Vision-Language Alignment

Open-vocabulary semantic segmentation requires adapting image-level vision-language models such as CLIP to dense pixel-level prediction, which is challenging due to the mismatch between hierarchical structure and semantic alignment in the embedding space. While recent works leverage hyperbolic geometry to model hierarchical relationships, they align embeddings across hierarchical levels but overlook semantic misalignment among embeddings within the same level. In this work, we propose HyRo, a hyperbolic fine-tuning framework that decouples hierarchical and semantic alignment in the Poincaré ball model. HyRo aligns hierarchical levels by adjusting the hyperbolic radius and refines semantic relationships through angular alignment using an orthogonal transformation that theoretically preserves the hyperbolic radius. Experiments on standard open-vocabulary semantic segmentation benchmarks demonstrate that HyRo achieves state-of-the-art performance over prior methods.

Chinese character categories are extremely large, and unseen characters frequently arise in open-world scenarios, making zero-shot Chinese character recognition an important yet challenging problem. Existing IDS-based retrieval methods usually encode a character image and its ideographic description sequence into a single global vector for matching. Although efficient, such holistic alignment often under-models local component differences. Moreover, directly introducing patch-token level fine-grained interaction suffers from both the noise of structural operators in IDS and the high cost of full-candidate retrieval.To address these issues, we propose a Global-Local Hierarchical Perception Network (GL-HPN), which jointly learns global and local representations of character images and IDS sequences within a unified cross-modal alignment framework. The global branch supports efficient coarse recall, while the local branch improves component-level discrimination through patch-token interaction. We further introduce a structure filtering mask to suppress structurally meaningful but visually non-entity IDS operators in local similarity aggregation. On top of this, we design a coarse-to-fine hierarchical inference strategy that performs global retrieval over the full candidate set and local reranking only on Top-$K$ candidates, followed by parameter-free multiplicative fusion of normalized posterior scores. Experimental results show that GL-HPN achieves competitive performance across multiple zero-shot splits, performs especially well under low-resource settings, and substantially reduces the inference cost of large-scale candidate retrieval.

Wireless capsule endoscopy (WCE) enables non-invasive visual assessment of the small bowel, but its clinical utility is constrained by the large volume of frames generated per examination and the difficulty of recognising subtle abnormalities under highly variable imaging conditions. Existing learning-based approaches for WCE are predominantly vision-only, often confined to narrow pathology sets, and show limited transfer across datasets and centres. To address these limitations, this study introduces CapCLIP, a domain-specific vision-language representation learning framework for WCE. CapCLIP aligns capsule endoscopy frames with clinically grounded textual descriptions derived from standardised nomenclature and pathology-aware caption templates, thereby learning embeddings that are both semantically informed and transferable. The proposed framework is evaluated against relevant open-source vision and vision-language foundation models under strict zero-shot conditions using unseen WCE datasets. Evaluation covers three downstream tasks: K-nearest neighbour classification, CLIP-style image-text classification, and text-to-image retrieval. Across these settings, CapCLIP consistently outperforms the compared baselines, with particularly strong gains in zero-shot image-text classification and cross-modal retrieval on out-of-distribution datasets. The results indicate that language-guided representation learning can improve both generalisation and semantic interpretability in WCE analysis. These findings position CapCLIP as a step toward foundation models tailored to capsule endoscopy and support the use of language-grounded WCE analysis.

Spatial intelligence in vision-language models (VLMs) attracts research interest with the practical demand to reason in the 3D world.Despite promising results, most existing methods follow the conventional 2D pipeline in VLMs and use pixel-aligned representations for the vision modality. However, correspondence-based models with implicit 3D scene understanding often fail to achieve spatial consistency, and representation-based models with 3D geometric priors lack efficiency in vision sequence serialization. To address this, we propose a Proxy3D method with compact yet comprehensive 3D proxy representations for the vision modality. Given only video frames as input, we employ semantic and geometric encoders to extract scene features and then perform their semantic-aware clustering to obtain a set of proxies in the 3D space. For representation alignment, we further curate the SpaceSpan dataset and apply multi-stage training to adopt the proposed 3D proxy representations with the VLM. When using shorter sequences for vision information, our method achieves competitive or state-of-the-art performance in 3D visual question answering, visual grounding and general spatial intelligence benchmarks.

Cameras and LiDAR degrade in rain, fog, and snow, while millimeter-wave radar remains largely unaffected. We align a radar encoder to frozen SigLIP vision embeddings and decode structured scene captions through a frozen vision-language model (VLM) with approximately 7M trainable parameters. On K-RADAR with held-out fog, light snow, and heavy snow sequences, all radar configurations outperform a camera baseline that collapses to over 90% hallucination. We identify a token-norm mismatch as the dominant failure mode when bridging radar to a frozen VLM and show that projector-output LayerNorm resolves it. Analysis of encoder complexity, caption format, and pooling strategy reveals tradeoffs that inform future radar-VLM pipeline design.

How to achieve vision-language (VL) tracking using natural language descriptions from a video sequence without relying on any bounding-box ground truth? In this work, we achieve this goal by tackling self-supervised VL tracking, which aims to evaluate tracking capabilities guided by natural language descriptions. We introduce \tracker, a novel self-supervised VL tracker that is capable of tracking any referred object by a language description. Unlike traditional methods that equally fuse all language and visual tokens, we propose an efficient Dynamic Token Aggregation Module, which treats each visual token unequally. The module consists of three main steps: i) Based on an anchor token, it selects multiple important target tokens from the template frame. ii) The selected target tokens are merged according to their attention scores and aggregated into the language tokens, thereby eliminating redundant visual token noise and enhancing semantic alignment. iii) Finally, the fused language tokens serve as guiding signals to extract potential target tokens from the search frame and propagate them to subsequent frames, enhancing temporal prompts and encouraging the tracker to autonomously learn instance tracking from unlabeled videos. This new modeling approach enables the effective self-supervised learning of language-guided tracking representations without the need for large-scale bounding box annotations. Extensive experiments on VL tracking benchmarks show that {\tracker} surpasses SOTA self-supervised methods.

Contrastive language-image pretraining (CLIP) suffers from two structural weaknesses: the symmetric InfoNCE loss discards the relative ordering among unmatched in-batch pairs, and global pooling collapses the visual representation into a semantic bottleneck that is poorly sensitive to fine-grained local structure. RANKCLIP partially addresses the first issue with a list-wise Plackett-Luce ranking-consistency loss, but its model is strictly first-order and inherits the second weakness untouched. We propose DINORANKCLIP, a pretraining framework that addresses both jointly. Our principal contribution is injecting a frozen DINOv3 teacher into the contrastive trunk through a dual-branch lightweight student and a multi-scale fusion module with channel-spatial attention, a self-attention refiner, and a conflict-aware gate that preserves the cross-modal alignment up to first order. Complementarily, we introduce a high-order Plackett-Luce ranking model in which the per-position utility is augmented with attention-parameterised pairwise and tuple-wise transition terms; the family contains CLIP and RANKCLIP as nested zero-order and first-order special cases, and the optimal order on every benchmark is $R^*=3$. The full empirical study -- order sweep, Fine-grained Probe on five datasets, four-node Modality-Gap analysis, six-variant Fusion ablation -- fits in 72 hours on a single eight-GPU H100 node and trains entirely on Conceptual Captions 3M. DINORANKCLIP consistently outperforms CLIP, CyCLIP, ALIP, and RANKCLIP under matched compute, with the largest relative gains on the fine-grained and out-of-distribution evaluations that most directly stress local structural reasoning.

Vehicle Re-identification (Re-ID) aims to retrieve the most similar image to a given query from images captured by non-overlapping cameras. Extending vehicle Re-ID from image-only queries to text-based queries enables retrieval in real-world scenarios where only a witness description of the target vehicle is available. In this paper, we propose PFCVR, a Part-level Fine-grained Cross-modal Vehicle Retrieval model for text-to-image vehicle re-identification. PFCVR constructs locally paired images and texts at the part level and introduces learnable part-query tokens that aggregate both part-specific and full-sentence context before aligning with visual part features. On top of this explicit local alignment, a bi-directional mask recovery module lets each modality reconstruct its masked content under the guidance of the other, implicitly bridging local correspondences into global feature alignment. Furthermore, we construct a new large-scale dataset called T2I-VeRW, which contains 14,668 images covering 1,796 vehicle identities with fine-grained part-level annotations. Experimental results on the T2I-VeRI dataset show that PFCVR achieves 29.2% Rank-1 accuracy, improving over the best competing method by +3.7% percentage points. On the newly proposed T2I-VeRW benchmark, PFCVR achieves 55.2% Rank-1 accuracy, outperforming a comprehensive set of recent state-of-the-art methods. Source code will be released on https://github.com/Event-AHU/Neuromorphic_ReID

Vision-Language Models (VLMs) increasingly power high-stakes applications, from medical imaging to autonomous systems, yet they routinely hallucinate, confidently describing content not present in the input. We investigate the root causes of these failure modes with a mechanistic analysis focusing on the decoder-based VLMs. We trace these failure modes to a geometric over-alignment: to bridge the modality gap required by attention mechanisms, decoder-based VLMs over-align visual embeddings with the text manifold, injecting a statistical linguistic bias that systematically overshadows fine-grained visual evidence. While prior work either aggressively closes this gap or suppresses hallucinations through expensive black-box decoding strategies, none addresses the underlying geometric cause. We provide the first quantitative characterization of this over-alignment, demonstrating that linguistic bias concentrates in the top principal components of a universal, dataset-agnostic text subspace. Building on this insight, we propose two complementary remedies: a training-free inference strategy and a bias-aware fine-tuning paradigm, both of which explicitly project out this subspace from visual representations. Our methods significantly reduce hallucinations across POPE, CHAIR, and AMBER benchmarks, and improve CLAIR scores on long-form captioning tasks, with the training-free variant adding no computational overhead over the base model.

Lifelong person re-identification (LReID) aims to train a generalizable model with sequentially collected data. However, such models often suffer from semantic drift, limited adaptability, and catastrophic forgetting as new domains emerge. Existing exemplar-free approaches largely rely on visual-only distillation or parameter regularization, while overlooking the potential of auxiliary modalities, such as text, to preserve semantic stability and enable incremental plasticity. We observe that the frozen text encoder in pretrained vision-language models can serve as a stable semantic anchor across domains. To decouple the roles of vision and text, we propose Prompt-Anchored vision-text Distillation (PAD), an asymmetric vision-text framework for semantic alignment and cross-domain generalization. On the textual side, we distill prompts to preserve vision-text alignment under a fixed semantic space, acting as a global semantic reference rather than a dominant learning signal. On the visual side, an EMA-based teacher with an adaptive prompt pool enables domain-wise adaptation by allocating new slots while freezing past ones. Extensive experiments show that PAD substantially outperforms state-of-the-art methods across seen and unseen domains, achieving a strong balance between stability and plasticity. Project page is available at https://github.com/zu-zi/PAD.

Image-Text-Matching (ITM) is one of the defacto methods of learning generalized representations from a large corpus in Vision and Language (VL). However, due to the weak association between the web-collected image-text pairs, models fail to show a fine-grained understanding of the combined semantics of these modalities. To address this issue we propose Hard Negative Captions (HNC): an automatically created dataset containing foiled hard negative captions for ITM training towards achieving fine-grained cross-modal comprehension in VL. Additionally, we provide a challenging manually-created test set for benchmarking models on a fine-grained cross-modal mismatch task with varying levels of compositional complexity. Our results show the effectiveness of training on HNC by improving the models' zero-shot capabilities in detecting mismatches on diagnostic tasks and performing robustly under noisy visual input scenarios. Also, we demonstrate that HNC models yield a comparable or better initialization for fine-tuning

Open-vocabulary object detection with vision-language models (VLMs) such as Grounding DINO suffers from performance degradation under test-time distribution shifts, primarily due to semantic misalignment between text embeddings and shifted visual embeddings of region proposals. While recent test-time adaptive object detection methods for VLM-based either rely on costly backpropagation or bypass semantic misalignment via external memory, none directly and efficiently align text and vision in a training-free manner. To address this, we propose Reward-Guided Semantic Evolution (RGSE), a training-free framework that directly refines the text embeddings at test time. Inspired by evolutionary search, RGSE treats text embedding adaptation as a semantic search process: it perturbs text embeddings as candidate variants, evaluates them via cosine similarity with current and historical high-confidence visual proposals as a reward signal, and fuses them into a refined embedding through reward-weighted averaging. Without any backpropagation, RGSE achieves state-of-the-art performance across multiple detection benchmarks while adding minimal computational overhead. Our code will be open source upon publication.

Open-vocabulary object detection aims to recognize objects from an open set of categories, which leverages vision-language models (VLMs) pre-trained on large-scale image-text data. The cooperative paradigm combines an object detector with a VLM to achieve zero-shot recognition of novel objects. However, VLMs pre-trained on full images often struggle to capture local object details, limiting their effectiveness when applied to region-level detection. We present Decoupled Adaptivity Training (DAT), a self-supervised fine-tuning approach to improve VLMs for cooperative model-based object detection. Given a cooperative model consists of a closed-set detector and a VLM, we first construct a region-aware pseudo-labeled dataset using a pre-trained closed-set object detector, in which regions corresponding to novel objects may be present but remain unlabeled or mislabeled. We then fine-tune the visual backbone of the VLM in a decoupled manner, which enhances local feature alignment while preserving global semantic knowledge via weight interpolation. DAT is a plug-and-play module that requires no inference overhead and fine-tunes less than 0.8M parameters. Experiments on the COCO and LVIS datasets show that DAT consistently improves detection performance on both novel and known categories, establishing a new state of the art in cooperative open-vocabulary detection.

Vision-Language Models (VLMs) have broad potential in privacy-sensitive domains such as healthcare and finance, yet strict data-sharing constraints render centralized training infeasible. Federated Learning mitigates this issue by enabling decentralized training, but practical deployments face challenges due to client heterogeneity in computational resources, application requirements, and model architectures. Under extreme model and data heterogeneity, replacing parameter aggregation with preference-based collaboration offers a more suitable interface, as it eliminates the need for direct parameter or data exchange. Motivated by this, we propose MoR, a federated alignment framework that combines GRPO with Mixture-of-Rewards for heterogeneous VLMs. In MoR, each client locally trains a reward model from local preference annotations, capturing specific evaluation signals without exposing raw data. To combine these heterogeneous supervision signals, MoR introduces a Mixture-of-Rewards mechanism with learned routing, which adaptively fuses client reward models according to the input and alignment objective. The server then optimizes a base VLM using GRPO with a KL penalty to a reference model, enabling preference alignment without requiring client models to share architectures or parameters. Experiments on diverse public vision-language benchmarks demonstrate that MoR consistently outperforms federated alignment baselines in generalization and cross-client adaptability. Our approach provides a scalable solution for privacy-preserving alignment of heterogeneous VLMs under federated settings.

Video Temporal Grounding (VTG) faces a cross-modal semantic gap that often leads to background features being incorrectly aligned with the query, while directly matching the query to moments results in insufficient discriminability and consistency of temporal semantics. To address this issue, we propose MLLM-Assisted Semantic-Relational Consistent Alignment (MASRA), a training-time MLLM-based optimization framework for VTG. MASRA leverages an MLLM during training to produce two forms of textual priors, namely event-level descriptions with temporal spans and clip-level captions, and instantiates two MLLM-assisted alignments. Event Semantic Temporal Alignment (ESTA) aligns temporal context with event semantics to explicitly strengthen the correspondence between semantics and temporal events and improve span-level separability. Local Relational Consistency Alignment (LRCA) constructs a textual relation matrix derived from clip-level captions and aligns it with the temporal feature similarity matrix in the model, enhancing temporal consistency while capturing local structural information. MASRA includes two simple supporting modules, semantic-guided enhancement and second-order relational attention, to better utilize the learned semantic context and relational structure. Moreover, we introduce Decoupled Alignment Interaction (DAI) with a context-aware codebook to adaptively absorb query-irrelevant semantics and alleviate the cross-modal gap. The MLLM is only invoked during training and is not used at inference. Extensive experiments show that MASRA outperforms existing methods, and ablation studies validate its effectiveness.

Long video understanding with multimodal language models suffers from three compounding bottlenecks: heavy decode cost to obtain dense RGB frames, token growth with frame count, and weak motion perception under sparse keyframe sampling. Existing remedies either prune visual tokens after decoding, which leaves the expensive RGB pipeline untouched, or they discretise codec motion vectors with a generative tokeniser, which couples motion modelling to a heavy pre-training stage. In this paper we present HIMMEL, a hierarchical video-language framework that allocates semantic and motion capacity along separate paths. A small set of sparse semantic anchor frames is routed to the expensive host ViT to ground object identity and scene layout, while the far denser inter-frame intervals are encoded by a lightweight compressed-domain tri-stream adapter that distils motion evidence from motion-vector maps, residual maps, and an I-frame context branch into aligned motion tokens. These tokens are injected into the LLM via a differentiable placeholder mechanism, after a dedicated Stage 1 contrastive alignment that places the motion representation in a geometry compatible with the frozen visual backbone. We further show that an InfoNCE alignment objective beats MSE regression by +1.5 pp. On Video-MME, HIMMEL surpasses the dense 32-frame baseline by +2.3 pp (61.2 to 63.5%) while using 3.6 times fewer context tokens and running end-to-end in 1.03 s per question instead of 2.75 s. Switching to a stronger Qwen3-VL-8B host pushes HIMMEL to 64.9%, narrowing the gap to much larger proprietary systems while keeping inference on a single NVIDIA H20 GPU. Extensive ablations across stream composition, motion-encoder family, fusion mode, alignment objective, anchor count, LoRA rank, and video duration confirm that the full tri-stream contributes to the observed gains, and that the benefit grows with video length.

Zero-shot recognition aims to classify an image by selecting the most compatible label description from a set of candidate classes without any task-specific supervision. In fine-grained settings, however, the relevant evidence often lies in localized parts, attributes, or textures rather than in the full image, making whole-image alignment suboptimal. Recent localized visual-text alignment methods address this by comparing class descriptions with multiple image regions, but they typically rely on large sets of random or redundant crops, increasing inference cost and introducing many highly redundant or weakly relevant candidates. Moreover, introducing semantic guidance too early can create an error-amplifying feedback process in which inaccurate intermediate predictions bias later localization and reinforce subsequent mistakes; we refer to this failure mode as the prediction loop. We propose LAGO (LAnguage-Guided adaptive Object-region focus), a framework for efficient and robust zero-shot localized visual-text alignment. LAGO first performs class-agnostic object-centric candidate discovery to obtain a stable visual initialization, and then applies adaptive language-guided refinement with the strength of semantic guidance controlled by intermediate confidence. It further combines object-level, contextual, and full-image evidence through an effective object-context dual-channel aggregation strategy. Extensive experiments show that LAGO consistently achieves state-of-the-art performance on standard zero-shot benchmarks and challenging distribution-shift settings, while requiring substantially fewer candidate regions at inference time.

A hierarchical cross-modal fusion model is proposed for vision-language question answering (VLQA) in industrial robotics, targeting the challenges of semantic ambiguity, complex environmental layouts, and domain-specific terminology common in modern manufacturing. The framework integrates advanced object detection, multi-scale visual encoding, syntactic parsing, and task-aware semantic attention to unite vision and language signals into a joint reasoning space. Region-based deep networks extract visual features, weighted embeddings aggregate, and recurrent neural parsing encodes sentence structures. Through fine-grained semantic alignment driven by adaptive fusion and cross-attention mechanisms, the system can handle operational queries, instruction steps, and anomaly detection with higher reliability. Compared to the existing VLQA benchmarks, validation experiments conducted on the IVQA and RIF benchmarks indicate improvements in semantic alignment, Top-1 accuracy, and robustness to ambiguous or procedural task queries. Ablation studies further quantify the impact of each architectural module, confirming the necessity of multi-level feature integration and context-driven gating for dependable industrial deployment. The technical advancements reported here provide core methodologies to improve the interpretability and operational effectiveness of industrial robots faced with diverse human-robot interaction tasks.

Vision-Language Models (VLMs) have enhanced traditional LLMs with visual capabilities through the integration of vision encoders. While recent works have explored various combinations of vision encoders and LLMs, there still lacks a principled understanding of what makes a vision encoder suitable for VLM alignment. In this paper, we systematically investigate this question via comprehensive experiments on a curated collection of 19 pre-trained vision encoders from diverse sources. We first demonstrate that common practices, such as choosing encoders with the largest size or highest zero-shot accuracy, consistently fail to identify optimal models. In fact, these metrics show only weak to moderate correlation with VLM performance. This intriguing finding begs a fundamental question: What factors of vision-encoders matter in VLM? Through comprehensive analysis, we identify that the structural similarity across modalities plays a crucial but previously overlooked role in vision-encoder selection, which we measure using the Gromov-Wasserstein distance as a proxy. From a theoretical perspective, we show that the learnability of cross-modality mapping can be provably associated with the Gromov-Wasserstein distance. Empirical verification on 60+ full VLM training runs shows that our proposed inference-only metric performs significantly better than alternative model selection strategies and exhibits a much stronger correlation with final VLM performance, thereby enabling efficient and effective prediction of VLM performance before full training.

Temporal Video Grounding (TVG) aims to localize temporal moments in an untrimmed video that semantically correspond to given natural language queries. Recently, Graph Convolutional Networks (GCN) have been widely adopted in TVG to model temporal relations among video clips and enhance contextual reasoning by constructing clip-level graphs. Despite their effectiveness, existing GCN-based TVG methods encounter three critical bottlenecks: 1) Most methods construct graph nodes using either static or dynamic features alone, resulting in incomplete visual representation and overlooking complementary semantics, 2) Most methods construct temporal graphs in a query-agnostic manner, leading to inefficient feature interaction within the temporal graph representation, and 3) Most methods often suffer from a single-granularity semantic matching, while direct training on complex temporal localization task may lead to slow convergence and suboptimal precision. To address these challenges, we propose Static and Dynamic Graph Alignment Network (SDGAN). First, SDGAN jointly exploits static and dynamic visual features to construct two complementary temporal graphs and performs Position-wise Nodes Alignment, enabling more expressive and robust visual representation. Second, SDGAN introduces Query-Clip Contrastive Learning and Adaptive Graph Modeling to explicitly align visual clips with their corresponding textual queries, yielding query-aware visual representations. Third, SDGAN incorporates multi-granularity temporal proposals within Progressive Easy-to-Hard Training Strategy, effectively bridging coarse-grained semantic localization and fine-grained temporal boundary refinement. Extensive experiments on three benchmark datasets demonstrate that SDGAN achieves superior performance across complex TVG scenarios. Codes and datasets are available at https://github.com/ZhanJieHu/SDGAN.

The visual modality of vision-language models (VLMs) is an underexplored attack surface for bypassing safety alignment. We introduce four jailbreak attacks exploiting the vision component: (1) encoding harmful instructions as visual symbol sequences with a decoding legend, (2) replacing harmful objects with benign substitutes (e.g., bomb -> banana) then prompting for harmful actions using the substitute term, (3) replacing harmful text in images (e.g., on book covers) with benign words while visual context preserves the original meaning, and (4) visual analogy puzzles whose solution requires inferring a prohibited concept. Evaluating across six frontier VLMs, our visual attacks bypass safety alignment and expose a cross-modality alignment gap: text-based safety training does not automatically generalize to harmful intent conveyed visually. For example, our visual cipher achieves 40.9% attack success on Claude-Haiku-4.5 versus 10.7% for an equivalent textual cipher. To further our insight into the attack mechanism, we present preliminary interpretability and mitigation results. These findings highlight that robust VLM alignment requires treating vision as a first-class target for safety post-training.

Radiology report generation (RRG) has emerged as a promising approach to alleviate radiologists' workload and reduce human errors by automatically generating diagnostic reports from medical images. A key challenge in RRG is achieving fine-grained alignment between complex visual features and the hierarchical structure of long-form radiology reports. Although recent methods have improved image-text representation learning, they often treat reports as flat sequences, overlooking their structured sections and semantic hierarchies. This simplification hinders precise cross-modal alignment and weakens RRG accuracy. To address this challenge, we propose RIHA (Report-Image Hierarchical Alignment Transformer), a novel end-to-end framework that performs multi-level alignment between radiological images and their corresponding reports across paragraph, sentence, and word levels. This hierarchical alignment enables more precise cross-modal mapping, essential for capturing the nuanced semantics embedded in clinical narratives. Specifically, RIHA introduces a Visual Feature Pyramid (VFP) to extract multi-scale visual features and a Text Feature Pyramid (TFP) to represent multi-granularity textual structures. These components are integrated through a Cross-modal Hierarchical Alignment (CHA) module, leveraging optimal transport to effectively align visual and textual features across various levels. Furthermore, we incorporate Relative Positional Encoding (RPE) into the decoder to model spatial and semantic relationships among tokens, enhancing the token-level alignment between visual features and generated text. Extensive experiments on two benchmark chest X-ray datasets, IU-Xray and MIMIC-CXR, demonstrate that RIHA outperforms existing state-of-the-art models in both natural language generation and clinical efficacy metrics.

In recent years, Multimodal Large Language Models (MLLMs) have achieved remarkable progress on a wide range of multimodal benchmarks. Despite these advances, most existing benchmarks mainly focus on single-image or multi-image comprehension. In real-world scenarios such as document reading, information is often presented as interleaved multimodel contexts. This requires MLLMs not only to recognize the content of individual images, but also to identify relevant textual and visual evidence, establish fine-grained alignments between them, and reason over these aligned signals in interleaved contexts based on contextual evidence. However, there is still a lack of systematic benchmarks for quantifying the fine-grained understanding ability of MLLMs in interleaved image-text contexts. To fill this gap, we propose COHERENCE, a benchmark designed to evaluate the ability of MLLMs to recover fine-grained image-text correspondences in interleaved multimodal contexts. COHERENCE covers interleaved image-text content from four representative domains and contains 6,161 high-quality questions. Moreover, we perform a six-type error analysis, enabling fine-grained attribution of failures in interleaved image-text understanding to the specific capabilities missing in current MLLMs.

Vision-language models have shown strong performance, but they often generalize poorly to specialized domains. While semi-supervised vision-language learning mitigates this limitation by leveraging a small set of labeled image-text pairs together with abundant unlabeled images, existing methods remain fundamentally pairwise and fail to model the global structure of multimodal representation manifolds. Most existing topology-based alignment methods rely on persistence diagram matching, which neither guarantees geometric alignment nor uses the image-text pairing information central to vision-language learning. We propose Topology-Aware Representation Alignment for Semi-Supervised Vision-Language Learning (ToMA), a framework that uses persistent homology to decompose pairwise relations into $H_0$-death and $H_1$-birth relations and aligns them across modalities through available cross-modal correspondences. This lightweight decomposition captures both connectivity and cycle-closing structure without constructing 2-simplices. Experiments show clear improvements on remote sensing, while fashion results reveal strong gains in classification and particularly strong retrieval improvements for difficult queries. Additional analysis shows that ToMA provides more consistent gains than alternative topology-based objectives and that lightweight $H_1$-birth edges provide useful higher-order structural signals.

Vision-language models (VLMs) are increasingly deployed as evaluators in tasks requiring nuanced image understanding, yet their reliability in scoring alignment between images and text descriptions remains underexplored. We investigate whether small, open-weight VLMs exhibit sycophantic behavior when evaluating image-text alignment: assigning high scores without grounding their judgments in visual evidence. To quantify this phenomenon, we introduce the Bluffing Coefficient (\bc), a metric that measures the mismatch between a model's score and its evidence recall. We evaluate six open-weight VLMs ranging from 450M to 8B parameters on a benchmark of 173,810 AI-generated character portraits paired with detailed textual descriptions. Our analysis reveals a significant inverse correlation between model size and sycophancy rate ($r = -0.96$, $p = 0.002$), with smaller models exhibiting substantially higher rates of unjustified high scores. The smallest model tested (LFM2-VL, 450M) produced sycophantic evaluations in 22.3% of cases, compared to 6.0% for the largest (LLaVA-1.6, 7B). These findings have direct implications for the deployment of small, open-weight VLMs as automated evaluators within attribute-rich, synthetic image evaluation tasks, where the gap between assigned scores and cited visual evidence is both measurable and consequential.

Sign language translation (SLT) converts continuous sign videos into spoken-language text, yet it remains challenging due to the intrinsic modality mismatch between visual signs and written text, particularly in gloss-free settings. Recent SLT systems increasingly adopt CLIP-like Vision-Language pretraining (VLP) for cross-modal alignment, but the random in-batch contrast provides few, batch-dependent negatives and may mislabel semantically similar (or even identical) pairs as negatives, introducing noisy and potentially inconsistent alignment supervision. In this work, we first conduct a preliminary trajectory-based analysis that tracks negative video-text similarity over training. The results show that only a small subset of negatives exhibits the desired behavior of being consistently pushed away, while the remaining negatives display heterogeneous and often non-decreasing similarity dynamics, suggesting that random in-batch negatives are frequently uninformative for effective alignment. Inspired by this, we propose Selective Contrastive Learning for SLT (SCL-SLT) with a Pair Selection (PS) strategy. PS scores candidate negatives using similarity dynamics from reference checkpoints and constructs mini-batches via a curriculum that progressively emphasizes more challenging negatives, thereby strengthening contrastive supervision while reducing the influence of noisy or semantically invalid negatives.

Neural networks exhibit a remarkable degree of representational convergence across diverse architectures, training objectives, and even data modalities. This convergence is predictive of alignment with brain representation. A recent hypothesis suggests this arises from learning the underlying structure in the environment in similar ways. However, it is unclear how individual stimuli elicit convergent representations across networks. An image can be perceived in multiple ways and expressed differently using words. Here, we introduce a methodology based on the Generalized Procrustes Algorithm to measure intra-modal representational convergence at the single-stimulus level. We applied this to vision models with distinct training objectives, selecting stimuli based on their degree of alignment (intra-modal dispersion). Crucially, we found that this intra-modal dispersion strongly modulates alignment between vision and language models (cross-modal convergence). Specifically, stimuli with low intra-modal dispersion (high agreement among vision models) elicited significantly higher cross-modal alignment than those with high dispersion, by up to a factor of two (e.g., in pairings of DINOv2 with language models). This effect was robust to stimulus selection criteria and generalized across different pairings of vision and language models. Measuring convergence at the single-stimulus level provides a path toward understanding the sources of convergence and divergence across modalities, and between neural networks and human neural representations.

Large Multimodal Models (LMMs) such as LLaVA are typically trained with an autoregressive language modeling objective, providing only indirect supervision to visual tokens. This often yields weak internal visual representations and brittle behavior under distribution shift. Inspired by recent progress on latent denoising for learning high-quality visual tokenizers, we show that the same principle provides an effective form of visual supervision for improving internal visual feature alignment and multimodal understanding in LMMs. We propose a latent denoising framework that corrupts projected visual tokens using a saliency-aware mixture of masking and Gaussian noising. The LMM is trained to denoise these corrupted tokens by recovering clean teacher patch features from hidden states at a selected intermediate LLM layer using a decoder. To prevent representation collapse, our framework also preserves the teacher's intra-image similarity structure and applies intra-image contrastive patch distillation. During inference, corruption and auxiliary heads are disabled, introducing no additional inference-time overhead. Across a broad suite of standard multimodal benchmarks, our method consistently improves visual understanding and reasoning over strong baselines, and yields clear gains on compositional robustness benchmarks (e.g., NaturalBench). Moreover, under ImageNet-C-style non-adversarial common corruptions applied to benchmark images, our method maintains higher accuracy and exhibits reduced degradation at both moderate and severe corruption levels. Our code is available at https://github.com/dhruvashp/latent-denoising-for-lmms.

Vision-Language-Action systems follow instructions to execute multi-step tasks in multimodal environments. Recent VLA approaches typically rely on post-hoc correction mechanisms or operate under fixed task decompositions and alignment schemes. However, once an intermediate step is mis-specified, local errors propagate through subsequent steps and eventually accumulate into cascading failures. To mitigate this compounding effect, we propose Predictive Alignment and Planning Architecture, a framework that uses prediction and contrast to adjust deviations across three levels: actions, subgoals, and trajectories. Semantic alignment is enforced at all levels using a Sinkhorn-based module and a Score-field module. The predictive correction and alignment jointly update the action generator during training, enabling it to adjust fine-grained steps to remain aligned with the overall intent. We further introduce two new metrics to quantify error propagation and recovery processes in tasks, capturing how mistakes spread and fade over long-horizon execution. Experiments show that ReCAPA achieves competitive results on embodied agent benchmarks such as VisualAgentBench, MineDojo, and AI2-THOR, outperforming strong proprietary and open-source Large Language Model baselines.

Vision-Language-Action Models (VLAs) inherit their visual and linguistic capabilities from Vision-Language Models (VLMs), yet most VLAs are built from off-the-shelf VLMs that are not adapted to the embodied domain, limiting their downstream performance. In this work, we propose EmbodiedMidtrain to bridge the gap between VLMs and VLAs. We first characterize the data distribution gap between them, showing that VLA data occupy compact regions that are largely separated from the broader VLM distribution, while the degree of alignment varies substantially both across and within VLM data sources. Then, we build a mid-training data engine that leverages a lightweight learnable proximity estimator to select the most VLA-aligned candidates from a large VLM pool, and produce EmbodiedMidtrain-1M, a VLM dataset that we further use to mid-train the VLM before downstream VLA fine-tuning. Experiments on both simulated robot manipulation benchmarks and real-world tasks show consistent VLA performance gains over off-the-shelf checkpoints across VLM backbones and VLA action heads. These gains extend to downstream tasks not used to guide data selection. Further analysis reveals that mid-training provides a stronger initialization for VLA fine-tuning, with gains emerging from the earliest steps and widening throughout training. Moreover, the data engine captures both dataset-level and sample-level alignment signals, favoring spatial reasoning over text-centric tasks while preserving the diversity of the VLM data. We will release all code, data and models for future research.