Search research

Online comments play a crucial role in shaping public sentiment and opinion dynamics on social media. However, evaluating their popularity remains challenging, not only because it depends on linguistic quality, originality, and emotional resonance, but also because stylistic preferences vary widely across platforms and user groups, causing the same comment to resonate differently in different communities. In this work, we present HotComment, a multimodal benchmark integrating video and text modalities that comprehensively quantifies popularity from three enhanced aspects: (1) Content Quality, which evaluates semantic similarity with ground-truth human comments and extends quality assessment through four interpretable dimensions; (2) Popularity Prediction, based on trends from models trained on real-world interaction data; and (3) User Behavior Simulation, which models the distribution of platform users and approximates \textbf{engagement scores} through an agent-based framework. Furthermore, we propose StyleCmt, inspired by social ripple effects, where multiple stylistic dimensions align to amplify socially resonant expressions and suppress incongruent ones.

We introduce DualFact, a dual-layer, multimodal factuality evaluation framework for procedural video captioning. DualFact separates factual correctness into conceptual facts, capturing abstract semantic roles (e.g., Action, Ingredient, Tool, Location), and contextual facts, capturing their grounded predicate-argument realizations in video. To support complete and role-consistent evaluation, DualFact incorporates implicit argument augmentation (VIA) and contrastive fact sets. We instantiate DualFact in two modes: DualFact-T, which verifies facts against textual evidence, and DualFact-V, which verifies facts against video-grounded visual evidence. Experiments on YouCook3-Fact and CraftBench-Fact show that state-of-the-art multimodal language models produce fluent but often factually incomplete captions, with systematic omissions and role-level inconsistencies. DualFact correlates more strongly with human factuality judgments than standard metrics, particularly for contextual facts, and reveals that caption-only evaluation overestimates hallucinations compared to video-grounded verification. Overall, DualFact offers an interpretable and human-aligned evaluation protocol that highlights persistent challenges in multimodal factual grounding, extending beyond surface-level fluency.

We present LatentDiff, a scalable framework for semantic dataset comparison that operates directly in the latent space of pretrained vision encoders. By combining sparse autoencoder-based divergence testing with density ratio estimation, LatentDiff identifies interpretable semantic differences between datasets at a fraction of the computational cost of caption-based alternatives. We also introduce Noisy-Diff, a benchmark capturing realistic sparse distribution shifts that cause existing methods to struggle. Experiments demonstrate that LatentDiff achieves superior accuracy while remaining robust to settings where an extremely small fraction of images (from 5% to <1% ) differ semantically.

Key-value (KV) caching is essential for large language model inference, yet its memory overhead poses a critical bottleneck for long-context generation. Existing eviction policies predominantly rely on empirical heuristics, lacking a rigorous theoretical foundation. This work rethinks KV cache eviction through the lens of the Information Bottleneck principle. Under a linear-Gaussian surrogate of attention, we derive a closed-form mutual information objective that characterizes the effective information capacity of a retained KV cache subset. This formulation reveals that a wide range of existing eviction strategies can be interpreted as different approximations of the same capacity-maximization principle. Guided by this insight, we introduce CapKV, a capacity-aware eviction method that directly targets information preservation via a log-determinant approximation using statistical leverage scores. This approach replaces heuristic selection with a theoretically grounded mechanism that preserves the maximum predictive signal. Extensive experiments across multiple models and long-context benchmarks show that CapKV consistently outperforms prior methods, achieving a better trade-off between memory efficiency and generational fidelity.

(Gen)AI diversity is not merely an ethical issue. From the perspective of geographic information science (GIScience), it could be interpreted as a function of uncertainty and as a form of cognitive bias, embedded in AI outputs. Recent work has sought to develop information-theoretic diversity measures and apply them to evaluate AI-chatbot outputs in a geographic context. As the AI ecosystem to which we are exposed on a daily basis becomes rapidly multimodal, we believe it is important to examine geographic diversity across various modalities. Focusing on images, this paper aims to fill this research gap. First, we select the GPT and DALL-E models as state-of-the-art examples and point out how assessing their geographic diversity involves various stages, including prompt revision and image generation. Then, taking inspiration from species diversity measures in ecological research, we incorporate similarity weighting into the measurement of geographic diversity. Next, we demonstrate how to evaluate geographic diversity in image generation through a case study. Our analysis reveals several counterintuitive findings. For instance, older models can exhibit greater geographic diversity despite producing lower-quality images, and prompt revision yields greater geographic diversity than image generation. At the same time, we observe explicit model homogeneity underlying the lack of geographic diversity, as the selected models consistently depict the same prototypical geo-specific feature or similar features. This is concerning, as it risks producing stereotypical representations of places.

In-Context Learning (ICL) allows LLMs to adapt to new tasks from a few demonstrations, but its reliability remains a concern: predictions are highly sensitive to both prompt design and the model's ability to understand the context, obscuring whether failures arise from data properties or model limitations. Uncertainty decomposition-separating aleatoric from epistemic sources-is particularly crucial in this setting, yet existing methods, designed for standard generation tasks, fail to capture the unique dynamics of ICL. To address this, we introduce a concept of self-function vectors, built upon Bayesian views and the mechanistic interpretability of ICL. These vectors leverage internal model representations to model the latent concept learned during in-context prompting, thereby enabling a direct estimation of aleatoric uncertainty within a Bayesian framework and circumventing the reliance on brittle input or decoding manipulations. Given the lack of established benchmarks and suitable evaluation protocols, we also propose the first and rigorous evaluation protocol, in which data is manipulated in controlled ways so as to quantify aleatoric uncertainty precisely and separately from epistemic uncertainty. With this new evaluation framework, initially grounded in synthetic tasks for conceptual development and subsequently extended to real-world datasets, we show that our proposed methodology can measure uncertainty of LLM predictions made under ICL more reliably than existing alternative methods. Moreover, we show it can be used as a practical tool for trustworthy-related applications, such as hallucination detection. Our findings pave a new direction for connecting the quantitative view of uncertainty with the mechanistic understanding of model behavior.

Do large language models (LLMs) truly acquire embodied cognition and cultural conventions from text? We introduce demonstratives, fundamental spatial expressions like "this/that" in English and "zhè/nà" in Chinese, as a novel probe for grounded knowledge. Using 6,400 responses from 320 native speakers, we establish a human baseline: English speakers reliably distinguish proximal-distal referents but struggle with perspective-taking, while Chinese speakers switch perspectives fluently but tolerate distal ambiguity. In contrast, five state-of-the-art LLMs fail to inherently understand the proximal-distal contrast and show no cultural differences, defaulting to English-centric reasoning. Our study contributes (i) a new task, based on demonstratives, as a new lens for evaluating embodied cognition and cultural conventions; (ii) empirical evidence of cross-cultural asymmetries in human interpretation; (iii) a new perspective on the egocentric-sociocentric debate, showing both orientations coexist but vary across languages; and (iv) a call to address individual variation in future model design.

Tree ensembles provide strong classification performance but usually behave as black-box models. Post-hoc interpretability techniques such as RuleCOSI+ extract a small ruleset that approximates the ensemble, but this simplification can leave the probabilities attached to the extracted rules unreliable. In particular, RuleCOSI+ assigns empirical class probabilities to the extracted rules and repeatedly uses those rule statistics during its greedy combination and simplification procedure. We present RCProb, a probabilistic extension that uses smoothed atomic class-conditional evidence for the expensive search stages and a support-adaptive mixture with an ensemble-informed m-estimate for the final rule probabilities. The method is evaluated on 18 binary and 5 multiclass datasets using random forest (RF) and gradient boosting machine (GBM) ensembles. Relative to RuleCOSI+, the median paired log-loss reduction is 71.9% for RF and 62.5% for GBM, with both differences remaining significant after Holm correction. The number of rules decreases by 38.7% for RF and 38.5% for GBM, while the primary tests do not detect a macro-F1 difference. Confidence-ECE also decreases for both ensembles, with statistical support for RF after correction. A separate controlled experiment with dedicated calibration data shows that native RCProb probabilities are competitive with RuleCOSI+ followed by temperature scaling, although additional post-hoc calibration can still improve RCProb. The results show that probability estimation is an important part of rule extraction and not only a post-processing step.

Viruses represent the most abundant biological entities on Earth and play a pivotal role in microbial ecosystems, yet, as prominent human pathogens, they are closely linked to human morbidity and mortality. Accurate identification of viral sequences from viral genome sequences is therefore essential, but existing genome-based classification models that largely relying on composition- or frequency-based subsequence features often suffer from limited interpretability and reduced accuracy, particularly on complex or imbalanced datasets. To address these limitations, we propose GeneNSPCla (Genomic Negative Sequential Pattern-based Classification), a novel viral classification framework based on Negative Sequential Patterns (NSPs) that extracts discriminative absence-based features from nucleotide sequences of RNA viral genomes. By transforming these NSPs into numerical feature vectors and integrating them into multiple supervised classifiers, GeneNSPCla effectively captures both presence and absence signals in viral sequences. Furthermore, we propose a negative pattern mining algorithm adapted for processing genomic data: GONPM+, which can discover longer and more biologically meaningful negative sequential patterns. The experimental results demonstrate that the average accuracy of GONPM+ in 8 classifiers has improved by 10.03% compared to the original negative pattern mining algorithm and by 24.75% compared to the positive pattern mining algorithm. These findings highlight the effectiveness of incorporating absence-based sequential information, providing a new and complementary perspective for viral genome analysis and classification.

Despite significant progress in Unified Multimodal Retrieval (UMR) powered by Large Multimodal Models (LMMs), existing embedding methods primarily focus on sample-level objectives via contrastive learning while overlooking the crucial subject-level semantics. This limitation hinders the model's ability to group semantically coherent subjects in complex multimodal queries, manifesting as semantic alignment deviation--where models fail to accurately localize salient text-referred regions in visual content. Moreover, without explicit guidance to model salient visual subjects, LMMs tend to over-rely on textual cues, resulting in visual modality neglect and suboptimal utilization of visual knowledge. To this end, we propose Salient Subject-Aware Multimodal Embedding (SSA-ME), a novel framework designed to enhance fine-grained representation learning through saliency-aware modeling. SSA-ME leverages LMMs and visual experts to identify and emphasize salient visual concepts in image-text pairs, and introduces a saliency-guided objective to better align cross-modal attention with semantically meaningful regions. Additionally, a feature regeneration module recalibrates visual features based on the derived saliency maps, ensuring a balanced and semantically coherent integration across modalities. Extensive experiments show that our method achieves state-of-the-art performance on the MMEB benchmark, demonstrating that incorporating subject-level modeling substantially improves multimodal retrieval. Comprehensive qualitative analyses further illustrate the interpretability and effectiveness of our approach.

Categorical structural optimization under aleatoric uncertainty is challenging because each design variable must be selected from a finite catalog of admissible instances, while each candidate design may require expensive stochastic finite-element evaluations. Existing latent-space optimization strategies can reduce the dimensionality of catalog attributes, but they often treat the reduced space as a continuous search domain. The resulting continuous optimum must then be rounded off to a nearby catalog instance, which may alter the objective value, constraint status, or physical interpretation of the design. To address this issue, this paper proposes the \textbf{C}ategorical \textbf{O}ptimization with \textbf{B}ayesian \textbf{A}nchored \textbf{L}atent \textbf{T}rust Regions (\textbf{COBALT}) framework for high-dimensional categorical Optimization Under Uncertainty. COBALT first embeds the physical catalog into a low-dimensional latent representation and locks the mapped instances as a discrete anchored graph. A data-independent random tree decomposition is then used to provide bounded-complexity additive modeling over high-dimensional categorical variables. On this anchored domain, an additive SAAS-GP surrogate is fitted to heteroscedastic MC-FEA observations, and a trust-region discrete graph acquisition search selects the next admissible catalog configuration without continuous relaxation or rounding-off. The proposed strategy is applied to robust design optimization of complex bar structures, considering structural weight, strain energy, and local buckling performance. By evaluating only valid catalog designs through the MC-FEA oracle, COBALT preserves physical admissibility throughout the active learning loop and improves the efficiency of robust categorical structural optimization.

We present four conceptual value-sensitive AI systems to examine how the presence of AI could influence praying experiences. Drawing on key values and practices associated with praying identified through a diary study, we designed AI systems intended to "assist" prayer practices. These designs were presented to participants through speculative design workbooks, serving as provocations to co-reflect on how the intervention of AI systems might shape their praying experiences. Our findings suggest that a sense of authenticity (or feeling a genuine connection to the divine) is a crucial value, while the presence of AI was often perceived as diminishing this authenticity, particularly when AI assumed too much agency in guiding praying practices. Based on our findings, we argue that AI system designs for deeply value-laden experiences should preserve users' agency in shaping their own experiences by maintaining interpretive openness, perhaps by leveraging AI's inexplicability as a resource for personal meaning-making or by recognizing non-use of AI as a legitimate design choice.

Embodied agents in safety-critical applications such as Vision-Language Navigation (VLN) rely on multiple interdependent capabilities (e.g., perception, memory, planning, decision), making failures difficult to localize and attribute. Existing testing methods are largely system-level and provide limited insight into which capability deficiencies cause task failures. We propose a capability-oriented testing approach that enables failure detection and attribution by combining (1) adaptive test case generation via seed selection and mutation, (2) capability oracles for identifying capability-specific errors, and (3) a feedback mechanism that attributes failures to capabilities and guides further test generation. Experiments show that our method discovers more failure cases and more accurately pinpoints capability-level deficiencies than state-of-the-art baselines, providing more interpretable and actionable guidance for improving embodied agents.

LLMs deployed for natural-language querying of analytical databases suffer from two intertwined failures - incorrect answers and confident hallucinations - both rooted in the same cause: the model is forced to infer business semantics that the schema does not encode. We test whether supplying those semantics as context closes the gap. We benchmark three frontier LLMs (Claude Opus 4.7, Claude Sonnet 4.6, GPT-5.4) on 100 natural-language questions over the Cleaned Contoso Retail Dataset in ClickHouse, using a paired single-shot protocol. Each model is evaluated twice: once given only the warehouse schema, and once given the schema plus a 4 KB hand-authored markdown document describing the dataset's measures, conventions, and disambiguation rules. Adding the document improves accuracy by +17 to +23 percentage points across all three models. With it, the three models are statistically indistinguishable (67.7-68.7%); without it, they are also indistinguishable (45.5-50.5%). Every cross-cluster comparison is significant at p < 0.01. The presence of the semantic-layer document accounts for essentially all of the significant variance; model choice within tier does not. We interpret this as a structural result: explicit business semantics suppress the dominant class of text-to-SQL errors not by making the model more capable, but by changing what the model is being asked to do.

Evaluating long document summaries remains the primary bottleneck in summarization research. Existing metrics correlate weakly with human judgments and produce aggregate scores without explaining deficiencies or guiding improvement, preventing effective refinement in applications requiring verifiable accuracy. We introduce LongSumEval, a unified framework bridging evaluation and generation through structured question-answering feedback. The framework operationalizes summary quality as answerability and factual alignment of question-answer pairs, generating interpretable scores and actionable feedback that identifies coverage gaps and factual inconsistencies. This resolves the misalignment where evaluation operates independently of generation objectives. Meta-evaluation of our QA-based evaluation module across seven benchmarks demonstrates substantially stronger agreement with human judgments compared to established metrics. Structured feedback enables significant quality improvements through self-refinement without retraining. By demonstrating that evaluation feedback can serve as executable instructions for generation, this work establishes a generalizable paradigm for aligning assessment with improvement, with direct implications for controllable text generation requiring verifiable accuracy and transparent quality control. All code and datasets will be released in GitHub for reproducibility.

Typographic prompt injection exploits vision language models' (VLMs) ability to read text rendered in images, posing a growing threat as VLMs power autonomous agents. Prior work typically focus on maximizing attack success rate (ASR) but does not explain \emph{why} certain renderings bypass safety alignment. We make two contributions. First, an empirical study across four VLMs including GPT-4o and Claude, twelve font sizes, and ten transformations reveals that multimodal embedding distance strongly predicts ASR ($r{=}{-}0.71$ to ${-}0.93$, $p{<}0.01$), providing an interpretable, model agnostic proxy. Since embedding distance predicts ASR, reducing it should improve attack success, but the relationship is mediated by two factors: perceptual readability (whether the VLM can parse the text) and safety alignment (whether it refuses to comply). Second, we use this as a red teaming tool: we directly maximize image text embedding similarity under bounded $\ell_\infty$ perturbations via CWA-SSA across four surrogate embedding models, stress testing both factors without access to the target model. Experiments across five degradation settings on GPT-4o, Claude Sonnet 4.5, Mistral-Large-3, and Qwen3-VL confirm that optimization recovers readability and reduces safety aligned refusals as two co-occurring effects, with the dominant mechanism depending on the model's safety filter strength and the degree of visual degradation.

Inertial Measurement Unit (IMU)-based Human Activity Recognition (HAR) aims to interpret and classify user behaviors from temporal motion signals. Recently, deep learning frameworks have advanced this task by learning and extracting discriminative spatiotemporal representations, significantly improving recognition performance. However, IMU-based HAR still faces several critical challenges, particularly limited training samples and static knowledge utilization, both of which severely hinder its large-scale deployment. In this paper, we introduce MoRA, the first Retrieval-Augmented Module specifically designed for motion series. It can be flexibly integrated into any existing HAR model, enhancing recognition performance while maintaining inference efficiency. To address issues such as information redundancy in retrieval results and rigid fusion strategies, we propose an uncertainty-adaptive fusion unit within MoRA. This unit leverages previous physical knowledge from IMU signals to dynamically adjust the fusion strategy between original outputs and retrieved information, enabling more robust recognition. Extensive experiments on ten real-world datasets demonstrate that MoRA significantly improves the performance of existing IMU-based HAR models, consistently delivering stable and effective gains. The source code of MoRA is available at: https://github.com/liavonpenn/mora.

Cellular differentiation is governed by gene regulatory networks, the high-dimensional stochastic biochemical systems that determine the transcriptional landscape and mediate cellular responses to signals and perturbations. Although single-cell RNA sequencing provides quantitative snapshots of the transcriptome, current methods for inferring gene-regulatory dynamics often lack mechanistic interpretability and fail to generalize to unseen conditions. Here we introduce Probability Flow Matching (PFM), a scalable framework for learning biophysically consistent stochastic processes directly from time-resolved single-cell measurements. Applying PFM to three hematopoiesis datasets, we show that models with similar interpolation accuracy can encode fundamentally different dynamics, with only biophysically consistent formulations accurately capturing mechanisms of lineage transitions, fate specification, and gene perturbation responses. We further demonstrate that PFM accommodates unbalanced populations, enabling simultaneous inference of cellular proliferation and death dynamics. Together, these results establish PFM as a flexible, scalable framework for integrating mechanistic modeling with single-cell omics.

Reinforcement learning (RL)-based post-training often improves the reasoning performance of large language models (LLMs) beyond the training domain, while supervised fine-tuning (SFT) frequently leads to general capabilities forgetting. However, the mechanisms underlying this contrast remain unclear. To bridge this gap, we present a feature-level mechanistic analysis methodology to probe RL generalization using a controlled experimental setup, where RL- and SFT-tuned models are trained from the same base model on identical data. Leveraging our interpretability framework, we align internal activations across models within a shared feature space and analyze how features evolve during post-training. We find that SFT rapidly introduces many highly specialized features that stabilize early in training, whereas RL induces more restrained and continually evolving feature changes that largely preserve base models' representations. Focusing on samples where RL succeeds but the base model fails, we identify a compact, task-agnostic set of features that directly mediate generalization across diverse tasks. Feature-level interventions confirm their causal role: disabling these features significantly degrades RL models' generalization performance, while amplifying them improves base models' performance. The code is available at https://github.com/danshi777/RL-generalization.

The Controller Area Network (CAN) is a safety-critical in-vehicle communication protocol that lacks built-in security mechanisms, making intrusion detection essential. Existing approaches predominantly formulate CAN intrusion detection as a classification task, mapping complex traffic patterns to attack labels. However, this formulation abstracts away the temporal and relational structure of CAN traffic and misaligns with real-world forensic workflows, which require systematic reasoning about traffic behavior. To address this gap, we introduce CAN-QA, the first benchmark that reformulates CAN traffic analysis as a question-answering (QA) task. CAN-QA converts raw CAN logs into temporally segmented windows and applies deterministic rule-based templates to generate natural-language questions paired with automatically derived ground-truth answers. The resulting dataset comprises 33,128 QA pairs across 10 categories, each targeting distinct semantic and temporal properties of CAN traffic. Using CAN-QA, we evaluate large language models across both True/False and multiple-choice formats. Our results indicate that, although these models capture superficial statistical regularities, they struggle with temporal reasoning, multi-condition inference, and higher-level behavioral interpretation. Our code is available at https://github.com/Kriiiiss/CAN-QA.

Multi-agent debate has been shown to improve reasoning in large language models (LLMs). However, it is compute-intensive, requiring generation of long transcripts before answering questions. To address this inefficiency, we develop a framework that distills multi-agent debate into a single LLM through a two-stage fine-tuning pipeline combining debate structure learning with internalization via dynamic reward scheduling and length clipping. Across multiple models and benchmarks, our internalized models match or exceed explicit multi-agent debate performance using up to 93% fewer tokens. We then investigate the mechanistic basis of this capability through activation steering, finding that internalization creates agent-specific subspaces: interpretable directions in activation space corresponding to different agent perspectives. We further demonstrate a practical application: by instilling malicious agents into the LLM through internalized debate, then applying negative steering to suppress them, we show that distillation makes harmful behaviors easier to localize and control with smaller reductions in general performance compared to steering base models. Our findings offer a new perspective for understanding multi-agent capabilities in distilled models and provide practical guidelines for controlling internalized reasoning behaviors. Code available at https://github.com/johnsk95/latent_agents

Shot Boundary Detection (SBD) aims to automatically identify shot changes and divide a video into coherent shots. While SBD was widely studied in the literature, existing methods often produce non-interpretable boundaries on transitions, miss subtle yet harmful discontinuities, and rely on noisy, low-diversity annotations and outdated benchmarks. To alleviate these limitations, we propose OmniShotCut to formulate SBD as structured relational prediction, jointly estimating shot ranges with intra-shot relations and inter-shot relations, by a shot query-based dense video Transformer. To avoid imprecise manual labeling, we adopt a fully synthetic transition synthesis pipeline that automatically reproduces major transition families with precise boundaries and parameterized variants. We also introduce OmniShotCutBench, a modern wide-domain benchmark enabling holistic and diagnostic evaluation. Experiments on the benchmarks demonstrate the effectiveness and generality of our method.

Linear activation steering is a powerful approach for eliciting the capabilities of large language models and specializing their behavior using limited labeled data. While effective, existing methods often apply a fixed steering strength to all tokens, resulting in inconsistent steering quality across diverse input prompts. In this work, we introduce Contextual Linear Activation Steering (CLAS), a method that dynamically adapts linear activation steering to context-dependent steering strengths. Across eleven steering benchmarks and four model families, it consistently outperforms standard linear activation steering and matches or exceeds the performance of ReFT and LoRA in settings with limited labeled data. We therefore propose CLAS as a scalable, interpretable, and accurate method for specializing and steering large language models.

Identifying the dynamical state variables of a system from high-dimensional observations is a central problem across physical sciences. The challenge is that the state variables are not directly observable and must be inferred from raw high-dimensional data without supervision. Here we introduce DySIB (Dynamical Symmetric Information Bottleneck) as a method to learn low-dimensional representations of time-series data by maximizing predictive mutual information between past and future observation windows while penalizing representation complexity. This objective operates entirely in latent space and avoids reconstruction of the observations. We apply DySIB to an experimental video dataset of a physical pendulum, where the underlying state space is known. The method, with hyperparameters of the learning architecture set self-consistently by the data, recovers a two-dimensional representation that matches the dimensionality, topology, and geometry of the pendulum phase space, with the learned coordinates aligning smoothly with the canonical angle and angular velocity. These results demonstrate, on a well-characterized experimental system, that predictive information in latent space can be used to recover interpretable dynamical coordinates directly from high-dimensional data.

Multimodal Large Language Models (MLLMs) have demonstrated strong image-level visual understanding and reasoning, yet their pixel-level perception across both images and videos remains limited. Foundation segmentation models such as the SAM series produce high-quality masks, but they rely on low-level visual prompts and cannot natively interpret complex conversational instructions. Existing segmentation MLLMs narrow this gap, but are usually specialized for either images or videos and rarely support both textual and visual prompts in one interface. We introduce X2SAM, a unified segmentation MLLM that extends any-segmentation capabilities from images to videos. Given conversational instructions and visual prompts, X2SAM couples an LLM with a Mask Memory module that stores guided vision features for temporally consistent video mask generation. The same formulation supports generic, open-vocabulary, referring, reasoning, grounded conversation generation, interactive, and visual grounded segmentation across image and video inputs. We further introduce the Video Visual Grounded (V-VGD) segmentation benchmark, which evaluates whether a model can segment object tracks in videos from interactive visual prompts. With a unified joint training strategy over heterogeneous image and video datasets, X2SAM delivers strong video segmentation performance, remains competitive on image segmentation benchmarks, and preserves general image and video chat ability.

The development of practical (multimodal) large language model assistants for Korean weather forecasters is hindered by the absence of a multidimensional, expert-level evaluation framework grounded in authoritative sources. To address this, we introduce K-MetBench, a diagnostic benchmark grounded in national qualification exams. It exposes critical gaps across four dimensions: expert visual reasoning of charts, logical validity via expert-verified rationales, Korean-specific geo-cultural comprehension, and fine-grained domain analysis. Our evaluation of 55 models reveals a profound modality gap in interpreting specialized diagrams and a reasoning gap where models hallucinate logic despite correct predictions. Crucially, Korean models outperform significantly larger global models in local contexts, demonstrating that parameter scaling alone cannot resolve cultural dependencies. K-MetBench serves as a roadmap for developing reliable, culturally aware expert AI agents. The dataset is available at https://huggingface.co/datasets/soyeonbot/K-MetBench .

Vision-language models (VLMs) are increasingly proposed as general-purpose tools for scientific data interpretation, yet their reliability on real astronomical observations across diverse modalities remains untested. We present AstroVLBench, a comprehensive benchmark comprising over 4,100 expert-verified instances across five tasks spanning optical imaging, radio interferometry, multi-wavelength photometry, time-domain light curves, and optical spectroscopy. Evaluating six frontier models, we find that performance is strongly modality-dependent: while one model (Gemini 3 Pro) emerges as the most consistently capable across tasks, task-specific strengths vary, and all models substantially underperform domain-specialized methods. Mechanistic ablations reveal that performance depends not only on directing attention to salient visual features but also on grounding those features in physical knowledge. Phenomenological prompts describing what to look for improve accuracy by sharpening model focus, but physical prompts explaining why those features matter perform better overall and yield more balanced classifications with reduced class-specific bias. Consistent with this picture, presenting the underlying one-dimensional measurements directly as numerical tables instead of rendered plots yields up to 13 percentage points improvement. Reasoning quality analysis further demonstrates that, without explicit physical grounding, models may reach correct predictions from phenomenologically plausible cues while providing physically imprecise justifications, establishing that accuracy alone is insufficient for trustworthy scientific deployment. These findings provide the first systematic, multi-modal baselines for VLMs in observational astronomy and identify the specific representation, grounding, and reasoning bottlenecks where current models fail.

RGB-Thermal (T) crowd counting aims to integrate visible-spectrum and thermal infrared information to improve the robustness of crowd density estimation in complex scenes. Although existing studies generally improve counting accuracy through cross-modal feature fusion, most current methods rely on implicit cross-modal fusion strategies and lack explicit modeling of local spatial discrepancies as well as fine-grained characterization of modality reliability at the positional level, thereby limiting the accuracy and interpretability of the fusion process. To address these issues, this paper proposes a two-stage fusion framework, RACANet, a Reliability-Aware Crowd Anchor Network for RGB-T crowd counting. First, we introduce a lightweight cross-modal alignment pretraining stage, which explicitly learns cross-modal semantic correspondences through crowd-prior supervision and local bidirectional soft matching. Then, based on the priors learned during pretraining, a Local Anchor Fusion Module (LAFM) is introduced in the formal training stage. This module generates local semantic anchors by aggregating features from highly reliable regions and further enables adaptive pixel-level feature redistribution with a local attention mechanism. In addition, we propose a discrepancy-aware consistency constraint to dynamically coordinate the reliability of regions where modal representations are consistent. Experiments conducted on two widely used benchmark datasets, RGBT-CC and Drone-RGBT, demonstrate that RACANet outperforms existing methods. The anonymous code is available at https://anonymous.4open.science/r/RACANet-9985.

Accurate trajectory prediction is fundamentally challenging due to high scene heterogeneity - the severe variance in motion velocity, spatial density, and interaction patterns across different real-world environments. However, most existing approaches typically train a single unified model, expecting a fixed-capacity architecture to generalize universally across all possible scenarios. This conventional model-centric paradigm is fundamentally flawed when confronting such extreme heterogeneity, inevitably leading to a severe generalization gap, degraded accuracy, and massive computational waste. To overcome this bottleneck, rather than refining restricted model-centric architectures, we propose selective learning, a novel scene-centric paradigm. It explicitly analyzes the characteristics of the underlying scene to dynamically route inputs to the most appropriate expert models. As a concrete implementation of this paradigm, we introduce SceneSelect. Specifically, SceneSelect utilizes unsupervised clustering on interpretable geometric and kinematic features to discover a latent scene taxonomy. A highly decoupled classification module is then trained to assign real-time inputs to these scene categories, and a highly extensible, plug-and-play scheduling policy automatically dispatches the trajectory sequence to the optimal expert predictor. Crucially, this decoupled design ensures excellent generalization capabilities, allowing seamless integration with different off-the-shelf models and robust adaptation across new datasets without requiring computationally expensive joint retraining. Extensive experiments on three public benchmarks (ETH-UCY, SDD, and NBA) demonstrate that our method consistently outperforms strong single-model and ensemble baselines, achieving an average improvement of 10.5%, showcasing the effectiveness of scene-aware selective learning.

Long-context ability, has become one of the most important iteration direction of next-generation Large Language Models, particularly in semantic understanding/reasoning, code agentic intelligence and recommendation system. However, the standard softmax attention exhibits quadratic time complexity with respect to sequence length. As the sequence length increases, this incurs substantial overhead in long-context settings, leading the training and inference costs of extremely long sequences deteriorate rapidly. Existing solutions mitigate this issue through two technique routings: i) Reducing the KV cache per layer, such as from the head-level compression GQA, and the embedding dimension-level compression MLA, but the KV cache remains linearly dependent on the sequence length at a 1:1 ratio. ii) Interleaving with KV Cache friendly architecture, such as local attention SWA, linear kernel GDN, but often involve trade-offs among KV Cache and long-context modeling effectiveness. Besides the two technique routings, we argue that there exists an intermediate path not well explored: {Maintaining a linear relationship between the KV cache and sequence length, but performing semantic-level compression through a specific ratio $k$}. This $O(n/k)$ path does not pursue a ``minimum KV cache'', but rather trades acceptable memory costs for complete, referential, and interpretable retention of long distant dependency. Motivated by this, we propose Kwai Summary Attention (KSA), a novel attention mechanism that reduces sequence modeling cost by compressing historical contexts into learnable summary tokens.