approach consistently outperforms baselines such as Center-of-HMD and Mean Gaze. These results demonstrate the effectiveness of predictive gaze modeling in reducing perceptual lag and enhancing natural interaction in VR environments where direct eye tracking is constrained.

Medical vision-language models (VLMs) achieve strong performance in diagnostic reporting and image-text alignment, yet their underlying reasoning mechanisms remain fundamentally correlational, exhibiting reliance on superficial statistical associations that fail to capture the causal pathophysiological mechanisms central to clinical decision-making. This limitation makes them fragile, prone to hallucinations, and sensitive to dataset biases. Retrieval-augmented generation (RAG) offers a partial remedy by grounding predictions in external knowledge. However, conventional RAG depends on semantic similarity, introducing new spurious correlations. We propose Multimodal Causal Retrieval-Augmented Generation, a framework that integrates causal inference principles with multimodal retrieval. It retrieves clinically relevant exemplars and causal graphs from external sources, conditioning model reasoning on counterfactual and interventional evidence rather than correlations alone. Applied to radiology report generation, diagnosis prediction, and visual question answering, it improves factual accuracy, robustness to distribution shifts, and interpretability. Our results highlight causal retrieval as a scalable path toward medical VLMs that think beyond pattern matching, enabling trustworthy multimodal reasoning in high-stakes clinical settings.

Creative writing has long been considered a uniquely human endeavor, requiring voice and style that machines could not replicate. This assumption is challenged by Generative AI that can emulate thousands of author styles in seconds with negligible marginal labor. To understand this better, we conducted a behavioral experiment where 28 MFA writers (experts) competed against three LLMs in emulating 50 critically acclaimed authors. Based on blind pairwise comparisons by 28 expert judges and 131 lay judges, we find that experts preferred human writing in 82.7% of cases under the in-context prompting condition but this reversed to 62% preference for AI after fine-tuning on authors' complete works. Lay judges, however, consistently preferred AI writing. Debrief interviews with expert writers revealed that their preference for AI writing triggered an identity crisis, eroding aesthetic confidence and questioning what constitutes "good writing." These findings challenge discourse about AI's creative limitations and raise fundamental questions about the future of creative labor.

Recent advances in multimodal large language models (MLLMs) have demonstrated impressive performance on existing low-level vision benchmarks, which primarily focus on generic images. However, their capabilities to perceive and assess portrait images, a domain characterized by distinct structural and perceptual properties, remain largely underexplored. To this end, we introduce Q-Bench-Portrait, the first holistic benchmark specifically designed for portrait image quality perception, comprising 2,765 image-question-answer triplets and featuring (1) diverse portrait image sources, including natural, synthetic distortion, AI-generated, artistic, and computer graphics images; (2) comprehensive quality dimensions, covering technical distortions, AIGC-specific distortions, and aesthetics; and (3) a range of question formats, including single-choice, multiple-choice, true/false, and open-ended questions, at both global and local levels. Based on Q-Bench-Portrait, we evaluate 20 open-source and 5 closed-source MLLMs, revealing that although current models demonstrate some competence in portrait image perception, their performance remains limited and imprecise, with a clear gap relative to human judgments. We hope that the proposed benchmark will foster further research into enhancing the portrait image perception capabilities of both general-purpose and domain-specific MLLMs.

As financial institutions increasingly adopt machine learning for credit risk assessment, the persistence of algorithmic bias remains a critical barrier to equitable financial inclusion. This study provides a comprehensive audit of structural gender bias within the Taiwan Credit Default dataset, specifically challenging the prevailing doctrine of "fairness through blindness." Despite the removal of explicit protected attributes and the application of industry standard fairness interventions, our results demonstrate that gendered predictive signals remain deeply embedded within non-sensitive features. Utilizing SHAP (SHapley Additive exPlanations), we identify that variables such as Marital Status, Age, and Credit Limit function as potent proxies for gender, allowing models to maintain discriminatory pathways while appearing statistically fair. To mathematically quantify this leakage, we employ an adversarial inverse modeling framework. Our findings reveal that the protected gender attribute can be reconstructed from purely non-sensitive financial features with an ROC AUC score of 0.65, demonstrating that traditional fairness audits are insufficient for detecting implicit structural bias. These results advocate for a shift from surface-level statistical parity toward causal-aware modeling and structural accountability in financial AI.

Sound source localization (SSL) demonstrates remarkable results in controlled settings but struggles in real-world deployment due to dual imbalance challenges: intra-task imbalance arising from long-tailed direction-of-arrival (DoA) distributions, and inter-task imbalance induced by cross-task skews and overlaps. These often lead to catastrophic forgetting, significantly degrading the localization accuracy. To mitigate these issues, we propose a unified framework with two key innovations. Specifically, we design a GCC-PHAT-based data augmentation (GDA) method that leverages peak characteristics to alleviate intra-task distribution skews. We also propose an Analytic dynamic imbalance rectifier (ADIR) with task-adaption regularization, which enables analytic updates that adapt to inter-task dynamics. On the SSLR benchmark, our proposal achieves state-of-the-art (SoTA) results of 89.0% accuracy, 5.3° mean absolute error, and 1.6 backward transfer, demonstrating robustness to evolving imbalances without exemplar storage.

As LLMs are increasingly integrated into clinical workflows, their tendency for sycophancy, prioritizing user agreement over factual accuracy, poses significant risks to patient safety. While existing evaluations often rely on subjective datasets, we introduce a robust framework grounded in medical MCQA with verifiable ground truths. We propose the Adjusted Sycophancy Score, a novel metric that isolates alignment bias by accounting for stochastic model instability, or "confusability". Through an extensive scaling analysis of the Qwen-3 and Llama-3 families, we identify a clear scaling trajectory for resilience. Furthermore, we reveal a counter-intuitive vulnerability in reasoning-optimized "Thinking" models: while they demonstrate high vanilla accuracy, their internal reasoning traces frequently rationalize incorrect user suggestions under authoritative pressure. Our results across frontier models suggest that benchmark performance is not a proxy for clinical reliability, and that simplified reasoning structures may offer superior robustness against expert-driven sycophancy.

This study proposes an efficient Densely Swin Hybrid (EDSH) framework for brain tumor MRI analysis, designed to jointly capture fine grained texture patterns and long range contextual dependencies. Two tumor aware experimental setups are introduced to address class-specific diagnostic challenges. The first setup employs a Boosted Feature Space (BFS), where independently customized DenseNet and Swint branches learn complementary local and global representations that are dimension aligned, fused, and boosted, enabling highly sensitive detection of diffuse glioma patterns by successfully learning the features of irregular shape, poorly defined mass, and heterogeneous texture. The second setup adopts a hierarchical DenseNet Swint architecture with Deep Feature Extraction have Dual Residual connections (DFE and DR), in which DenseNet serves as a stem CNN for structured local feature learning, while Swin_t models global tumor morphology, effectively suppressing false negatives in meningioma and pituitary tumor classification by learning the features of well defined mass, location (outside brain) and enlargments in tumors (dural tail or upward extension). DenseNet is customized at the input level to match MRI spatial characteristics, leveraging dense residual connectivity to preserve texture information and mitigate vanishing-gradient effects. In parallel, Swint is tailored through task aligned patch embedding and shifted-window self attention to efficiently capture hierarchical global dependencies. Extensive evaluation on a large-scale MRI dataset (stringent 40,260 images across four tumor classes) demonstrates consistent superiority over standalone CNNs, Vision Transformers, and hybrids, achieving 98.50 accuracy and recall on the test unseen dataset.

We propose a drone signal out-of-distribution (OOD) detection algorithm based on discriminability-driven spatial-channel selection with a gradient norm. Time-frequency image features are adaptively weighted along both spatial and channel dimensions by quantifying inter-class similarity and variance based on protocol-specific time-frequency characteristics. Subsequently, a gradient-norm metric is introduced to measure perturbation sensitivity for capturing the inherent instability of OOD samples, which is then fused with energy-based scores for joint inference. Simulation results demonstrate that the proposed algorithm provides superior discriminative power and robust performance via SNR and various drone types.

We propose a drone signal out-of-distribution detection (OODD) algorithm based on the cognitive fusion of Zadoff-Chu (ZC) sequences and time-frequency images (TFI). ZC sequences are identified by analyzing the communication protocols of DJI drones, while TFI capture the time-frequency characteristics of drone signals with unknown or non-standard communication protocols. Both modalities are used jointly to enable OODD in the drone remote identification (RID) task. Specifically, ZC sequence features and TFI features are generated from the received radio frequency signals, which are then processed through dedicated feature extraction module to enhance and align them. The resultant multi-modal features undergo multi-modal feature interaction, single-modal feature fusion, and multi-modal feature fusion to produce features that integrate and complement information across modalities. Discrimination scores are computed from the fused features along both spatial and channel dimensions to capture time-frequency characteristic differences dictated by the communication protocols, and these scores will be transformed into adaptive attention weights. The weighted features are then passed through a Softmax function to produce the signal classification results. Simulation results demonstrate that the proposed algorithm outperforms existing algorithms and achieves 1.7% and 7.5% improvements in RID and OODD metrics, respectively. The proposed algorithm also performs strong robustness under varying flight conditions and across different drone types.

The vision-language-action (VLA) paradigm has enabled powerful robotic control by leveraging vision-language models, but its reliance on large-scale, high-quality robot data limits its generalization. Generative world models offer a promising alternative for general-purpose embodied AI, yet a critical gap remains between their pixel-level plans and physically executable actions. To this end, we propose the Tool-Centric Inverse Dynamics Model (TC-IDM). By focusing on the tool's imagined trajectory as synthesized by the world model, TC-IDM establishes a robust intermediate representation that bridges the gap between visual planning and physical control. TC-IDM extracts the tool's point cloud trajectories via segmentation and 3D motion estimation from generated videos. Considering diverse tool attributes, our architecture employs decoupled action heads to project these planned trajectories into 6-DoF end-effector motions and corresponding control signals. This plan-and-translate paradigm not only supports a wide range of end-effectors but also significantly improves viewpoint invariance. Furthermore, it exhibits strong generalization capabilities across long-horizon and out-of-distribution tasks, including interacting with deformable objects. In real-world evaluations, the world model with TC-IDM achieves an average success rate of 61.11 percent, with 77.7 percent on simple tasks and 38.46 percent on zero-shot deformable object tasks. It substantially outperforms end-to-end VLA-style baselines and other inverse dynamics models.

Real-world visualization tasks involve complex, multi-modal requirements that extend beyond simple text-to-chart generation, requiring reference images, code examples, and iterative refinement. Current systems exhibit fundamental limitations: single-modality input, one-shot generation, and rigid workflows. While LLM-based approaches show potential for these complex requirements, they introduce reliability challenges including catastrophic failures and infinite loop susceptibility. To address this gap, we propose MultiVis-Agent, a logic rule-enhanced multi-agent framework for reliable multi-modal and multi-scenario visualization generation. Our approach introduces a four-layer logic rule framework that provides mathematical guarantees for system reliability while maintaining flexibility. Unlike traditional rule-based systems, our logic rules are mathematical constraints that guide LLM reasoning rather than replacing it. We formalize the MultiVis task spanning four scenarios from basic generation to iterative refinement, and develop MultiVis-Bench, a benchmark with over 1,000 cases for multi-modal visualization evaluation. Extensive experiments demonstrate that our approach achieves 75.63% visualization score on challenging tasks, significantly outperforming baselines (57.54-62.79%), with task completion rates of 99.58% and code execution success rates of 94.56% (vs. 74.48% and 65.10% without logic rules), successfully addressing both complexity and reliability challenges in automated visualization generation.

We report the design, protocol, and outcomes of a student reproducibility hackathon focused on replicating the results of three influential MRI reconstruction papers: (a) MoDL, an unrolled model-based network with learned denoising; (b) HUMUS-Net, a hybrid unrolled multiscale CNN+Transformer architecture; and (c) an untrained, physics-regularized dynamic MRI method that uses a quantitative MR model for early stopping. We describe the setup of the hackathon and present reproduction outcomes alongside additional experiments, and we detail fundamental practices for building reproducible codebases.

As climate-related hazards intensify, conventional early warning systems (EWS) disseminate alerts rapidly but often fail to trigger timely protective actions, leading to preventable losses and inequities. We introduce Climate RADAR (Risk-Aware, Dynamic, and Action Recommendation system), a generative AI-based reliability layer that reframes disaster communication from alerts delivered to actions executed. It integrates meteorological, hydrological, vulnerability, and social data into a composite risk index and employs guardrail-embedded large language models (LLMs) to deliver personalized recommendations across citizen, volunteer, and municipal interfaces. Evaluation through simulations, user studies, and a municipal pilot shows improved outcomes, including higher protective action execution, reduced response latency, and increased usability and trust. By combining predictive analytics, behavioral science, and responsible AI, Climate RADAR advances people-centered, transparent, and equitable early warning systems, offering practical pathways toward compliance-ready disaster resilience infrastructures.

Quantization is an effective technique for reducing the storage footprint and computational costs of Large Language Models (LLMs), but it often results in performance degradation. Existing post-training quantization methods typically use small, English-only calibration sets; however, their impact on multilingual models remains underexplored. We systematically evaluate eight calibration settings (five single-language and three multilingual mixes) on two quantizers (GPTQ, AWQ) on data from 10 languages. Our findings reveal a consistent trend: non-English and multilingual calibration sets significantly improve perplexity compared to English-only baselines. Specifically, we observe notable average perplexity gains across both quantizers on Llama3.1 8B and Qwen2.5 7B, with multilingual mixes achieving the largest overall reductions of up to 3.52 points in perplexity. Furthermore, our analysis indicates that tailoring calibration sets to the evaluation language yields the largest improvements for individual languages, underscoring the importance of linguistic alignment. We also identify specific failure cases where certain language-quantizer combinations degrade performance, which we trace to differences in activation range distributions across languages. These results highlight that static one-size-fits-all calibration is suboptimal and that tailoring calibration data, both in language and diversity, plays a crucial role in robustly quantizing multilingual LLMs.

With the widespread adoption of Graphical User Interface (GUI) agents for automating GUI interaction tasks, substantial research focused on improving GUI perception to ground task instructions into concrete action steps. However, the step execution capability of these agents has gradually emerged as a new bottleneck for task completion. In particular, existing GUI agents often adopt overly simplified strategies for handling swipe interactions, preventing them from accurately replicating human-like behavior. To address this limitation, we decompose human swipe gestures into multiple quantifiable dimensions and propose an automated pipeline SwipeGen to synthesize human-like swipe interactions through GUI exploration. Based on this pipeline, we construct and release the first benchmark for evaluating the swipe execution capability of GUI agents. Furthermore, leveraging the synthesized data, we propose GUISwiper, a GUI agent with enhanced interaction execution capabilities. Experimental results demonstrate that GUISwiper achieves a swipe execution accuracy of 69.07%, representing a 214% improvement over existing VLM baselines.

This paper discusses the internal behavior of Transformer language models. Many recent pre-trained models have been reported to exhibit only slight changes in the angular distance between the input and output hidden state vectors in the middle Transformer layers, despite a disproportionately large ``jump'' in the angular distance occurring in or around the final Transformer layer. To characterize this, we first introduce a quantitative metric for the jump strength around the final layer, and then demonstrate its prevalence across many open-weight models, as well as its amplification throughout pre-training. Assuming such jumps indicate an undesirable property, we propose the jump-suppressing regularizer (JREG) which penalizes this jump during pre-training, thereby encouraging more balanced capability usage across the middle layers. Empirical evaluations of three model sizes of Llama-based models, trained with the proposed JREG method, reveal improved task performance compared to the baseline without altering the model architecture.

Quest2ROS2 is an open-source ROS2 framework for bi-manual teleoperation designed to scale robot data collection. Extending Quest2ROS, it overcomes workspace limitations via relative motion-based control, calculating robot movement from VR controller pose changes to enable intuitive, pose-independent operation. The framework integrates essential usability and safety features, including real-time RViz visualization, streamlined gripper control, and a pause-and-reset function for smooth transitions. We detail a modular architecture that supports "Side-by-Side" and "Mirror" control modes to optimize operator experience across diverse platforms. Code is available at: https://github.com/Taokt/Quest2ROS2.

Large language model (LLM)-based agents have been successfully deployed in many tool-augmented settings, but their scalability is fundamentally constrained by context length. Existing context-folding methods mitigate this issue by summarizing past interactions, yet they are typically designed for single-query or single-intent scenarios. In more realistic user-centric dialogues, we identify two major failure modes: (i) they irreversibly discard fine-grained constraints and intermediate facts that are crucial for later decisions, and (ii) their summaries fail to track evolving user intent, leading to omissions and erroneous actions. To address these limitations, we propose U-Fold, a dynamic context-folding framework tailored to user-centric tasks. U-Fold retains the full user--agent dialogue and tool-call history but, at each turn, uses two core components to produce an intent-aware, evolving dialogue summary and a compact, task-relevant tool log. Extensive experiments on $τ$-bench, $τ^2$-bench, VitaBench, and harder context-inflated settings show that U-Fold consistently outperforms ReAct (achieving a 71.4% win rate in long-context settings) and prior folding baselines (with improvements of up to 27.0%), particularly on long, noisy, multi-turn tasks. Our study demonstrates that U-Fold is a promising step toward transferring context-management techniques from single-query benchmarks to realistic user-centric applications.