We propose Puffin-World, a unified multimodal architecture that integrates physical understanding, spatial simulation, and 3D world generation and reconstruction without relying on external offline modules. To reliably construct and interact with 3D worlds, our framework jointly models three native world states: physics (gravity field and latitude), geometry (depth), and appearance (image), together with a unified Omni-Camera representation that supports diverse tasks and flexible motions. Beyond modeling these states, we introduce a strategy for propagating physical dynamics across future frames. By grounding absolute camera properties in the real world, Puffin-World enables physically consistent and visually stable world generation. We further couple appearance and geometry within a single generative process, jointly synthesizing each future view and reconstructing its underlying geometry. This unified paradigm enables interleaved closed-loop applications requiring synergy across multiple tasks, including mimic and self-calibrated world exploration. To scale Puffin-World to complex scenarios, we construct Puffin-16M, comprising 15 million vision-language-camera triplets and 1 million trajectories featuring various and challenging motions. To foster further research in this area, we released the code, models, and datasets.
Landslides are widespread geological hazards, yet their automated detection and mapping in remote sensing imagery remain challenging because of their irregular morphology, ambiguous spectral signatures, and substantial domain shifts across imaging platforms. To overcome these challenges, we propose EarthLD, a vision-language-guided diffusion framework for open-world landslide understanding, enabling unified landslide recognition, mapping, and trigger interpretation. At its core, EarthLD formulates landslide understanding as a diffusion process that progressively infers the presence, spatial extent, and pixel-level boundaries of landslides from noisy latent representations. This probabilistic formulation enables the model to jointly perform image-level landslide recognition and mapping while characterizing predictive uncertainty. By integrating visual observations with contextual knowledge in the denoising process, EarthLD distinguishes diverse landslides from backgrounds, produces confidence-aware predictions for suspected regions, and maps landslide ranges. We additionally construct a global-scale open-world landslide benchmark by systematically harmonizing multiple publicly available remote sensing data collected by diverse institutions. Extensive experiments across regions, sensors, and triggering events demonstrate that EarthLD consistently outperforms existing landslide detection methods, highlighting its potential as a unified and robust solution for global geological-hazard monitoring and emergency response.
Indirect speech acts (ISAs) require pragmatic reasoning over context, as directive intent can- not be inferred from surface form alone. Prior text-based studies and existing multimodal benchmarks largely overlook this requirement, focusing instead on explicitly encoded context or perceptual recognition, and thus underex- plore context-dependent pragmatic understand- ing, particularly in high-context languages such as Korean. We introduce READI, a multimodal benchmark for evaluating ISA understanding through integrated reasoning over visual con- text and dialogue. READI models graded in- directness grounded in pragmatic theory and formulates the task as vision-based pragmatic question answering (V-PQA), supporting cross- lingual evaluation in English and Korean. Ex- periments show that even state-of-the-art multi- modal models struggle with visually grounded indirect speech acts, with performance declin- ing as indirectness increases, underscoring the need for benchmarks that explicitly target con- textual pragmatic reasoning.
Referring single-object tracking enables language-grounded target initialization and subsequent tracking by jointly leveraging semantic cues and visual templates. The core difficulty is to use language differently across stages: it is indispensable for grounding but can induce semantic drift during tracking when overemphasized. Meanwhile, current methods often require costly vision-language alignment training. We present LVTrack, a pure transformer framework that introduces a mode-conditioned Gated Feature Injector to adaptively regulate textual guidance and alleviate semantic drift. Together with targeted adaptations, it directly harnesses a frozen vision-language pretrained model, greatly reducing training cost and preserving strong language understanding. To further improve temporal localization, LVTrack integrates hybrid relative-absolute positional encodings with a lightweight memory mechanism and optimizes autoregressive box prediction using a Gaussian-smoothed KL loss. Extensive experiments on standard benchmarks demonstrate that LVTrack achieves strong performance.
Kishor Datta Gupta, Ahmed Rafi Hasan, Md. Mahfuzur Rahman +2cs.CV
Locating a specific object instance in a cluttered scene using a single reference image and a short description, and reporting when that instance is absent, large vision-language models usually address this task. We ask whether the same capability is available far more cheaply, from representations already learned by a world-model pretraining objective. We present WALDO, a one-shot exemplar- and language-conditioned detection head with 3.4M trainable parameters that reads frozen V-JEPA 2.1 features to jointly predict object localization and target presence, with no gradient on the backbone. Because exemplar-conditioned supervision is scarce, we synthesize training episodes from instance annotations, mining exemplars from ground-truth boxes and constructing absence cases that exclude the referenced instance while leaving same-category distractors in view. This is easy to get wrong: in the obvious implementation, crop size alone predicts the label, and a head trained on it reaches 0.9998 absence AUROC without ever consulting the exemplar, and we report the negative controls that close the shortcut. On 35 held-out cluttered scenes, WALDO achieves a 0.461 catalogue AP@50, compared to 0.306 for a prompted Grounding DINO baseline under an identical scorer. Substituting DINOv3 for V-JEPA under a matched 576-token grid drops within-category absence AUROC from 0.880 to 0.726 and instance AP@50 from 0.201 to 0.141, isolating the pretraining objective rather than input resolution as the source of the gain. Instance-level Success@1, however, reaches only 0.190 against a 0.190 category-chance floor: world-model features transfer to localization precision and absence detection but not to instance identity.
Time series forecasting models operate on raw numerical sequences, lacking the semantic knowledge that domain experts implicitly leverage, such as the physical meaning of each variable, its statistical behavior, and its temporal dynamics. Recent efforts to bridge this gap fall into two camps. Some rely on large language models at inference time, which is computationally expensive. Others apply uniform textual prompts at the dataset level, ignoring the heterogeneous semantics across individual variates. We propose SAGE (Seeing and Augmenting with Grounded Encoding), an end-to-end CLIP-based framework that jointly models temporal, cross-variable, textual, and visual information. The CLIP text encoder processes frequency-enhanced patches and variable tokens, while gated residual paths inject variable-specific descriptions and statistical descriptors. In parallel, the frozen CLIP vision encoder aligns rendered series with temporal representations through a training-only contrastive objective. This dual use of CLIP adds complementary semantic and visual supervision without placing an LLM in the forecasting loop. Across eight long-term benchmarks and M4, SAGE achieves state-of-the-art accuracy. Ablations confirm complementary gains from multimodal alignment and variable-level knowledge.
Text-guided drone geo-localization aims to identify a target region in a large-scale image gallery from a natural-language description. Existing methods mainly formulate this task as direct matching between an open-ended text query and candidate images. However, incomplete queries and highly similar candidates often make global cross-modal matching insufficient for reliable fine-grained localization. We propose UniGeo, a unified multimodal large language model (MLLM) for text-guided drone geo-localization. Built on a shared vision-language framework, UniGeo jointly supports geo-semantic understanding, cross-view semantic generation, and candidate-level verification. Specifically, it establishes stable correspondences among local scene elements, spatial relations, and language descriptions through geo-semantic learning, and further models semantic mappings between drone and satellite views through cross-view generation. Based on these capabilities, a plug-and-play verification module performs fine-grained discrimination among highly confusable candidates. We further introduce a multi-stage training strategy that progressively learns geo-semantic understanding, cross-view generation, and candidate verification, improving adaptation to text-guided geo-localization. Experiments demonstrate consistent improvements across multiple retrieval backbones. On GeoText-1652, UniGeo improves R@10 and mAP by 13.59 and 2.83 percentage points, respectively, validating its effectiveness for fine-grained text-guided drone geo-localization.
Gait recognition has achieved remarkable progress, yet existing methods remain confined to rigid visual matching and often overlook the potential of natural language instructions for interactive retrieval. In this paper, we introduce Composed Gait Retrieval (CoGR), a novel task that retrieves a target gait sequence based on a reference sequence and a natural language modification query. To address the absence of existing datasets for this task, we design an automated annotation pipeline powered by large vision-language models (VLMs) to construct the first gait-language datasets: Language-Augmented CCPG and Language-Augmented CASIA-B. Building on this, we propose ComposeGait, an identity-anchored composition framework designed to prevent the identity drift that arises when generic composed retrieval follows the instruction but returns the wrong person. Its Part-aware Identity Adapter (PIA) aggregates multi-frame, part-aware identity evidence into a sample-specific ID token. We inject the ID tokens into both branches of a shared Q-Former to preserve identity, while excluding the ID-token outputs from the final retrieval embeddings. Joint identity and task-adapted composed-retrieval objectives optimize this space end to end. We evaluate ComposeGait on both benchmarks and show that it achieves the best R@1 among the compared methods, reaching 72.38% on Language-Augmented CCPG and 83.61% on Language-Augmented CASIA-B. These results establish ComposeGait as a strong baseline for CoGR. The datasets and code will be made publicly available.
Reducing annotation requirements remains a key challenge in developing robust medical object detectors. To address this, Vision-Language (VL) object detection methods leverage grounding text information to enable powerful zero-shot and few-shot object detectors in the natural image domain [1, 2, 3, 4]. However, transferring these methods to the medical domain is challenging due to the absence of comparable quality and quantity of the grounding data. Regardless, significant contextual and non-imaging information exists in medical images that remains underutilized. Few-shot learning (FSL) techniques partially address this limitation but struggle to general ize to unseen medical findings and require extensive retraining when new findings are introduced [5, 6]. To overcome these challenges, we extend our prior EM-DETR framework [7] and introduce a scalable FS detection approach designed for efficient abnormality detection in Chest X-Ray (CXR) images under minimal supervision. The proposed architecture incorporates exemplar-based feature generation and domain-aware contrastive optimization, enabling effective adaptation to novel disease findings without exhaustive retraining. Our method achieves near state-of-the-art (SOTA) detection performance using less than 10% of the annotated data, demonstrating its potential for practical, annotation-efficient clinical deployment across both proprietary and public CXR datasets.
Incremental Object Detection (IOD) aims to enable detectors to continuously learn novel categories while preserving previously acquired knowledge. However, existing methods suffer from two forms of \textbf{class knowledge coupling}: class boundary erosion induced by shared parameter updates and class representation entanglement arising from mixed feature encoding. We argue that effective incremental learning requires class-specific computational pathways that enable isolated parameter updates and separated class-wise injection. To this end, we propose \textbf{C$^2$Path}, a class-conditional pathway decoupling framework for vision-language incremental object detection that leverages token-level class cues to establish dedicated and updatable computational pathways for different categories. Specifically, C$^2$Path introduces a category expert library and a class-conditional decoupling module. The expert library consists of learnable low-rank computational nodes that capture category-specific knowledge, while the decoupling module generates class-aware routing signals to dynamically compose \textit{ClassLoRA} adapters from these experts, thereby forming class-specific computational pathways for isolated updates and separated injection across categories. Extensive experiments on COCO 2017 under multiple incremental learning settings demonstrate that C$^2$Path consistently outperforms state-of-the-art methods, providing an effective and scalable solution for continual category expansion in vision-language detectors.
Guangyuan Dong, Chuang Liu, Haoyu Wang +10cs.MM cs.AI cs.CV cs.GR
Multimodal information systems increasingly route generated visual content back through the same vision-language index that informed its production, so the output must remain retrievable by the queries it was meant to serve. When the scene contains entities at vastly different scales, existing language-guided generators condition on a single, globally pooled text embedding and quietly drop scale-specific concepts, breaking concept-query retrieval even when pixel fidelity is high. We formalise this failure as semantic collapse and propose CERES, a closed-loop multimodal indexing framework that builds a three-level semantic pyramid, mines implicit concepts via a co-occurrence-aware router, performs scale-routed cross-attention into a lightweight U-Net generator, and verifies coverage by re-indexing the generated image with the same frozen VLM. A continuously differentiable soft-Jaccard coverage objective returns dense gradients to the 0.39 M-parameter generator under explicit non-degeneracy conditions, and coverage is verified by an independent DINOv2 linear probe trained only on external scene and object labels. On four pansharpening benchmarks across seven settings, CERES delivers the new state of the art with the largest gains where scale variation is most extreme (+4.64% relative Q2n and +9.7 mAP for DOTA detection). It also improves concept-query retrieval Recall@5 by +14.0 points and image-text mean reciprocal rank by 0.19 over the strongest baseline, showing that the closed loop preserves queryable content rather than self-referential feature consistency.
Estimating interpretable conditional-dependence structures from multimodal visual-linguistic features remains largely unexplored. We propose CM-GLasso (Cross-Modal Graphical Lasso), a framework that bridges vision-language representation learning and sparse Gaussian Graphical Models. CM-GLasso introduces three key components: (i) a text visualization strategy that renders class-attribute descriptions as images and processes them through the same SigLIP-2 vision encoder as natural images, yielding prototype-indexed patch-level attention footprints in a shared feature coordinate system; (ii) a cross-attention distillation mechanism that condenses high-dimensional patches into a small set of semantic graph nodes, whose attention-footprint similarities yield cross-modal structural priors for non-uniform L1 penalization; (iii) a joint ADMM formulation that estimates shared and class-specific precision components within a single convex objective, avoiding the need to first estimate and then decompose separate class-wise graphs. The learned sparse graph topologies directly support a parameter-free, precision-based classification rule and a lightweight topology-aware segmentation head. Extensive experiments on eight benchmarks demonstrate that CM-GLasso achieves competitive or superior performance compared with strong feature-based and task-specific baselines. Under the matched controlled protocol, it attains the highest average classification accuracy (91.97%) and the highest segmentation mIoU among the controlled baselines on VOC (74.75%) and ADE20K (64.01%), while also yielding explicit sparse conditional-dependence graphs with common-specific decomposition.
Unified multimodal models (UMMs) are motivated by the hope that understanding and generation reinforce each other but controlled ablations repeatedly find that adding a generation objective leaves understanding flat. Joint-training studies cannot settle the disagreement: with overlapping supervision, a gain cannot be attributed to the architecture rather than the data. To further investigate the relationship between the two directions in UMMs, we separate them by construction. A novel visual entity, a rendered 3D asset paired with a pseudo-word screened for absence from the frozen model's behavior, is bound through exactly one task direction, and the untrained direction is then measured. We find that the channel is real in both directions, but the directions differ in kind: generation training installs a name the model can only match among candidates; understanding training installs one it can also produce. What governs cross-task usability is where the binding enters the shared computation. An alignment probe predicts export across 36 configurations (Spearman $ρ= +0.68$). That objective's alignment term, maximized in closed form over activations with every weight frozen, makes a concept drawable when injected at layer 7 of 28 and is indistinguishable from the base model from layer 14 on, while the weight-based version of the same edit peaks at layers 10-14. In an observational series of four models, this window appears only where the understanding pathway is a semantic vision encoder, suggesting that unified weights are not enough: the two directions must share a semantic format at the entry point. Exploiting the rule, a mid-stack alignment objective acquires the concept for a $0.1\%$ relative loss of the model's general text-to-image ability, against $41\%$ for the standard generative route. Our code is at https://github.com/Zane-ZYQiu/entry-point-umm.
Out-of-distribution (OOD) detection remains challenging for image classifiers, especially when near-OOD samples lie close to in-distribution (ID) class boundaries. Recent vision-language detectors improve OOD detection through class semantics, local prompting, or LLM-generated outlier concepts, but seldom use language as explicit boundary evidence between confusing ID classes. We propose Pairwise Witness Local Rejection (PWLR), which uses an MLLM offline to describe visible local cues that favor one ID class over a specific rival class. These cue phrases are then screened with ID-only data under a frozen vision-language backbone, so that only reliable local verifiers are kept. At inference, PWLR first retains a small set of globally plausible classes, then checks whether any of them is locally supported against its most relevant rivals, and finally combines this pairwise local evidence with the global class score through calibration. Experiments on ImageNet-100 far-OOD, cleaner/challenging OOD and near-OOD benchmarks show that PWLR consistently improves strong vision-language baselines across multiple backbones. Source code will be released.
Wafa Al Ghallabi, Ritesh Thawkar, Sara Ghaboura +6cs.CV
Magnetic Resonance Imaging (MRI) interpretation is fundamental to clinical decision-making, requiring radiologists to integrate multi-view anatomical planes across sequential timepoints while precisely localizing interval changes. However, existing vision-language benchmarks remain confined to single-timepoint, single-view interpretation, failing to capture the temporal-spatial reasoning essential to radiologic practice. We introduce the Time-Aware Multi-View MRI Benchmark, an evaluation framework unifying multi-view anatomical input, temporal reasoning across longitudinal scans, and structured localization guidance. The benchmark comprises 3,920 expert-verified question-answer pairs derived from 890 patients across over 3,200 longitudinal MRI timepoints, drawn from seven clinical cohorts covering glioblastoma, neurodegeneration, vestibular schwannoma, and brain metastases, in open-ended, multiple-choice, and binary formats, requiring models to identify anatomical regions of maximal change, characterize progression across sequences and views, and provide structured guidance specifying boundaries, imaging features, and confounders. Experiments across 16 vision-language models reveal moderate temporal alignment but systematic failure on change direction recognition and volumetric quantification, while multi-view inputs improve spatial localization yet degrade temporal reasoning in compact architectures. Our benchmark provides a systematic framework for evaluating progression tracking, interval change localization, and temporal ordering, which are essential for clinical deployment. Code, evaluation splits, and the dataset are available at: https://github.com/wafaAlghallabi/Time-Aware-MRI.
As Large Vision-Language Models increasingly aim to integrate visual generation and understanding within a single parameter space, evaluating such structural unification in a cohesive manner remains a critical challenge. Current evaluation protocols predominantly treat generative and discriminative capabilities as separate tasks, leaving a gap in system-level evaluation for unified multimodal models (UMMs). In this work, we propose Self-Generative-Understanding (SGU), a novel, annotation-free evaluation framework that probes the integrated capabilities of unified models through a semantic closed-loop challenge. Without requiring new annotations, SGU leverages the dual understanding-and-generation abilities of UMMs by asking them to first perceive an image and produce a textual description, subsequently reconstruct a visual context based on that description, and finally perform reasoning over the self-generated output. This pipeline provides a zero-cost testbed that yields an integrated performance score specifically tailored for evaluating UMMs as unified systems. Extensive experiments show that even high-performing UMMs often struggle to reason over their own generated contexts, revealing limitations that are not captured by separate evaluations of understanding or generation alone. Our work provides a complementary holistic evaluation framework and offers a foundation for benchmarking the development of next-generation unified multimodal models.
Humans converge on shared names for novel, hard-to-describe objects through repeated interaction, a process psycholinguists call lexical entrainment. Leading vision-language models fail at this: recent empirical work documents that they do not shorten references, reuse successful expressions, or maintain stable pact state across turns. We present a framework that addresses the gap by externalizing pact state into three explicit, inspectable sets of referent-object bindings ($Γ, Ξ, Ω$), updated by a dynamic-semantics context-change rule. The symbolic layer sits on top of a lightweight perceptual-alignment pipeline that grounds noisy human referring expressions in crowd-sourced imagery via SIFT homographies and the Universal Quality Index. Evaluated on the Stanford Repeated Reference Game corpus (over 15{,}000 director-matcher utterances on abstract tangram stimuli), the framework places the correct target in its top-5 hypothesis set 83.56% of the time from a single director utterance. Human matcher top-1 accuracy on the same corpus is approximately 77-80%. We also report results on a held-out condition in which obvious tangram-adjacent images are excluded from the retrieved set, which provides a more conservative measurement of the grounding signal. Ablations isolate the contribution of each component: SIFT alignment, UQI, query preprocessing, and image augmentation. The central contribution is the combination: a transparent, auditable symbolic layer that recovers the structure of lexical entrainment turn by turn, paired with a perceptual channel whose behavior can be examined ablation by ablation. We also discuss in detail what the framework does not do. It is not interactive, it does not close the loop with the director, and its retrieval-driven perceptual channel is vulnerable to a class of leakage effects that we quantify and bound rather than wave away.
Metaphor enables the understanding of abstract concepts through cross-domain mappings while conveying affective attitudes. In multimodal scenarios, visual and textual information jointly construct Target--Source mappings, requiring both conceptual understanding and cross-modal reasoning. However, existing benchmarks mainly evaluate metaphor understanding through isolated subtasks and lack evidence-grounded explanations, making it difficult to assess whether models establish mappings grounded in visual and textual cues.To address these limitations, we introduce M$^3$R-Bench, a unified and evidence-grounded benchmark containing 1,000 image--text instances with human-verified annotations. Guided by Conceptual Metaphor Theory and theories of nonliteral language understanding, M$^3$R-Bench provides joint annotations for metaphor occurrence, Target--Source mapping, sentiment, and stage-wise explanations following ``evidence identification--mapping establishment--sentiment inference.''Evaluations on M$^3$R-Bench reveal that existing models often overlook visual evidence, rely on superficial textual cues, and produce inaccurate Target--Source mappings, exposing a cross-modal evidence--mapping mismatch. To address this mismatch, we propose M$^3$R-Reasoner, which combines curriculum-based reasoning supervision with task-aware reinforcement learning to align model reasoning with metaphor interpretation. Experiments show that, with only an 8B-parameter backbone, M$^3$R-Reasoner outperforms larger proprietary MLLMs across four unified-task metrics and improves Visual Evidence and Sentiment Justification scores over GPT-5.5 by 28.45 and 30.11 points, respectively, while surpassing Claude-Sonnet-4.6 by 8.00 points in mean rubric score. The dataset and code are available at https://github.com/hongshi4/M3R-Bench.
Label-free reliability for vision-language models rests on invariance: perturb the input and a faithful reader's answer should not change. This has a known blind spot, a systematic misreading survives the perturbation and gets certified wrong, which we show is computable, not just real: an error is invisible to an edit exactly when the two commute, so the errors a suite cannot reach form its joint centralizer, a set that shrinks as edits are added and can be written down rather than guessed at. We act on the complementary relation, equivariance: edit a figure's data and the correct answer must change by a computable amount. Two matched edits are provably complete for affine reading errors; no suite of swap edits is complete for label permutations, and cyclic relabeling closes most of that gap. We instantiate the theory as the Equivariance-Consistency Score, a label-free, training-free detector, and release REND-EQUIV, pairing matched invariance and equivariance sets over identical data. The predicted ordering holds across three models and a hand-labeled population immune to the one circularity in how it is selected; a second invariance-family method confirms the blind spot belongs to the relation, not to any implementation; and cyclic relabeling delivers its predicted gain on a matched real sample. The same characterization explains a reported inversion of this ordering in the classifier metamorphic-testing literature: detectability is a joint property of the relation and the fault class, never of the relation alone.
Training-free open-vocabulary segmentation remains limited by a missing inference abstraction. Frozen vision-language features are produced at patch level, yet dense prediction requires a unit that simultaneously governs feature interaction, spatial support, contextual recovery, and retrieval-based correction. We present SCI-CLIP, a segment-centric inference framework built around the principle that the same region abstraction should organize all stages of dense open-vocabulary prediction. SCI-CLIP first induces a region-consistent interaction graph over frozen visual tokens, then reconstructs dense features by propagating values over this graph, augmenting them with selective cross-window support only where local evidence is insufficient. The same segment abstraction is subsequently used to construct and query an offline reference memory, aligning exemplar retrieval with the units on which prediction is made. SCI-CLIP turns frozen CLIP-style features into spatially coherent, context-aware, and retrieval-compatible dense predictions without any training. SCI-CLIP consistently improves the structural quality of dense predictions, the robustness of contextual reasoning, and the alignment of exemplar-based correction, yielding stronger open-vocabulary segmentation across eight benchmarks. Project code is available at: https://github.com/mzamini92/SCICLIP.
Yang Yang, Qinyu Zhao, Mouxiang Chen +5cs.CV cs.CL
Existing scaling strategies for Multimodal Large Language Models (MLLMs) typically expand either model parameters or sequential inference computation, incurring substantial memory or latency overhead. More importantly, most existing methods fail to alter the rigid, fixed computation allocation between the Vision Transformer and the Large Language Model components, limiting task-specific optimization. To address this, we introduce the Parallel Vision-Language (ParVL) scaling framework for MLLMs, which scales parallel computation by reusing the existing ViT and LLM backbone parameters across multiple vision and language branches. This framework raises a central question: given a fixed backbone parameter budget, how should additional shared-backbone computation be allocated between the vision and language modalities? We instantiate each parallel computational stream with branch-specific prefix parameters over a shared backbone, and train the entire model end-to-end via full-parameter supervised fine-tuning on roughly 13B tokens. We systematically study the computation-allocation trade-off between the ViT encoder and LLM decoder. ParVL improves overall multimodal performance over same-recipe single-branch baselines, and the best evaluated vision--language allocation varies across tasks. Code is available at https://github.com/YangYangGirl/ParVL.
Human-interpretable computer-aided diagnosis is crucial for clinical decision making. Concept-based models excel by providing transparent reasoning and enabling post-hoc, clinician-in-the-loop interventions. However, their rigid dataset-specific adaptation inherently restricts cross-site generalization. Applying them across diverse modalities, such as dermoscopic and clinical photographs, is challenging due to heterogeneous concept taxonomies varying in availability, granularity, and semantics across cohorts. Consequently, adapting Foundation Vision-Language Models (FVLMs) demands costly label engineering and repeated post-training. Existing intervention mechanisms remain rigidly tied to predefined concepts, lacking adaptability and hindering scalable dermatology CAD deployment. To address these bottlenecks, we propose UniCon, an open-linguistic unified concept learning framework for multimodal interpretable vision-language diagnosis. UniCon resolves these challenges through three contributions: (1) A shared semantic representation space via a unified concept prototype codebook, seamlessly coordinating heterogeneous concept systems across modalities without dataset-specific retraining. (2) Open-linguistic based multi-faceted semantic specifications to overcome sparse textual label limitations, improving boundary sensitivity in uncertain clinical contexts. (3) A robust, cross-site adjustable intervention interface powered by reliability-gated bottleneck aggregation, enabling consistent reasoning and transferable clinician corrections. Extensive experiments demonstrate that beyond securing top-tier diagnostic accuracy, UniCon successfully bridges disparate clinical taxonomies, unlocking unprecedented cross-site intervention capabilities. Code is available at https://github.com/wuchengyu123/UniCon.
Vision-language MoE batches contain different numbers of image and text tokens. Image resolution, image count, tiling, and prompt length all change this token mix. We call the standard token-level Switch auxiliary loss Std-Aux. Std-Aux balances only the mixed load, so large image and text load errors can cancel at one mix. On our main model, the same trained router shows more than a fivefold change in load imbalance across image resolutions. We hold the image and text load profiles fixed and derive the exact load curve as the token mix varies. The image-text load gap controls sensitivity to the token mix. Physical preprocessing can also change the conditional profiles. The fixed-profile law excludes such changes. To design a remedy, we examine the router input structure. Image and text occupy distinct regions, while visual tokens group strongly by source image. The modality boundary motivates separate image and text terms. The image boundary motivates one equal-weight routing instance per image. ReBA, or Relax Within, Balance Across, implements both choices. Across four split backbones, ReBA lowers load on every reported benchmark input while keeping mean task accuracy comparable to Std-Aux. ReBA also lowers average load over the tested range and worst physical load under resolution and tiling shifts. Code is available at https://github.com/ZiangWu-77/ReBA.
To maintain common ground in cooperative conversation, humans iteratively update their beliefs as conversation participants share new information; participants who are epistemically vigilant detect when new information conflicts with prior beliefs and take steps to repair these conflicts. In order for AI systems to serve as reliable partners in complex cooperative tasks, they must similarly weigh incoming information against their own private evidence and shared context and appropriately surface inconsistencies when they arise. To measure the epistemic vigilance of vision-language models in cooperative settings, we present an information-asymmetric, dialog-based "spot-the-difference" task. Two models are privately shown one image each, and must determine through conversation whether the images are identical or, if not, identify the difference. Models routinely fail at this: they frequently overlook key evidence in their private image in favor of agreeing with their conversational partner, even when their agreement is unwarranted. We relate these violations of epistemic vigilance to the broader behavior of sycophancy, which manifests itself in cooperative goal-oriented dialog as over-accommodation and weak evidential grounding. Our results show that model steering to reduce sycophancy with a vector learned from task-agnostic sycophancy examples can reduce epistemic vigilance-related errors, making models more faithful reporters of their evidence, and in turn, more reliable partners in information-asymmetric cooperative tasks.
Monocular depth estimation (MDE) faces challenges with non-Lambertian surfaces and adverse weather conditions due to the visual ambiguities inherent in single-image limited information. Existing works address them in isolation via image inpainting or augmentation, yielding limited robustness gains. Language, as a powerful complementary modality to vision, is demonstrated to enhance the visual perception capabilities of vision-language models (VLMs) via detailed long captions. However, prior language-integrated MDE methods fail to fully harness this potential due to short text input with limited information, coarse global text feature learning, and limited language guidance during depth decoding. To address these limitations, we propose CapDepth, a novel framework for robust MDE that leverages guidance from detailed long captions to alleviate visual ambiguities in both challenging scenarios. First, we design a detailed long caption input template that explicitly conveys rich spatial relationships among multiple atom sentences. Second, a dynamic caption encoder is introduced to extract fine-grained depth-relevant text features via progressive masked attention. Finally, we propose a text-adaptive decoder that guides enhanced depth decoding with text features via stable adaptive layer normalization. Extensive experiments validate the efficacy of CapDepth, which outperforms state-of-the-art methods, achieving depth error reductions of 25.0% on non-Lambertian surfaces and 22.0% under adverse weather conditions.
Multimodal Large Language Models (MLLMs) have achieved strong performance on a wide range of vision-language tasks, but often fail under imperfect or shifted contexts. A reliable MLLM should refuse truly out-of-context (OOC) questions with subject-level context shifts while still answering shifted in-context (Shifted IC) questions with non-subject context shifts. Existing benchmarks mainly target OOC or visually unanswerable questions, but overlook answerable Shifted IC cases and cover limited OOC shifts. To fill this gap, we present MMOOC, a large-scale benchmark for evaluating refusal and robust answering abilities of MLLMs. MMOOC contains over 41K image-question pairs, including answerable Shifted IC cases and unanswerable OOC cases, spanning three question formats, eight shift types and six visual scenarios, with data quality ensured through MLLM-based filtering and human verification. We evaluate model responses using Accuracy and Refusal Rate, and further introduce an LLM-as-a-Judge metric to assess the correctness of model reasoning. Experiments on diverse MLLMs show that current models still struggle to balance answer-ability and refusal under shifted contexts. We further analyze key failure patterns and show that post-training can improve robustness. MMOOC will be made publicly available.
Dense vision-language understanding, including object localization, region recognition, and open-vocabulary semantic segmentation, requires associating language concepts with spatially grounded visual regions. CLIP provides a strong foundation for these tasks by learning a shared image-text embedding space from large-scale contrastive pre-training. However, its image-level objective aligns text with a CLS-derived global representation, leaving local vision-language correspondence only indirectly constrained. Existing methods either introduce additional supervision, external models, or task-specific adaptation, while training-free approaches mainly recover dense responses from existing patch features without examining where local semantics become most accessible within CLIP. We introduce TraceCLIP, a training-free framework that recovers latent patch-level semantic evidence by isolating the patch-specific terms written into the CLS attention output. TraceCLIP further converts contribution-derived semantic responses into a semantic-geodesic topology gate that calibrates final-layer patch affinity for dense feature reconstruction. Diagnostic experiments show that these contribution features exhibit strong local semantic discrimination and text-conditioned spatial alignment. On eight zero-shot semantic segmentation benchmarks, TraceCLIP achieves gains of 1.3 to 4.5 points in average mIoU over the strongest prior training-free methods across both backbones and background settings, without additional training, external vision foundation models, or region-level supervision. More broadly, these findings suggest that spatially localized semantics may remain accessible within the internal construction of globally aligned representations.
Kimi Team, Tongtong Bai, Yifan Bai +399cs.CL cs.LG
We introduce Kimi K3, a 2.8T parameter Mixture-of-Experts model with 104 billion activated parameters, native vision capabilities, and a 1-million-token context window. Kimi K3 is built on Kimi Delta Attention and Attention Residuals, which improve information flow across sequence length and model depth. Together with Stable LatentMoE, which effectively activates 16 of 896 routed experts per token, and refined training and data recipes, these advances yield an approximately 2.5x improvement in overall scaling efficiency over Kimi K2. Post-training highlights reinforcement learning across general, agentic, and coding domains and multiple reasoning-effort levels, enabling compositional generalization and robust long-horizon execution. At 2.8T scale, Kimi K3 is supported by infrastructure advances in multiple areas: algorithm-system co-design for KDA, perfectly balanced expert-parallel training with efficient memory management, million-token agentic RL with persistent rollout and sandbox states, and deployment innovations. Extensive evaluations show that Kimi K3 achieves frontier-level performance across long-horizon coding, agentic, knowledge, reasoning, and vision tasks. While its overall performance still trails the most powerful proprietary models, namely Claude Fable 5 and GPT-5.6 Sol, Kimi K3 consistently outperforms other open and proprietary models evaluated in our suite. We release the full Kimi K3 model weights to facilitate future research and accelerate the broader deployment and adoption of frontier intelligence.
Image captioning is a primary task in vision--language research, yet assessing how faithfully a caption preserves image semantics without relying on reference captions remains unsettled. Prevailing evaluations rely on human-annotated references, whose content reflects annotator intent and captioning proficiency. In this paper, we study a reconstruction-based principle for caption evaluation: a caption is as good as its capacity to enable reconstruction of the original image. However, because captioning inherently compresses visual information, it is impossible to recover all details, and pixel-wise comparison between reconstructed and source images is neither feasible nor meaningful. Through our in-depth analysis of the nature of captions, whose fundamental purpose is to transmit the semantic content of an image, we propose a revised principle: a caption is as good as its capacity to enable a reconstruction that is semantically equivalent to the original. To assess semantic equivalence, we test whether the reconstruction matches the original image across a suite of downstream vision--language tasks, yielding a reference-free, task-conditioned caption score. We characterize component-dependent limitations and introduce the lower-cost Captioning Turing Test Dataset (CTTD) surrogate.
Collision anticipation in autonomous driving requires not only accurate early warnings but also interpretable reasoning about what risk factors are being tracked and how risk evolves over time. Existing methods fall short in this regard: feature-driven models are opaque, post-hoc explanations often lack fidelity, and concept-based methods are mostly designed for static recognition rather than dynamic driving scenes. We propose CARA (Concept-Aware Risk Attention), an intrinsically interpretable spatio-temporal framework for collision anticipation. CARA derives domain-grounded risk concepts from accident narratives, aligns them with video frames via vision-language similarity, and organizes them into evolving concept trajectories. These trajectories provide explicit risk evidence that guides spatial attention, temporal attention, and anticipation, allowing semantic concepts to directly influence both where the model attends and how it predicts risk over time. By treating semantic risk factors as dynamic intermediate evidence rather than auxiliary post-hoc explanations, CARA tightly couples interpretability with the predictive process. Extensive experiments on three benchmarks show that CARA consistently improves anticipation accuracy and warning earliness over strong baselines, while providing sparse and semantically grounded concept evidence.