Current Multimodal Large Language Models exhibit very strong performance for several demanding tasks. While commercial MLLMs deliver acceptable performance in low-resource languages, comparable results remain unattained within the open science community. In this paper, we aim to develop a strong MLLM for a low-resource language, namely Basque. For that purpose, we develop our own training and evaluation image-text datasets. Using two different Large Language Models as backbones, the Llama-3.1-Instruct model and a Basque-adapted variant called Latxa, we explore several data mixtures for training. We show that: i) low ratios of Basque multimodal data (around 20%) are already enough to obtain solid results on Basque benchmarks, and ii) contrary to expected, a Basque instructed backbone LLM is not required to obtain a strong MLLM in Basque. Our results pave the way to develop MLLMs for other low-resource languages by openly releasing our resources.

Accurately estimating the normalization term (also known as the partition function) in the contrastive loss is a central challenge for training Contrastive Language-Image Pre-training (CLIP) models. Conventional methods rely on large batches for approximation, demanding substantial computational resources. To mitigate this issue, prior works introduced per-sample normalizer estimators, which are updated at each epoch in a blockwise coordinate manner to keep track of updated encoders. However, this scheme incurs optimization error that scales with the ratio of dataset size to batch size, limiting effectiveness for large datasets or small batches. To overcome this limitation, we propose NeuCLIP, a novel and elegant optimization framework based on two key ideas: (i) $\textbf{reformulating}$ the contrastive loss for each sample $\textbf{via convex analysis}$ into a minimization problem with an auxiliary variable representing its log-normalizer; and (ii) $\textbf{transforming}$ the resulting minimization over $n$ auxiliary variables (where $n$ is the dataset size) via $\textbf{variational analysis}$ into the minimization over a compact neural network that predicts the log-normalizers. We design an alternating optimization algorithm that jointly trains the CLIP model and the auxiliary network. By employing a tailored architecture and acceleration techniques for the auxiliary network, NeuCLIP achieves more accurate normalizer estimation, leading to improved performance compared with previous methods. Extensive experiments on large-scale CLIP training, spanning datasets from millions to billions of samples, demonstrate that NeuCLIP outperforms previous methods. Code is available at https://github.com/Optimization-AI/NeuCLIP.

Semantic segmentation has achieved great success in ideal conditions. However, when facing extreme conditions (e.g., insufficient light, fierce camera motion), most existing methods suffer from significant information loss of RGB, severely damaging segmentation results. Several researches exploit the high-speed and high-dynamic event modality as a complement, but event and RGB are naturally heterogeneous, which leads to feature-level mismatch and inferior optimization of existing multi-modality methods. Different from these researches, we delve into the edge secret of both modalities for resilient fusion and propose a novel Edge-awareness Semantic Concordance framework to unify the multi-modality heterogeneous features with latent edge cues. In this framework, we first propose Edge-awareness Latent Re-coding, which obtains uncertainty indicators while realigning event-RGB features into unified semantic space guided by re-coded distribution, and transfers event-RGB distributions into re-coded features by utilizing a pre-established edge dictionary as clues. We then propose Re-coded Consolidation and Uncertainty Optimization, which utilize re-coded edge features and uncertainty indicators to solve the heterogeneous event-RGB fusion issues under extreme conditions. We establish two synthetic and one real-world event-RGB semantic segmentation datasets for extreme scenario comparisons. Experimental results show that our method outperforms the state-of-the-art by a 2.55% mIoU on our proposed DERS-XS, and possesses superior resilience under spatial occlusion. Our code and datasets are publicly available at https://github.com/iCVTEAM/ESC.

The convincingness of an argument does not only depend on its structure (logos), the person who makes the argument (ethos), but also on the emotion that it causes in the recipient (pathos). While the overall intensity and categorical values of emotions in arguments have received considerable attention in the research community, we argue that the emotion an argument evokes in a recipient is subjective. It depends on the recipient's goals, standards, prior knowledge, and stance. Appraisal theories lend themselves as a link between the subjective cognitive assessment of events and emotions. They have been used in event-centric emotion analysis, but their suitability for assessing argument convincingness remains unexplored. In this paper, we evaluate whether appraisal theories are suitable for emotion analysis in arguments by considering subjective cognitive evaluations of the importance and impact of an argument on its receiver. Based on the annotations in the recently published ContArgA corpus, we perform zero-shot prompting experiments to evaluate the importance of gold-annotated and predicted emotions and appraisals for the assessment of the subjective convincingness labels. We find that, while categorical emotion information does improve convincingness prediction, the improvement is more pronounced with appraisals. This work presents the first systematic comparison between emotion models for convincingness prediction, demonstrating the advantage of appraisals, providing insights for theoretical and practical applications in computational argumentation.

Metaphors and metaphorical language (MLs) play an important role in healthcare communication between clinicians, patients, and patients' family members. In this work, we focus on Dutch language data from cancer patients. We extract metaphors used by patients using two data sources: (1) cancer patient storytelling interview data and (2) online forum data, including patients' posts, comments, and questions to professionals. We investigate how current state-of-the-art large language models (LLMs) perform on this task by exploring different prompting strategies such as chain of thought reasoning, few-shot learning, and self-prompting. With a human-in-the-loop setup, we verify the extracted metaphors and compile the outputs into a corpus named HealthQuote.NL. We believe the extracted metaphors can support better patient care, for example shared decision making, improved communication between patients and clinicians, and enhanced patient health literacy. They can also inform the design of personalized care pathways. We share prompts and related resources at https://github.com/4dpicture/HealthQuote.NL

Hallucination in large language models (LLMs) remains a critical barrier to their safe deployment. For hallucination detection to be practical in real-world scenarios, the use of efficient small models is essential to ensure low latency and minimal resource consumption. However, existing methods rely on fixed verification strategies, where simply tuning small models to mimic fixed verification trajectories fails to capture the adaptability required for diverse hallucination patterns, thereby inducing planning instability. To address this limitation, we propose a ``Learning to Evaluate and Adaptively Plan'' (LEAP) framework, which shifts hallucination detection from fixed execution to dynamic strategy learning. Specifically, LEAP first employs a powerful teacher model to iteratively explore and refine verification strategies through a failure-driven loop. This dynamic planning capability is then distilled into an efficient student model, augmented by a novel proactive correction mechanism that enables the model to evaluate and optimize its verification strategy before execution. Experiments on three benchmarks demonstrate that LEAP outperforms state-of-the-art methods, offering an effective and scalable solution for reliable hallucination detection.

Reconstructing real-world objects from multi-view images is essential for applications in 3D editing, AR/VR, and digital content creation. Existing methods typically prioritize either geometric accuracy (Multi-View Stereo) or photorealistic rendering (Novel View Synthesis), often decoupling geometry and appearance optimization, which hinders downstream editing tasks. This paper advocates an unified treatment on geometry and appearance optimization for seamless Gaussian-mesh joint optimization. More specifically, we propose a novel framework that simultaneously optimizes mesh geometry (vertex positions and faces) and vertex colors via Gaussian-guided mesh differentiable rendering, leveraging photometric consistency from input images and geometric regularization from normal and depth maps. The obtained high-quality 3D reconstruction can be further exploit in down-stream editing tasks, such as relighting and shape deformation. Our code will be released in https://github.com/zhejia01/TexGuided-GS2Mesh

Critical appraisal of scientific literature is an essential skill in the biomedical field. While large language models (LLMs) can offer promising support in this task, their reliability remains limited, particularly for critical reasoning in specialized domains. We introduce CareMedEval, an original dataset designed to evaluate LLMs on biomedical critical appraisal and reasoning tasks. Derived from authentic exams taken by French medical students, the dataset contains 534 questions based on 37 scientific articles. Unlike existing benchmarks, CareMedEval explicitly evaluates critical reading and reasoning grounded in scientific papers. Benchmarking state-of-the-art generalist and biomedical-specialized LLMs under various context conditions reveals the difficulty of the task: open and commercial models fail to exceed an Exact Match Rate of 0.5 even though generating intermediate reasoning tokens considerably improves the results. Yet, models remain challenged especially on questions about study limitations and statistical analysis. CareMedEval provides a challenging benchmark for grounded reasoning, exposing current LLM limitations and paving the way for future development of automated support for critical appraisal.

Human-interpretable predictions are essential for deploying AI in medical imaging, yet most interpretable-by-design (IBD) frameworks require concept annotations for training data, which are costly and impractical to obtain in clinical contexts. Recent attempts to bypass annotation, such as zero-shot vision-language models or concept-generation frameworks, struggle to capture domain-specific medical features, leading to poor reliability. In this paper, we propose a novel Prior-guided Concept Predictor (PCP), a weakly supervised framework that enables concept answer prediction without explicit supervision or reliance on language models. PCP leverages class-level concept priors as weak supervision and incorporates a refinement mechanism with KL divergence and entropy regularization to align predictions with clinical reasoning. Experiments on PH2 (dermoscopy) and WBCatt (hematology) show that PCP improves concept-level F1-score by over 33% compared to zero-shot baselines, while delivering competitive classification performance on four medical datasets (PH2, WBCatt, HAM10000, and CXR4) relative to fully supervised concept bottleneck models (CBMs) and V-IP.

Cyber-physical systems increasingly rely on foundational models, such as Large Language Models (LLMs) and Vision-Language Models (VLMs) to increase autonomy through enhanced perception, inference, and planning. However, these models also introduce new types of errors, such as hallucinations, over-generalizations, and context misalignments, resulting in incorrect and flawed decisions. To address this, we introduce the concept of Cognition Envelopes, designed to establish reasoning boundaries that constrain AI-generated decisions while complementing the use of meta-cognition and traditional safety envelopes. As with safety envelopes, Cognition Envelopes require practical guidelines and systematic processes for their definition, validation, and assurance. In this paper we describe an LLM/VLM-supported pipeline for dynamic clue analysis within the domain of small autonomous Uncrewed Aerial Systems deployed on Search and Rescue (SAR) missions, and a Cognition Envelope based on probabilistic reasoning and resource analysis. We evaluate the approach through assessing decisions made by our Clue Analysis Pipeline in a series of SAR missions. Finally, we identify key software engineering challenges for systematically designing, implementing, and validating Cognition Envelopes for AI-supported decisions in cyber-physical systems.

Adam [Kingma & Ba, 2015] is the de facto optimizer in deep learning, yet its theoretical understanding remains limited. Prior analyses show that Adam favors solutions aligned with $\ell_\infty$-geometry, but these results are restricted to the full-batch regime. In this work, we study the implicit bias of incremental Adam (using one sample per step) for logistic regression on linearly separable data, and show that its bias can deviate from the full-batch behavior. As an extreme example, we construct datasets on which incremental Adam provably converges to the $\ell_2$-max-margin classifier, in contrast to the $\ell_\infty$-max-margin bias of full-batch Adam. For general datasets, we characterize its bias using a proxy algorithm for the $β_2 \to 1$ limit. This proxy maximizes a data-adaptive Mahalanobis-norm margin, whose associated covariance matrix is determined by a data-dependent dual fixed-point formulation. We further present concrete datasets where this bias reduces to the standard $\ell_2$- and $\ell_\infty$-max-margin classifiers. As a counterpoint, we prove that Signum [Bernstein et al., 2018] converges to the $\ell_\infty$-max-margin classifier for any batch size. Overall, our results highlight that the implicit bias of Adam crucially depends on both the batching scheme and the dataset, while Signum remains invariant.

Autonomous navigation under natural language instructions represents a crucial step toward embodied intelligence, enabling complex task execution in environments ranging from industrial facilities to domestic spaces. However, language-driven 3D navigation for Unmanned Aerial Vehicles (UAVs) requires precise spatial reasoning, a capability inherently lacking in current zero-shot Vision-Language Models (VLMs) which often generate ambiguous outputs and cannot guarantee geometric feasibility. Furthermore, existing Vision-Language Navigation (VLN) methods are predominantly tailored for 2.5D ground robots, rendering them unable to generalize to the unconstrained 3D spatial reasoning required for aerial tasks in small-scale, cluttered environments. In this paper, we present SoraNav, a novel framework enabling zero-shot VLM reasoning for UAV task-centric navigation. To address the spatial-semantic gap, we introduce Multi-modal Visual Annotation (MVA), which encodes 3D geometric priors directly into the VLM's 2D visual input. To mitigate hallucinated or infeasible commands, we propose an Adaptive Decision Making (ADM) strategy that validates VLM proposals against exploration history, seamlessly switching to geometry-based exploration to avoid dead-ends and redundant revisits. Deployed on a custom PX4-based micro-UAV, SoraNav demonstrates robust real-world performance. Quantitative results show our approach significantly outperforms state-of-the-art baselines, increasing Success Rate (SR) by 25.7% and navigation efficiency (SPL) by 17.3% in 2.5D scenarios, and achieving improvements of 39.3% (SR) and 24.7% (SPL) in complex 3D scenarios.

Public research results on large-scale supervised finetuning of AI agents remain relatively rare, since the collection of agent training data presents unique challenges. In this work, we argue that the bottleneck is not a lack of underlying data sources, but that a large variety of data is fragmented across heterogeneous formats, tools, and interfaces. To this end, we introduce the agent data protocol (ADP), a light-weight representation language that serves as an "interlingua" between agent datasets in diverse formats and unified agent training pipelines downstream. The design of ADP is expressive enough to capture a large variety of tasks, including API/tool use, browsing, coding, software engineering, and general agentic workflows, while remaining simple to parse and train on without engineering at a per-dataset level. In experiments, we unified a broad collection of 13 existing agent training datasets into ADP format, and converted the standardized ADP data into training-ready formats for multiple agent frameworks. We performed SFT on these data, and demonstrated an average performance gain of ~20% over corresponding base models, and delivers state-of-the-art or near-SOTA performance on standard coding, browsing, tool use, and research benchmarks, without domain-specific tuning. All code and data are released publicly, in the hope that ADP could help lower the barrier to standardized, scalable, and reproducible agent training.

Sarcasm is a complex form of figurative language in which the intended meaning contradicts the literal one. Its prevalence in social media and popular culture poses persistent challenges for natural language understanding, sentiment analysis, and content moderation. With the emergence of multimodal large language models, sarcasm detection extends beyond text and requires integrating cues from audio and vision. We present MuSaG, the first German multimodal sarcasm detection dataset, consisting of 33 minutes of manually selected and human-annotated statements from German television shows. Each instance provides aligned text, audio, and video modalities, annotated separately by humans, enabling evaluation in unimodal and multimodal settings. We benchmark nine open-source and commercial models, spanning text, audio, vision, and multimodal architectures, and compare their performance to human annotations. Our results show that while humans rely heavily on audio in conversational settings, models perform best on text. This highlights a gap in current multimodal models and motivates the use of MuSaG for developing models better suited to realistic scenarios. We release MuSaG publicly to support future research on multimodal sarcasm detection and human-model alignment.

LaViRA: Zero-shot Vision-and-Language Navigation in Continuous Environments (VLN-CE) requires an agent to navigate unseen environments based on natural language instructions without any prior training. Current methods face a critical trade-off: either rely on environment-specific waypoint predictors that limit scene generalization, or underutilize the reasoning capabilities of large models during navigation. We introduce LaViRA, a simple yet effective zero-shot framework that addresses this dilemma by decomposing action into a coarse-to-fine hierarchy: Language Action for high-level planning, Vision Action for middle-level perceptual grounding, and Robot Action for low-level control. This modular decomposition allows us to leverage the distinct strengths of different scales of Multimodal Large Language Models (MLLMs) at each stage, creating a system that is powerful in its reasoning, grounding and practical control. LaViRA significantly outperforms existing state-of-the-art methods on the VLN-CE benchmark, demonstrating superior generalization capabilities in unseen environments, while maintaining transparency and efficiency for real-world deployment. Project page: https://robo-lavira.github.io/lavira-zs-vln/

We present Kaleido, a subject-to-video~(S2V) generation framework, which aims to synthesize subject-consistent videos conditioned on multiple reference images of target subjects. Despite recent progress in S2V generation models, existing approaches remain inadequate at maintaining multi-subject consistency and at handling background disentanglement, often resulting in lower reference fidelity and semantic drift under multi-image conditioning. These shortcomings can be attributed to several factors. Primarily, the training dataset suffers from a lack of diversity and high-quality samples, as well as cross-paired data, i.e., paired samples whose components originate from different instances. In addition, the current mechanism for integrating multiple reference images is suboptimal, potentially resulting in the confusion of multiple subjects. To overcome these limitations, we propose a dedicated data construction pipeline, incorporating low-quality sample filtering and diverse data synthesis, to produce consistency-preserving training data. Moreover, we introduce Reference Rotary Positional Encoding (R-RoPE) to process reference images, enabling stable and precise multi-image integration. Extensive experiments across numerous benchmarks demonstrate that Kaleido significantly outperforms previous methods in consistency, fidelity, and generalization, marking an advance in S2V generation.

Honesty alignment-the ability of large language models (LLMs) to recognize their knowledge boundaries and express calibrated confidence-is essential for trustworthy deployment. Existing methods either rely on training-free confidence estimation (e.g., token probabilities, self-consistency) or training-based calibration with correctness annotations. While effective, achieving universal honesty alignment with training-based calibration requires costly, large-scale labeling. To support annotation-efficient training, we introduce Elicitation-Then-Calibration (EliCal), a two-stage framework that first elicits internal confidence using inexpensive self-consistency supervision, then calibrates this confidence with a small set of correctness annotations. To support a large-scale study, we release HonestyBench, a benchmark covering ten free-form QA datasets with 560k training and 70k evaluation instances annotated with correctness and self-consistency signals. Experiments show that EliCal achieves near-optimal alignment with only 1k correctness annotations (0.18% of full supervision) and better alignment performance on unseen MMLU tasks than the calibration-only baseline, offering a scalable solution toward universal honesty alignment in LLMs.

Pretrained multi-modal large language models (MLLMs) demonstrate strong performance on diverse multimodal tasks, but remain limited in reasoning capabilities for domains where annotations are difficult to collect. In this work, we focus on artificial image domains such as charts, rendered documents, and webpages, which are abundant in practice yet lack large-scale human annotated reasoning datasets. We introduce COGS (COmposition-Grounded data Synthesis), a data-efficient framework for equipping MLLMs with advanced reasoning abilities from a small set of seed questions. The key idea is to decompose each seed question into primitive perception and reasoning factors, which can then be systematically recomposed with new images to generate large collections of synthetic question-answer pairs. Each generated question is paired with subquestions and intermediate answers, enabling reinforcement learning with factor-level process rewards. Experiments on chart reasoning show that COGS substantially improves performance on unseen questions, with the largest gains on reasoning-heavy and compositional questions. Moreover, training with a factor-level mixture of different seed data yields better transfer across multiple datasets, suggesting that COGS induces generalizable capabilities rather than dataset-specific overfitting. We further demonstrate that the framework extends beyond charts to other domains such as webpages.

The fields of explainable AI and mechanistic interpretability aim to uncover the internal structure of neural networks, with circuit discovery as a central tool for understanding model computations. Existing approaches, however, rely on manual inspection and remain limited to toy tasks. Automated interpretability offers scalability by analyzing isolated features and their activations, but it often misses interactions between features and depends strongly on external LLMs and dataset quality. Transcoders have recently made it possible to separate feature attributions into input-dependent and input-invariant components, providing a foundation for more systematic circuit analysis. Building on this, we propose WeightLens and CircuitLens, two complementary methods that go beyond activation-based analysis. WeightLens interprets features directly from their learned weights, removing the need for explainer models or datasets while matching or exceeding the performance of existing methods on context-independent features. CircuitLens captures how feature activations arise from interactions between components, revealing circuit-level dynamics that activation-only approaches cannot identify. Together, these methods increase interpretability robustness and enhance scalable mechanistic analysis of circuits while maintaining efficiency and quality.

Contrastive Vision-Language Models (VLMs) have demonstrated strong zero-shot capabilities. However, their cross-modal alignment remains biased toward English due to limited multilingual multimodal data. Recent multilingual extensions have alleviated this gap but enforce instance-level alignment while neglecting the global geometry of the shared embedding space. We address this problem by introducing ToMCLIP (Topological Alignment for Multilingual CLIP), a topology-aware framework aligning embedding spaces with topology-preserving constraints. The proposed method applies persistent homology to define a topological alignment loss and approximates persistence diagram with theoretical error bounds using graph sparsification strategy. This work validates the proposed approach, showing enhanced structural coherence of multilingual representations, higher zero-shot accuracy on the CIFAR-100, and stronger multilingual retrieval performance on the xFlickr&CO. Beyond VLMs, the proposed approach provides a general method for incorporating topological alignment into representation learning. Code is available at https://github.com/junwon0/ToMCLIP.git.

Music learners can greatly benefit from tools that accurately detect errors in their practice. Existing approaches typically compare audio recordings to music scores using heuristics or learnable models. This paper introduces LadderSym, a novel Transformer-based method for music error detection. LadderSym is guided by two key observations about the state-of-the-art approaches: (1) late fusion limits inter-stream alignment and cross-modality comparison capability; and (2) reliance on score audio introduces ambiguity in the frequency spectrum, degrading performance in music with concurrent notes. To address these limitations, LadderSym introduces (1) a two-stream encoder with inter-stream alignment modules to improve audio comparison capabilities and error detection F1 scores, and (2) a multimodal strategy that leverages both audio and symbolic scores by incorporating symbolic representations as decoder prompts, reducing ambiguity and improving F1 scores. We evaluate our method on the MAESTRO-E and CocoChorales-E datasets by measuring the F1 score for each note category. Compared to the previous state of the art, LadderSym more than doubles F1 for missed notes on MAESTRO-E (26.8% -> 56.3%) and improves extra note detection by 14.4 points (72.0% -> 86.4%). Similar gains are observed on CocoChorales-E. Furthermore, we also evaluate our models on real data we curated. This work introduces insights about comparison models that could inform sequence evaluation tasks for reinforcement learning, human skill assessment, and model evaluation. Code: https://github.com/ben2002chou/LadderSYM

Vision-Language Models (VLMs) have shown remarkable capabilities in spatial reasoning, yet they remain fundamentally limited to qualitative precision and lack the computational precision required for real-world robotics. Current approaches fail to leverage metric cues from depth sensors and camera calibration, instead reducing geometric problems to pattern recognition tasks that cannot deliver the centimeter-level accuracy essential for robotic manipulation. We present TIGeR (Tool-Integrated Geometric Reasoning), a novel framework that transforms VLMs from perceptual estimators to geometric computers by enabling them to generate and execute precise geometric computations through external tools. Rather than attempting to internalize complex geometric operations within neural networks, TIGeR empowers models to recognize geometric reasoning requirements, synthesize appropriate computational code, and invoke specialized libraries for exact calculations. To support this paradigm, we introduce TIGeR-300K, a comprehensive tool-invocation-oriented dataset covering point transformations, pose estimation, and spatial compatibility verification, complete with tool invocation sequences and intermediate computations. Through a two-stage training pipeline combining supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT) with our proposed hierarchical reward design, TIGeR achieves SOTA performance on geometric reasoning benchmarks while demonstrating centimeter-level precision in real-world robotic manipulation tasks.

Real-world robotic agents must act under partial observability and long horizons, where key cues may appear long before they affect decision making. However, most modern approaches rely solely on instantaneous information, without incorporating insights from the past. Standard recurrent or transformer models struggle with retaining and leveraging long-term dependencies: context windows truncate history, while naive memory extensions fail under scale and sparsity. We propose ELMUR (External Layer Memory with Update/Rewrite), a transformer architecture with structured external memory. Each layer maintains memory embeddings, interacts with them via bidirectional cross-attention, and updates them through an Least Recently Used (LRU) memory module using replacement or convex blending. ELMUR extends effective horizons up to 100,000 times beyond the attention window and achieves a 100% success rate on a synthetic T-Maze task with corridors up to one million steps. In POPGym, it outperforms baselines on more than half of the tasks. On MIKASA-Robo sparse-reward manipulation tasks with visual observations, it nearly doubles the performance of strong baselines, achieving the best success rate on 21 out of 23 tasks and improving the aggregate success rate across all tasks by about 70% over the previous best baseline. These results demonstrate that structured, layer-local external memory offers a simple and scalable approach to decision making under partial observability. Code and project page: https://elmur-paper.github.io/.

While modern visual generation models excel at creating aesthetically pleasing natural images, they struggle with producing or editing structured visuals like charts, diagrams, and mathematical figures, which demand composition planning, text rendering, and multimodal reasoning for factual fidelity. To address this, we present the first comprehensive, systematic investigation of this domain, encompassing data construction, model training, and an evaluation benchmark. First, we construct a large-scale dataset of 1.3 million high-quality structured image pairs derived from executable drawing programs and augmented with chain-of-thought reasoning annotations. Building on it, we train a unified model that integrates a VLM with FLUX.1 Kontext via a lightweight connector for enhanced multimodal understanding. A three-stage training curriculum enables progressive feature alignment, knowledge infusion, and reasoning-augmented generation, further boosted by an external reasoner at inference time. Finally, we introduce StructBench, a novel benchmark for generation and editing with over 1,700 challenging instances, and an accompanying evaluation metric, StructScore, which employs a multi-round Q\&A protocol to assess fine-grained factual accuracy. Evaluations of 15 models reveal that even leading closed-source systems remain far from satisfactory. Our model attains strong editing performance, and inference-time reasoning yields consistent gains across diverse architectures. By releasing the dataset, model, and benchmark, we aim to advance unified multimodal foundations for structured visuals.

Modeling the dynamics of cellular differentiation is fundamental to advancing the understanding and treatment of diseases associated with this process, such as cancer. With the rapid growth of single-cell datasets, this has also become a particularly promising and active domain for machine learning. Current state-of-the-art models, however, rely on computationally expensive optimal transport preprocessing and multi-stage training, while also not discovering explicit gene interactions. To address these challenges we propose Cell-Mechanistic Neural Networks (Cell-MNN), an encoder-decoder architecture whose latent representation is a locally linearized ODE governing the dynamics of cellular evolution from stem to tissue cells. Cell-MNN is fully end-to-end (besides a standard PCA pre-processing) and its ODE representation explicitly learns biologically consistent and interpretable gene interactions. Empirically, we show that Cell-MNN achieves competitive performance on single-cell benchmarks, surpasses state-of-the-art baselines in scaling to larger datasets and joint training across multiple datasets, while also learning interpretable gene interactions that we validate against the TRRUST database of gene interactions.

Although reinforcement learning (RL) has emerged as a promising approach for improving vision-language models (VLMs) and multimodal large language models (MLLMs), current methods rely heavily on manually curated datasets and costly human verification, which limits scalable self-improvement in multimodal systems. To address this challenge, we propose Vision-Zero, a label-free, domain-agnostic multi-agent self-play framework for self-evolving VLMs through competitive visual games generated from arbitrary image inputs. Specifically, Vision-Zero encompasses three main attributes: (1) Strategic Self-Play Framework: Vision-Zero trains VLMs in "Who Is the Spy"-style games, where the models engage in strategic reasoning and actions across multiple roles. Through interactive gameplay, models autonomously generate their training data without human annotation. (2) Gameplay from Arbitrary Images: Unlike existing gamified frameworks, Vision-Zero can generate games from arbitrary images, thereby enhancing the model's reasoning ability across diverse domains and showing strong generalization to different tasks. We demonstrate this versatility using three distinct types of image datasets: CLEVR-based synthetic scenes, charts, and real-world images. (3) Sustainable Performance Gain: We introduce Iterative Self-Play Policy Optimization (Iterative-SPO), a novel training algorithm that alternates between Self-Play and reinforcement learning with verifiable rewards (RLVR), mitigating the performance plateau often seen in self-play-only training and achieving sustained long-term improvements. Despite using label-free data, Vision-Zero achieves state-of-the-art performance on reasoning, chart question answering, and vision-centric understanding tasks, surpassing other annotation-based methods. Models and code have been released at https://github.com/wangqinsi1/Vision-Zero.

As machine learning systems increasingly train on self-annotated data, they risk reinforcing errors and becoming echo chambers of their own beliefs. We model this phenomenon by introducing a learning-theoretic framework: Online Learning in the Replay Setting. In round $t$, the learner outputs a hypothesis $\hat{h}_t$; the adversary then reveals either the true label $f^\ast(x_t)$ or a replayed label $\hat{h}_i(x_t)$ from an earlier round $i < t$. A mistake is counted only when the true label is shown, yet classical algorithms such as the SOA or the halving algorithm are easily misled by the replayed errors. We introduce the Extended Threshold dimension, $\mathrm{ExThD}(\mathcal{H})$, and prove matching upper and lower bounds that make $\mathrm{ExThD}(\mathcal{H})$ the exact measure of learnability in this model. A closure-based learner makes at most $\mathrm{ExThD}(\mathcal{H})$ mistakes against any adaptive adversary, and no algorithm can perform better. For stochastic adversaries, we prove a similar bound for every intersection-closed class. The replay setting is provably harder than the classical mistake bound setting: some classes have constant Littlestone dimension but arbitrarily large $\mathrm{ExThD}(\mathcal{H})$. Proper learning exhibits an even sharper separation: a class is properly learnable under replay if and only if it is (almost) intersection-closed. Otherwise, every proper learner suffers $Ω(T)$ errors, whereas our improper algorithm still achieves the $\mathrm{ExThD}(\mathcal{H})$ bound. These results give the first tight analysis of learning against replay adversaries, based on new results for closure-type algorithms.

Deep Research Agents (DRAs) can autonomously conduct complex investigations and generate comprehensive reports, demonstrating strong real-world potential. However, existing evaluations mostly rely on close-ended benchmarks, while open-ended deep research benchmarks remain scarce and typically neglect personalized scenarios. To bridge this gap, we introduce Personalized Deep Research Bench (PDR-Bench), the first benchmark for evaluating personalization in DRAs. It pairs 50 diverse research tasks across 10 domains with 25 authentic user profiles that combine structured persona attributes with dynamic real-world contexts, yielding 250 realistic user-task queries. To assess system performance, we propose the PQR Evaluation Framework, which jointly measures Personalization Alignment, Content Quality, and Factual Reliability. Our experiments on a range of systems highlight current capabilities and limitations in handling personalized deep research. This work establishes a rigorous foundation for developing and evaluating the next generation of truly personalized AI research assistants.

Tool agents interact with users through multi-turn dialogues to accomplish various tasks. Recent studies have adopted user simulation methods to develop these agents in multi-turn settings. However, existing user simulators tend to be agent-friendly, exhibiting only cooperative behaviors, failing to train and test agents against non-collaborative users in the real world. We propose a novel user simulator architecture that simulates four categories of non-collaborative behaviors: requesting unavailable services, digressing into tangential conversations, expressing impatience, and providing incomplete utterances. Our user simulator can simulate challenging and natural non-collaborative behaviors while reliably delivering all intents and information necessary to accomplish the task. Our experiments on MultiWOZ and τ-bench reveal significant performance degradation in state-of-the-art tool agents when encountering non-collaborative users, as well as agent weaknesses under each non-collaborative condition such as escalated hallucinations and dialogue breakdowns. Our findings point to the need for methods that can improve agent robustness to the wide range of user behaviors encountered in deployment. We release the extensible simulation framework to help the community develop and stress-test tool agents under realistic conditions within their own service domains. Our code is available at https://github.com/holi-lab/NCUser.

Class Incremental Learning (CIL) requires models to continuously learn new classes without forgetting previously learned ones, while maintaining stable performance across all possible class sequences. In real-world settings, the order in which classes arrive is diverse and unpredictable, and model performance can vary substantially across different sequences. Yet mainstream evaluation protocols calculate mean and variance from only a small set of randomly sampled sequences. Our theoretical analysis and empirical results demonstrate that this sampling strategy fails to capture the full performance range, resulting in biased mean estimates and a severe underestimation of the true variance in the performance distribution. We therefore contend that a robust CIL evaluation protocol should accurately characterize and estimate the entire performance distribution. To this end, we introduce the concept of extreme sequences and provide theoretical justification for their crucial role in the reliable evaluation of CIL. Moreover, we observe a consistent positive correlation between inter-task similarity and model performance, a relation that can be leveraged to guide the search for extreme sequences. Building on these insights, we propose EDGE (Extreme case-based Distribution & Generalization Evaluation), an evaluation protocol that adaptively identifies and samples extreme class sequences using inter-task similarity, offering a closer approximation of the ground-truth performance distribution. Extensive experiments demonstrate that EDGE effectively captures performance extremes and yields more accurate estimates of distributional boundaries, providing actionable insights for model selection and robustness checking. Our code is available at https://github.com/AIGNLAI/EDGE.