The automatic understanding of video content is advancing rapidly. Empowered by deeper neural networks and large datasets, machines are increasingly capable of understanding what is concretely visible in video frames, whether it be objects, actions, events, or scenes. In comparison, humans retain a unique ability to also look beyond concrete entities and recognize abstract concepts like justice, freedom, and togetherness. Abstract concept recognition forms a crucial open challenge in video understanding, where reasoning on multiple semantic levels based on contextual information is key. In this paper, we argue that the recent advances in foundation models make for an ideal setting to address abstract concept understanding in videos. Automated understanding of high-level abstract concepts is imperative as it enables models to be more aligned with human reasoning and values. In this survey, we study different tasks and datasets used to understand abstract concepts in video content. We observe that, periodically and over a long period, researchers have attempted to solve these tasks, making the best use of the tools available at their disposal. We advocate that drawing on decades of community experience will help us shed light on this important open grand challenge and avoid ``re-inventing the wheel'' as we start revisiting it in the era of multi-modal foundation models.

In many planning applications, we might be interested in finding plans that minimally modify the initial state to achieve the goals. We refer to this concept as plan disruption. In this paper, we formally introduce it, and define various planning-based compilations that aim to jointly optimize both the sum of action costs and plan disruption. Experimental results in different benchmarks show that the reformulated task can be effectively solved in practice to generate plans that balance both objectives.

Combined Algorithm Selection and Hyperparameter Optimization (CASH) has been fundamental to traditional AutoML systems. However, with the advancements of pre-trained models, modern ML workflows go beyond hyperparameter optimization and often require fine-tuning, ensembling, and other adaptation techniques. While the core challenge of identifying the best-performing model for a downstream task remains, the increasing heterogeneity of ML pipelines demands novel AutoML approaches. This work extends the CASH framework to select and adapt modern ML pipelines. We propose PS-PFN to efficiently explore and exploit adapting ML pipelines by extending Posterior Sampling (PS) to the max k-armed bandit problem setup. PS-PFN leverages prior-data fitted networks (PFNs) to efficiently estimate the posterior distribution of the maximal value via in-context learning. We show how to extend this method to consider varying costs of pulling arms and to use different PFNs to model reward distributions individually per arm. Experimental results on one novel and two existing standard benchmark tasks demonstrate the superior performance of PS-PFN compared to other bandit and AutoML strategies. We make our code and data available at https://github.com/amirbalef/CASHPlus.

Multilingual text embeddings are often assumed to encode meaning in a perspective-independent semantic space, yielding stable similarity judgments across tasks and languages. Our results show that this assumption does not hold in practice. We introduce SEA-BED, a large-scale benchmark covering 10 Southeast Asian (SEA) languages and diverse embedding tasks, designed to systematically examine how embedding performance varies across tasks, languages, and language-task combinations. Across extensive evaluations, we observe that no single model performs uniformly well across SEA languages; task difficulty differs markedly within languages, and success on one task does not reliably generalize to others. Language-task analyses further reveal highly non-uniform performance landscapes, where performance varies across different language-task combinations. These findings call for closer attention to performance measurements that provide an expansive view across languages and tasks to uncover inconsistencies in semantic representation. Based on these observations, we provide insights for future model development, including data, algorithmic, and architectural considerations.

The effectiveness of unlabeled data in Semi/Self-Supervised Learning (SSL) depends on appropriate assumptions for specific scenarios, thereby enabling the selection of beneficial unsupervised pretext tasks. However, existing research has paid limited attention to assumptions in SSL, resulting in practical situations where the compatibility between the unsupervised pretext tasks and the target scenarios can only be assessed after training and validation. This paper centers on the assumptions underlying unsupervised pretext tasks and explores the feasibility of preemptively estimating the impact of unsupervised pretext tasks at low cost. Through rigorous derivation, we show that the impact of unsupervised pretext tasks on target performance depends on three factors: assumption learnability with respect to the model, assumption reliability with respect to the data, and assumption completeness with respect to the target. Building on this theory, we propose a low-cost estimation method that can quantitatively estimate the actual target performance. We build a benchmark of over one hundred pretext tasks and demonstrate that estimated performance strongly correlates with the actual performance obtained through large-scale training and validation.

Lifting-based 3D human pose estimation infers 3D joints from 2D keypoints but generalizes poorly because $(x,y)$ coordinates alone are an ill-posed, sparse representation that discards geometric information modern foundation models can recover. We propose \emph{AugLift}, which changes the representation format of lifting from 2D coordinates to a 6D geometric descriptor via two modules: (1) an \emph{Uncertainty-Aware Depth Descriptor} (UADD) -- a compact tuple $(c, d, d_{\min}, d_{\max})$ extracted from a confidence-scaled neighborhood of an off-the-shelf monocular depth map -- and (2) a scale normalization component that handles train/test distance shifts. AugLift requires no new sensors, no new data collection, and no architectural changes beyond widening the input layer; because it operates at the representation level, it is composable with any lifting architecture or domain generalization technique. In the detection setting, AugLift reduces cross-dataset MPJPE by $10.1$% on average across four datasets and four lifting architectures while improving in-distribution accuracy by $4.0$%; post-hoc analysis shows gains concentrate on novel poses and occluded joints. In the ground-truth 2D setting, combining AugLift with PoseAug's differentiable domain generalization achieves state-of-the-art cross-dataset performance ($62.4$\,mm on 3DHP, $92.6$\,mm on 3DPW; $14.5$% and $22.2$% over PoseAug), demonstrating that foundation-model depth provides genuine geometric signal complementary to explicit 3D augmentation. Code will be made publicly available.

Although artificial neural networks are often described as brain-inspired, their representations typically rely on continuous activations, such as the continuous latent variables in variational autoencoders (VAEs), which limits their biological plausibility compared to the discrete spike-based signaling in real neurons. Extensions like the Poisson VAE introduce discrete count-based latents, but their equal mean-variance assumption fails to capture overdispersion in neural spikes, leading to less expressive and informative representations. To address this, we propose NegBio-VAE, a negative-binomial latent-variable model with a dispersion parameter for flexible spike count modeling. NegBio-VAE preserves interpretability while improving representation quality and training feasibility via novel KL estimation and reparameterization. Experiments on four datasets demonstrate that NegBio-VAE consistently achieves superior reconstruction and generation performance compared to competing single-layer VAE baselines, and yields robust, informative latent representations for downstream tasks. Extensive ablation studies are performed to verify the model's robustness w.r.t. various components. Our code is available at https://github.com/co234/NegBio-VAE.

As humans move toward collaborating with coordinated robot teams, understanding how these teams coordinate and fail is essential for building trust and ensuring safety. However, exposing human collaborators to coordination failures during early-stage development is costly and risky, particularly in high-stakes domains such as healthcare. We adopt an agent-simulation approach in which all team roles, including the supervisory manager, are instantiated as LLM agents, allowing us to diagnose coordination failures before humans join the team. Using a controllable healthcare scenario, we conduct two studies with different hierarchical configurations to analyze coordination behaviors and failure patterns. Our findings reveal that team structure, rather than contextual knowledge or model capability, constitutes the primary bottleneck for coordination, and expose a tension between reasoning autonomy and system stability. By surfacing these failures in simulation, we prepare the groundwork for safe human integration. These findings inform the design of resilient robot teams with implications for process-level evaluation, transparent coordination protocols, and structured human integration. Supplementary materials, including codes, task agent setup, trace outputs, and annotated examples of coordination failures and reasoning behaviors, are available at: https://byc-sophie.github.io/mas-to-mars/.

Large language models (LLMs) are commonly adapted to downstream tasks using parameter-efficient fine-tuning (PEFT) or in-context learning (ICL). Recently, ICL-driven embedding-based adaptation has been proposed as a distinct task adaptation paradigm. It derives task-specific embeddings from intermediate activations using few-shot prompts and injects them during inference. Despite its conceptual appeal, this approach has not demonstrated consistent performance gains over PEFT or ICL, and its empirical advantages have been limited in practice. We propose Soft head-selection for ICL-derived Task Embeddings (SITE), a gradient-based method that identifies task-relevant attention heads to enable effective task embedding injection. Across various types of open-ended generation, reasoning, and natural language understanding tasks, SITE significantly outperforms prior embedding-based adaptation methods and few-shot ICL, while using substantially fewer trainable parameters than PEFT. Experiments on 12 LLMs ranging from 4B to 70B parameters demonstrate the generality of our approach, and intra-task and inter-task activation patching analyses further provide new mechanistic insights by revealing strong task dependence in attention head functionality.

Discrete diffusion models have recently emerged as strong alternatives to autoregressive language models, matching their performance through large-scale training. However, inference-time control remains relatively underexplored. In this work, we study how to steer generation toward desired rewards without retraining the models. Prior methods typically resample or filter within a single denoising trajectory, optimizing rewards step-by-step without trajectory-level refinement. We introduce particle Gibbs sampling for diffusion language models (PG-DLM), an inference-time algorithm enabling trajectory-level refinement. PG-DLM constructs a Markov chain over full denoising trajectories and applies a conditional sequential Monte Carlo kernel to resample them. By doing so, PG-DLM introduces a new scaling axis, the number of refinement iterations, which is unavailable to prior methods. Increasing iterations remains effective even as gains from adding more parallel samples saturate. Furthermore, PG-DLM enables adaptive compute allocation by performing additional iterations only when needed, leading to further efficiency gains. We derive theoretical guarantees for convergence and variance bounds, and analyze trade-offs across different scaling axes. Empirically, PG-DLM outperforms prior methods across compute budgets on reward-guided generation tasks. On GSM8K, it achieves 90.07% accuracy with 2.9 particles on average and 94.47% accuracy with 16 particles.

Ensuring consistent product quality in modern manufacturing is crucial, particularly in safety-critical applications. Conventional quality control approaches, reliant on manually defined thresholds and features, lack adaptability to the complexity and variability inherent in production data and necessitate extensive domain expertise. Conversely, data-driven methods, such as machine learning, demonstrate high detection performance but typically function as black-box models, thereby limiting their acceptance in industrial environments where interpretability is paramount. This paper introduces a methodology for industrial fault detection, which is both data-driven and transparent. The approach integrates a supervised machine learning model for multi-class fault classification, Shapley Additive Explanations for post-hoc interpretability, and a do-main-specific visualisation technique that maps model explanations to operator-interpretable features. Furthermore, the study proposes an evaluation methodology that assesses model explanations through quantitative perturbation analysis and evaluates visualisations by qualitative expert assessment. The approach was applied to the crimping process, a safety-critical joining technique, using a dataset of univariate, discrete time series. The system achieves a fault detection accuracy of 95.9 %, and both quantitative selectivity analysis and qualitative expert evaluations confirmed the relevance and inter-pretability of the generated explanations. This human-centric approach is designed to enhance trust and interpretability in data-driven fault detection, thereby contributing to applied system design in industrial quality control.

Multimodal sarcasm understanding is a high-order cognitive task. Although large language models (LLMs) have shown impressive performance on many downstream NLP tasks, growing evidence suggests that they struggle with sarcasm understanding. In this paper, we propose Commander-GPT, a modular decision routing framework inspired by military command theory. Rather than relying on a single LLM's capability, Commander-GPT orchestrates a team of specialized LLM agents where each agent will be selectively assigned to a focused sub-task such as keyword extraction, sentiment analysis, etc. Their outputs are then routed back to the commander, which integrates the information and performs the final sarcasm judgment. To coordinate these agents, we introduce three types of centralized commanders: (1) a trained lightweight encoder-based commander (e.g., multi-modal BERT); (2) four small autoregressive language models, serving as moderately capable commanders (e.g., DeepSeek-VL); (3) two large LLM-based commander (Gemini Pro and GPT-4o) that performs task routing, output aggregation, and sarcasm decision-making in a zero-shot fashion. We evaluate Commander-GPT on the MMSD and MMSD 2.0 benchmarks, comparing five prompting strategies. Experimental results show that our framework achieves 4.4% and 11.7% improvement in F1 score over state-of-the-art (SoTA) baselines on average, demonstrating its effectiveness.

Ultra-long generation by large language models (LLMs) is a widely demanded scenario, yet it remains a significant challenge due to their maximum generation length limit and overall quality degradation as sequence length increases. Previous approaches, exemplified by LongWriter, typically rely on ''teaching'', which involves supervised fine-tuning (SFT) on synthetic long-form outputs. However, this strategy heavily depends on synthetic SFT data, which is difficult and costly to construct, often lacks coherence and consistency, and tends to be overly artificial and structurally monotonous. In this work, we propose an incentivization-based approach that, starting entirely from scratch and without relying on any annotated or synthetic data, leverages reinforcement learning (RL) to foster the emergence of ultra-long, high-quality text generation capabilities in LLMs. We perform RL training starting from a base model, similar to R1-Zero, guiding it to engage in reasoning that facilitates planning and refinement during the writing process. To support this, we employ specialized reward models that steer the LLM towards improved length control, writing quality, and structural formatting. Experimental evaluations show that our LongWriter-Zero model, trained from Qwen2.5-32B, consistently outperforms traditional SFT methods on long-form writing tasks, achieving state-of-the-art results across all metrics on WritingBench and Arena-Write, and even surpassing 100B+ models such as DeepSeek R1 and Qwen3-235B. We open-source our data and model checkpoints under https://huggingface.co/THU-KEG/LongWriter-Zero-32B

Accurate real-time estimation of human movement dynamics, including internal joint moments and muscle forces, is essential for applications in clinical diagnostics and sports performance monitoring. Inertial measurement units (IMUs) provide a minimally intrusive solution for capturing motion data, particularly when used in sparse sensor configurations. However, current real-time methods rely on supervised learning, where a ground truth dataset needs to be measured with laboratory measurement systems, such as optical motion capture. These systems are known to introduce measurement and processing errors and often fail to generalize to real-world or previously unseen movements, necessitating new data collection efforts that are time-consuming and impractical. To overcome these limitations, we propose SSPINNpose, a self-supervised, physics-informed neural network that estimates joint kinematics and kinetics directly from IMU data, without requiring ground truth labels for training. We run the network output through a physics model of the human body to optimize physical plausibility and generate virtual measurement data. Using this virtual sensor data, the network is trained directly on the measured sensor data instead of a ground truth. When compared to optical motion capture, SSPINNpose is able to accurately estimate joint angles and joint moments at an RMSD of 8.7 deg and 4.9 BWBH%, respectively, for walking and running at speeds up to 4.9 m/s at a latency of 3.5 ms. Furthermore, the framework demonstrates robustness across sparse sensor configurations and can infer the anatomical locations of the sensors. These results underscore the potential of SSPINNpose as a scalable and adaptable solution for real-time biomechanical analysis in both laboratory and field environments.

Gesture recognition is an important research area in the field of computer vision. Most gesture recognition efforts focus on close-set scenarios, thereby limiting the capacity to effectively handle unseen or novel gestures. We aim to address class-incremental gesture recognition, which entails the ability to accommodate new and previously unseen gestures over time. Specifically, we introduce a Prototype-Guided Pseudo Feature Replay (PGPFR) framework for data-free class-incremental gesture recognition. This framework comprises four components: Pseudo Feature Generation with Batch Prototypes (PFGBP), Variational Prototype Replay (VPR) for old classes, Truncated Cross-Entropy (TCE) for new classes, and Continual Classifier Re-Training (CCRT). To tackle the issue of catastrophic forgetting, the PFGBP dynamically generates a diversity of pseudo features in an online manner, leveraging class prototypes of old classes along with batch class prototypes of new classes. Furthermore, the VPR enforces consistency between the classifier's weights and the prototypes of old classes, leveraging class prototypes and covariance matrices to enhance robustness and generalization capabilities. The TCE mitigates the impact of domain differences of the classifier caused by pseudo features. Finally, the CCRT training strategy is designed to prevent overfitting to new classes and ensure the stability of features extracted from old classes. Extensive experiments conducted on two widely used gesture recognition datasets, namely SHREC 2017 3D and EgoGesture 3D, demonstrate that our approach outperforms existing state-of-the-art methods by 11.8\% and 12.8\% in terms of mean global accuracy, respectively. The code is available on https://github.com/sunao-101/PGPFR-3/.

Visual recognition relies on understanding the semantics of image tokens and their complex interactions. Mainstream self-attention methods, while effective at modeling global pair-wise relations, fail to capture high-order associations inherent in real-world scenes and often suffer from redundant computation. Hypergraphs extend conventional graphs by modeling high-order interactions and offer a promising framework for addressing these limitations. However, existing hypergraph neural networks typically rely on static and hard hyperedge assignments, which lead to redundant hyperedges and overlooking the continuity of visual semantics. In this work, we present Soft Hypergraph Neural Networks (SoftHGNN), a lightweight plug-and-play hypergraph computation method for late-stage semantic reasoning in existing vision pipelines. Our SoftHGNN introduces the concept of soft hyperedges, where each vertex is associated with hyperedges via continuous and differentiable participation weights rather than hard binary assignments. These weights are produced by measuring similarities between vertex features and a small set of learnable hyperedge prototypes, yielding input-adaptive and semantically rich soft hyperedges. Using soft hyperedges as the medium for message aggregation and dissemination, SoftHGNN enriches feature representations with high-order contextual associations. To further enhance efficiency when scaling up the number of soft hyperedges, we incorporate a sparse hyperedge selection mechanism that activates only the top-k important hyperedges, along with a load-balancing regularizer to ensure adequate and balanced hyperedge utilization. Experimental results across three tasks on five datasets demonstrate that SoftHGNN efficiently captures high-order associations in visual scenes, achieving significant performance improvements. The code is available at: https://github.com/Mengqi-Lei/SoftHGNN.

Identifying controlled direct effects (CDEs) is crucial across numerous scientific domains. While existing methods can identify these effects from causal directed acyclic graphs (DAGs), the true DAG is often unknown in practice. Essential graphs, which represent a Markov equivalence class of DAGs characterized by the same set of conditional independencies, provide a more practical and realistic alternative, and the PC algorithm is one of the most widely used method to learn them using conditional independence tests. However, learning the full essential graph is computationally intensive and relies on strong, untestable assumptions. In this work, we adapt the PC algorithm to recover only the portion of the graph needed for identifying CDEs. In particular, we introduce the local essential graph (LEG), a graph structure defined relative to a target variable, and present LocPC, an algorithm that learns the LEG using solely local conditional independence tests. Building on this, we develop LocPC-CDE, which extracts precisely the portion of the LEG that is both necessary and sufficient for identifying a CDE. Compared to global methods, our algorithms require less conditional independence tests and operate under weaker assumptions while maintaining theoretical guarantees. We illustrate the effectiveness of our approach on synthetic and real data.

We present a method to dynamically deform 3D garments, in the form of a 3D polygon mesh, based on body shape, motion, and physical cloth material properties. Considering physical cloth properties allows to learn a physically grounded model, with the advantage of being more accurate in terms of physically inspired metrics such as strain or curvature. Existing work studies pose-dependent garment modeling to generate garment deformations from example data, and possibly data-driven dynamic cloth simulation to generate realistic garments in motion. We propose D-Garment, a learning-based approach trained on new data generated with a physics-based simulator. Compared to prior work, our 3D generative model learns garment deformations conditioned by physical material properties, which allows to model loose cloth geometry, especially for large deformations and dynamic wrinkles driven by body motion. Furthermore, the model can be efficiently fitted to observations captured using vision sensors such as 3D point clouds. We leverage the capability of diffusion models to learn flexible and powerful generative priors by modeling the 3D garment in a 2D parameter space independently from the mesh resolution. This representation allows to learn a template-specific latent diffusion model. This allows to condition global and local geometry with body and cloth material information. We quantitatively and qualitatively evaluate D-Garment on both simulations and data captured with a multi-view acquisition platform. Compared to recent baselines, our method is more realistic and accurate in terms of shape similarity and physical validity metrics. Code and data are available for research purposes at https://dumoulina.github.io/d-garment/

We demonstrate that large language models' (LLMs) mathematical reasoning is culturally sensitive: testing 14 models from Anthropic, OpenAI, Google, Meta, DeepSeek, Mistral, and Microsoft across six culturally adapted variants of the GSM8K benchmark, we find accuracy drops ranging from 0.3% (Claude 3.5 Sonnet) to 5.9% (LLaMA 3.1-8B) when math problems are embedded in unfamiliar cultural contexts--even when the underlying mathematical logic remains unchanged. These statistically significant performance reductions (p < 0.01, confirmed through McNemar tests) reveal that mathematical reasoning in LLMs is not culturally neutral. To create these variants for Haiti, Moldova, Pakistan, Solomon Islands, Somalia, and Suriname, we systematically replaced cultural entities (names, foods, places, etc.) in 1,198 GSM8K questions while preserving all mathematical operations and numerical values. Our quantitative error analysis of 18,887 instances reveals that cultural adaptation affects broader reasoning patterns, with mathematical reasoning errors comprising 54.7% and calculation errors 34.5% of failures. Interestingly, cultural familiarity can enhance performance: Mistral Saba outperforms some larger models on Pakistan-adapted problems due to Middle Eastern and South Asian training data exposure. This study underscores the need for more diverse training data to ensure robust LLM performance across global contexts.

The Segment Anything Model (SAM) has revolutionized image and video segmentation with its powerful zero-shot capabilities. However, its massive parameter scale and high computational demands hinder efficient deployment on resource-constrained edge devices. While Post-Training Quantization (PTQ) offers a practical solution, existing methods still fail to handle four critical quantization challenges: (1) ill-conditioned weights; (2) skewed and long-tailed post-GELU activations; (3) pronounced inter-channel variance in linear projections; and (4) exponentially scaled and heterogeneous attention scores. To mitigate these bottlenecks, we propose AHCQ-SAM, an accurate and hardware-compatible PTQ framework featuring four synergistic components: (1) Activation-aware Condition Number Reduction (ACNR), which regularizes weight matrices via a proximal point algorithm to suppress ill-conditioning; (2) Hybrid Log-Uniform Quantization (HLUQ), which combines power-of-two and uniform quantizers to capture skewed post-GELU activations; (3) Channel-Aware Grouping (CAG), which clusters channels with homogeneous statistics to achieve high accuracy with minimal hardware overhead; and (4) Logarithmic Nonlinear Quantization (LNQ), which utilizes logarithmic transformations to adaptively adjust quantization resolution for exponential and heterogeneous attention scores. Experimental results demonstrate that AHCQ-SAM outperforms current methods on SAM. Compared with the SOTA method, it achieves a 15.2% improvement in mAP for 4-bit SAM-B with Faster R-CNN on the COCO dataset. Furthermore, we establish a PTQ benchmark for SAM2, where AHCQ-SAM yields a 14.01% improvement in J&F for 4-bit SAM2-Tiny on the SA-V Test dataset. Finally, FPGA-based implementation validates the practical utility of AHCQ-SAM, delivering a 7.12x speedup and a 6.62x power efficiency improvement over the floating-point baseline.

Path regularization has shown to be a very effective regularization to train neural networks, leading to a better generalization property than common regularizations i.e. weight decay, etc. We propose a first near-complete (as will be made explicit in the main text) nonasymptotic generalization theory for multilayer neural networks with path regularizations for general learning problems. In particular, it does not require the boundedness of the loss function, as is commonly assumed in the literature. Our theory goes beyond the bias-variance tradeoff and aligns with phenomena typically encountered in deep learning. It is therefore sharply different from other existing nonasymptotic generalization error bounds. More explicitly, we propose an explicit generalization error upper bound for multilayer neural networks with $σ(0)=0$ and sufficiently broad Lipschitz loss functions, without requiring the width, depth, or other hyperparameters of the neural network to approach infinity, a specific neural network architecture (e.g., sparsity, boundedness of some norms), a particular optimization algorithm, or boundedness of the loss function, while also taking approximation error into consideration. A key feature of our theory is that it also considers approximation errors. In particular, we solve an open problem proposed by Weinan E et. al. regarding the approximation rates in generalized Barron spaces. Furthermore, we show the near-minimax optimality of our theory for regression problems with ReLU activations. Notably, our upper bound exhibits the famous double descent phenomenon for such networks, which is the most distinguished characteristic compared with other existing results. We argue that it is highly possible that our theory reveals the true underlying mechanism of the double descent phenomenon.

Model reuse offers a solution to the challenges of segmentation in biomedical imaging, where high data annotation costs remain a major bottleneck for deep learning. However, although many pretrained models are released through challenges, model zoos, and repositories, selecting the most suitable model for a new dataset remains difficult due to the lack of reliable model ranking methods. We introduce the first black-box-compatible framework for unsupervised and source-free ranking of semantic and instance segmentation models based on the consistency of predictions under perturbations. While ranking methods have been studied for classification and a few segmentation-related approaches exist, most target related tasks such as transferability estimation or model validation and typically rely on labelled data, feature-space access, or specific training assumptions. In contrast, our method directly addresses the repository setting and applies to both semantic and instance segmentation, for zero-shot reuse or after unsupervised domain adaptation. We evaluate the approach across a wide range of biomedical segmentation tasks in both 2D and 3D imaging, showing that our estimated rankings strongly correlate with true target-domain model performance rankings.

The effectiveness of instruction fine-tuning for Large Language Models is fundamentally constrained by the quality and efficiency of training datasets. This work introduces Low-Confidence Gold (LCG), a novel filtering framework that employs centroid-based clustering and confidence-guided selection for identifying valuable instruction pairs. Through a semi-supervised approach using a lightweight classifier trained on representative samples, LCG curates high-quality subsets while preserving data diversity. Experimental evaluation demonstrates that models fine-tuned on LCG-filtered subsets of 6K samples achieve superior performance compared to existing methods, with substantial improvements on MT-bench and consistent gains across comprehensive evaluation metrics. The framework's efficacy while maintaining model performance establishes a promising direction for efficient instruction tuning.All open-source assets are publicly available at https://github.com/Lizruletheworld/Low-Confidence_Gold.

As Large Language Models (LLMs) can now process extremely long contexts, efficient inference over these extended inputs has become increasingly important, especially for emerging applications like LLM agents that highly depend on this capability. Speculative decoding (SD) offers a promising lossless acceleration technique compared to lossy alternatives such as quantization and model cascades. However, most state-of-the-art SD methods are trained on short texts (typically fewer than 4k tokens), making them unsuitable for long-context scenarios. Specifically, adapting these methods to long contexts presents three key challenges: (1) the excessive memory demands posed by draft models due to large Key-Value (KV) cache; (2) performance degradation resulting from the mismatch between short-context training and long-context inference; and (3) inefficiencies in tree attention mechanisms when managing long token sequences. This work introduces LongSpec, a framework that addresses these challenges through three core innovations: a memory-efficient draft model with a constant-sized KV cache; novel position indices that mitigate the training-inference mismatch; and an attention aggregation strategy that combines fast prefix computation with standard tree attention to enable efficient decoding. Experimental results confirm the effectiveness of LongSpec, achieving up to a 3.26x speedup over strong Flash Attention baselines across five long-context understanding datasets, as well as a 2.25x reduction in wall-clock time on the AIME24 long reasoning task with the QwQ model, demonstrating significant latency improvements for long-context applications. The code is available at https://github.com/sail-sg/LongSpec.

Agents for computer use (ACUs) are an emerging class of systems capable of executing complex tasks on digital devices -- such as desktops, mobile phones, and web platforms -- given instructions in natural language. These agents can automate tasks by controlling software via low-level actions like mouse clicks and touchscreen gestures. However, despite rapid progress, ACUs are not yet mature for everyday use. In this survey, we investigate the state-of-the-art, trends, and research gaps in the development of practical ACUs. We provide a comprehensive review of the ACU landscape, introducing a unifying taxonomy spanning three dimensions: (I) the domain perspective, characterizing agent operating contexts; (II) the interaction perspective, describing observation modalities (e.g., screenshots, HTML) and action modalities (e.g., mouse, keyboard, code execution); and (III) the agent perspective, detailing how agents perceive, reason, and learn. We review 87 ACUs and 33 datasets across foundation model-based and classical approaches through this taxonomy. Our analysis identifies six major research gaps: insufficient generalization, inefficient learning, limited planning, low task complexity in benchmarks, non-standardized evaluation, and a disconnect between research and practical conditions. To address these gaps, we advocate for: (a) vision-based observations and low-level control to enhance generalization; (b) adaptive learning beyond static prompting; (c) effective planning and reasoning methods and models; (d) benchmarks that reflect real-world task complexity; (e) standardized evaluation based on task success; (f) aligning agent design with real-world deployment constraints. Together, our taxonomy and analysis establish a foundation for advancing ACU research toward general-purpose agents for robust and scalable computer use.

Influenced by the complexity of volumetric imaging, there is a shortage of established datasets useful for benchmarking volumetric deep-learning models. As a consequence, new and existing models are not easily comparable, limiting the development of architectures optimized specifically for volumetric data. To counteract this trend, we introduce MozzaVID -- a large, clean, and versatile volumetric classification dataset. Our dataset contains X-ray computed tomography (CT) images of mozzarella microstructure and enables the classification of 25 cheese types and 149 cheese samples. We provide data in three different resolutions, resulting in three dataset instances containing from 591 to 37,824 images. While targeted for developing general-purpose volumetric algorithms, the dataset also facilitates investigating the properties of mozzarella microstructure. The complex and disordered nature of food structures brings a unique challenge, where a choice of appropriate imaging method, scale, and sample size is not trivial. With this dataset, we aim to address these complexities, contributing to more robust structural analysis models and a deeper understanding of food structure. The dataset can be explored through: https://papieta.github.io/MozzaVID/

While text-to-video diffusion models have advanced significantly, creating coherent long-form content remains unreliable due to stochastic sampling artifacts. This necessitates generating multiple candidates, yet verifying them creates a severe bottleneck; manual review is unscalable, and existing automated metrics lack the adaptability and speed required for runtime monitoring. Another critical issue is the trade-off between evaluation quality and run-time performance: metrics that best capture human-like judgment are often too slow to support iterative generation. These challenges, originating from the lack of an effective evaluation, motivate our work toward a novel solution. To address this, we propose a scalable automated verification framework for long-form video. First, we introduce the MSG(Multi-Scene Generation) score, a hierarchical attention-based metric that adaptively evaluates narrative and visual consistency. This serves as the core verifier within our CGS (Candidate Generation and Selection) framework, which automatically identifies and filters high-quality outputs. Furthermore, we introduce Implicit Insight Distillation (IID) to resolve the trade-off between evaluation reliability and inference speed, distilling complex metric insights into a lightweight student model. Our approach offers the first comprehensive solution for reliable and scalable long-form video production.

Despite significant advancements, segmentation based on deep neural networks in medical and surgical imaging faces several challenges, two of which we aim to address in this work. First, acquiring complete pixel-level segmentation labels for medical images is time-consuming and requires domain expertise. Second, typical segmentation pipelines cannot detect out-of-distribution (OOD) pixels, leaving them prone to spurious outputs during deployment. In this work, we propose a novel segmentation approach which broadly falls within the positive-unlabelled (PU) learning paradigm and exploits tools from OOD detection techniques. Our framework learns only from sparsely annotated pixels from multiple positive-only classes and does not use any annotation for the background class. These multi-class positive annotations naturally fall within the in-distribution (ID) set. Unlabelled pixels may contain positive classes but also negative ones, including what is typically referred to as \emph{background} in standard segmentation formulations. To the best of our knowledge, this work is the first to formulate multi-class segmentation with sparse positive-only annotations as a pixel-wise PU learning problem and to address it using OOD detection techniques. Here, we forgo the need for background annotation and consider these together with any other unseen classes as part of the OOD set. Our framework can integrate, at a pixel-level, any OOD detection approaches designed for classification tasks. To address the lack of existing OOD datasets and established evaluation metric for medical image segmentation, we propose a cross-validation strategy that treats held-out labelled classes as OOD. Extensive experiments on both multi-class hyperspectral and RGB surgical imaging datasets demonstrate the robustness and generalisation capability of our proposed framework.

In reinforcement learning (RL), temporal difference (TD) error is known to be related to the firing rate of dopamine neurons. It has been observed that each dopamine neuron does not behave uniformly, but each responds to the TD error in an optimistic or pessimistic manner, interpreted as a kind of distributional RL. To explain such a biological data, a heuristic model has also been introduced with learning rates asymmetric for the positive and negative TD errors. However, this heuristic model is not theoretically-grounded and unknown whether it can work as a RL algorithm. This paper therefore introduces a novel theoretically-grounded model with optimism and pessimism, which is derived from control as inference. In combination with ensemble learning, a distributional value function as a critic is estimated from regularly introduced optimism and pessimism. Based on its central value, a policy in an actor is improved. This proposed algorithm, so-called DROP (distributional and regular optimism and pessimism), is compared on dynamic tasks. Although the heuristic model showed poor learning performance, DROP demonstrated excellent performance in all tasks with high generality. In addition, DROP achieved learning performance comparable to the state-of-the-art algorithms. In other words, it was suggested that DROP is a new model that can elicit the potential contributions of optimism and pessimism.

In this paper, we build a reinforcement learning framework to study how children compose numbers using base-ten blocks. Studying numerical cognition in toddlers offers a powerful window into the learning process itself, because numbers sit at the intersection of language, logic, perception, and culture. Specifically, we utilize state of the art (SOTA) reinforcement learning algorithms and neural network architectures to understand how variations in linguistic instructions can affect the learning process. Our results also show that instructions providing explicit action guidance are a more effective learning signal for RL agents to construct numbers. Furthermore, we identify an effective curriculum for ordering numerical-composition examples during training, resulting in faster convergence and improved generalization to unseen data. These findings highlight the role of language and multi-modal signals in numerical cognition and provide hypotheses for designing effective instructional strategies for early childhood education.