Maksim Evdokimov, Matvey Ivanov, Dmitrii Tsiupin +3cs.CL
Extracting structured fields from hundreds of millions of documents annually remains costly in regulated industries: bespoke OCR cascades cover only a fraction of workflows, privacy rules preclude external models, and existing open-source VLMs that clear quality thresholds cost more to serve than human annotation. We present a deployed document-understanding system built on a Mixture-of-Experts VLM (35B total, 3B active), fine-tuned on in-house production data mixed with open-domain documents curated by a Difficulty-Aware pipeline for layout diversity, fact-extractability, and cross-model consistency. Fitting on a single H100 and serving heterogeneous workflows via prompting, the model leads all deployable (non-reasoning) baselines up to an order of magnitude larger. A quality-adjusted cost analysis, with confirmation and correction costs calibrated from production telemetry, shows it reduces expected costs by over 80% against the human baseline and by more than 50% against the best competing open-source model, while larger baselines remain economically unviable.
Fine-tuning a pretrained LLM into a vision-language model (VLM) can erode the backbone's text capability, with the damage concentrated on tasks that require following exact output rules, such as instruction following, chain-of-thought reasoning graded on a strictly parsed final answer, and similar evaluations with strict graders. We trace this gap to attention-sink corruption: VL fine-tuning perturbs the early sink position that anchors a large fraction of attention probability, and how well the base LLM preserves its sink tracks how much of the affected capability survives adaptation. Building on this view, we introduce Sink Strength, a single scalar computed on the base LLM in a few seconds on a single GPU that predicts post-VL degradation without any VL training. It consistently tracks relative degradation across the six VLM-LLM pairs and multiple format-sensitive tasks. Complementing this diagnostic, we find that post-pretraining QK-RMSNorm injection fails to reproduce the protection of native QK-RMSNorm, while several off-the-shelf weight-merging settings fail to recover the lost capability after VL training. These negative results underscore the value of screening backbones with Sink Strength before VL training and narrow the intervention space toward head-selective training-time protection.
Communicating across cultures is inherently challenging, especially through culturally dense and ambiguous formats like memes. While people expect large language models (LLMs) to hold promise for bridging such gaps, existing benchmark datasets often fail to capture the cultural context necessary for accurate interpretation. To address this, we introduce MemeBridge, a curated dataset centered on U.S.-originated memes, designed to capture two complementary perspectives: (1) how Chinese participants interpret these memes, and (2) how U.S. participants anticipate how people from other cultures might misunderstand them. Here, context refers to implicit cultural knowledge, including background beliefs, norms, and shared assumptions that shape meme comprehension. The dataset was constructed via a multi-stage crowdsourcing pipeline with rigorous validation, including human agreement checks and GPT-based classification verification. Each meme is annotated with sentiment, emotion, cultural significance, and knowledge type, providing rich supervision for downstream tasks. Notably, we observe that the anticipated misunderstandings from U.S. participants are often inaccurate, highlighting the asymmetries in cultural understanding and the challenges of adopting perspectives beyond one's own. This bidirectional framing, which focuses on both expression and perception, enables more nuanced benchmarking of cross-cultural comprehension. Our probing of multiple LLMs reveals that while models developed in different cultural contexts exhibit partial cross-cultural understanding, they often struggle with sophisticated interpretations. By contrast, fine-tuning with MemeBridge improves model performance, underscoring the value of culturally grounded resources for training and evaluating LLMs in globally diverse settings.
Aditi Sarker, Rafi Ibn Sultan, Hui Zhu +2cs.CV cs.AI cs.LG
Large Vision-Language Models (LVLMs) are prone to hallucinations: they fluently describe objects, attributes, and scenes that are not in the image. We connect part of this failure to a measurable property of their representations, feature instability, where mild semantics-preserving perturbations of the input cause large changes in the learned embeddings; hallucination rates rise together with this variability. Existing stability-motivated remedies are explicit, in the sense that they intervene at inference time through latent steering or constrained decoding, and pay for it on every query. We propose implicit stabilization instead: perturbation-invariance is built into the model weights during fine-tuning, and nothing extra runs at deployment. Our framework, INFUSE, first stabilizes visual and textual representations around perturbation-averaged and ground-truth anchors, then aligns the stabilized representations across modalities with bidirectional contrastive objectives. We prove that the anchor's root-mean-square deviation from the perturbation-mean representation shrinks at rate $1/\sqrt{K}$ in the number of views, and that under a Lipschitz decoder, this bounds how much any perturbation can change the model's hallucination behavior. On LLaVA-1.5, LLaVA-1.6, and Qwen3-VL-8B-Instruct, INFUSE reduces AMBER CHAIR by 46-63% relative to each base model, improves ObjHal, MMHal, HallusionBench, and POPE, and preserves VQA-v2 and TextVQA, all with no inference-time overhead.
Chain-of-thought (CoT) reasoning has dramatically improved large language models (LLMs) by allowing them to decompose problems into intermediate steps. While CoT is widely effective for linguistic tasks, text-only CoT forces models to serialize visual problems into awkward prose. Although architectural solutions exist to process visual inputs, the community lacks a massive, multi-step, self-corrected dataset to teach models how to build and maintain internal visual workspaces when solving purely textual reasoning problems. To address this limitation, we introduce CoVA-SFT, a highly structured corpus of 51.9K samples containing over 222K multimodal reasoning steps across 5 distinct layout families and 17 complex tasks, and CoVA-Bench, a companion benchmark of 1,700 held-out test samples spanning the same tasks for reproducible evaluation. By providing explicit rationale formulations, agentic renderings, and verification loops, CoVA-SFT teaches multimodal language models to interleave text and visual abstractions. We validate the dataset by demonstrating that models fine-tuned on CoVA-SFT outperform all interleaved CoT baselines by more than 2x on average on CoVA-Bench, though they still fall short of strong text-only CoT baselines, highlighting open challenges for future work.
Naiming Liu, Zhiheng Wu, Shuning Wang +3cs.CV cs.AI
Recent vision language models (VLMs) have achieved strong progress in video understanding. However, most existing video QA research and benchmarks still follow an offline, single-round paradigm, overlooking realistic interactions where users may interrupt the model during answer generation. To address this gap, we formulate the task of Online Video Question Answering under Interruption and introduce OVIBench, the first standardized benchmark for evaluating VLMs in this setting. OVIBench categorizes interruptions into three types: Cancellation, False Trigger, Correction and supports both open-ended and multiple-choice evaluations. To enable large-scale and reproducible testing, we develop an offline simulation protocol that reproduces interruption during generation under a unified temporal setup, together with a multi-dimensional metric suite for assessing interruption understanding and response generation. Experiments demonstrate that OVIBench effectively distinguishes models' interruption-handling abilities, especially in following correction requests. Finally, we construct a train set OVI-Train for interruption-aware fine-tuning. Models fine-tuned on this dataset achieve significant gains on OVIBench, validating the effectiveness of our benchmark and data design. OVIBench, OVI-Train, and the evaluation code will be released.
Dense-caption retrieval has recently been improved by introducing segmentation, edge maps, LLM-filtered captions, and cross-modal modules into contrastive fine-tuning. However, these methods largely inherit the same InfoNCE objective, whose optimization can prematurely saturate under a strong pre-trained initialization: on dense captions, the loss falls below 10^{-3} on 80% of batches within the first epoch, while its gradient reaches exact zero in fp32 in 47% of measurements. We find that this behavior is closely related to the large number of near-duplicate captions in dense-caption benchmarks, where a few highly similar negatives remain unresolved after the easy majority has already been separated. As a remedy, we introduce HN-CLIP, which uses the text encoder's own text-text geometry to construct per-negative adaptive similarity margins. Specifically, a detached caption-similarity matrix is added to the negative logits, assigning larger margins to more similar captions without mining, synthesizing, or resampling negatives. The resulting objective requires only one caption-similarity matrix and a masked logit addition during training, with no auxiliary data, additional parameters, offline preprocessing, or inference-time overhead. Extensive experiments on four dense-caption retrieval benchmarks show that HN-CLIP improves over the strongest competitors by +2.4--+4.3 R@1 while training 2.4x faster than GOAL and 5.4x faster than StructXLIP. Moreover, the proposed objective improves all six tested fine-tuning frameworks on the in-domain benchmarks and reaches the strongest full-data baseline with only 20% of the training data.
This study investigates the challenge of ambiguity faced by Vision-Language Models (VLMs) in understanding spatial semantics. Spatial cognition, shaped by cognitive psychology, spatial science, and cultural context, often assigns directionality to objects. However, natural language descriptions of spatial relations frequently omit explicit reference frames, leading to semantic ambiguity and potentially serious errors for embodied AI robots. Existing VLMs, due to insufficient training on reference frames and object orientations, often produce inconsistent responses. To address this issue, we construct a new dataset, AlloEgo-View, comprising (image, query, view-specific answer) triplets that capture key object relations from both allocentric and egocentric perspectives. The view-specific descriptions follow a structured spatial representation that annotate detailed scene descriptions, reference and target objects, their orientations, reference frames, and view types. Building on AlloEgo-View, we develop AlloEgo-VLM, a framework to disambiguate allocentric and egocentric reference frames, even under ambiguous queries, and to be easily integrated into existing VLMs via supervised fine-tuning. Furthermore, we deploy our framework onto an embodied robotic platform within NVIDIA Isaac Sim to validate its real-world feasibility in open-ended object searching tasks. Experiments highlight the limitations of current VLMs in handling view-specific queries and demonstrate the strong disambiguation ability of AlloEgo-VLM.
Vision-language models (VLMs) remain unreliable on chart questions that require reasoning over visual quantities, and this weakness is usually attributed to a reasoning deficit and addressed with more reasoning supervision. We ask whether the difficulty lies in reasoning itself, or in the simpler skills that reasoning operates on: reading the plotted elements (\emph{perception}), locating them and binding them to their labels (\emph{grounding}), and performing single-step computations such as ranking, totals, and differences (\emph{simple reasoning}). We introduce \textbf{ChartProbe}, a diagnostic framework whose probes are generated directly from the code that renders each chart, so every gold answer is exact by construction, needs no human annotation, and attributes each failure to a single skill. ChartProbe enables an intervention prior work does not attempt: instead of synthesizing complex-reasoning data, we withhold complex questions and reasoning traces entirely, fine-tune on one simple skill at a time, and measure transfer to held-out complex-reasoning questions. Across three open-weight VLMs, supervising the simpler skills alone produces large gains on complex-reasoning questions the model never trained on: where these skills are weak and the model can be taught to read the image, training them recovers much of complex reasoning at no reasoning-data cost. The gains hold across three out-of-distribution settings: an unseen chart type (pie charts), a human-written benchmark disjoint from our images and templates (ChartQA), and a non-chart visual domain (CLEVR). Complex visual reasoning can therefore improve without complex-reasoning supervision.
Object hallucination in multimodal large language models arises when language priors and corpus co-occurrence bias outweigh the visual evidence, with nothing tying an individual object mention to what the image shows. Most remedies intervene at decoding time without training, yet under a unified protocol their benefit is confined to short captions;supervised fine-tuning (SFT) on a detail- rich corpus lengthens captions, but over forty percent still name absent objects. This paper proposes Dual-Stream Cross-Anchor Correction (DSCC). Unlike work that post-processes decoding, DSCC is the first to inject object-level visual anchors into the language model itself during fine- tuning: a perception stream aligns object-level hidden states at an intermediate layer to frozen text anchors by a bidirectional contrastive objective; a cognition stream lets deeper layers query those anchors by cross-attention at every generation step; and a two-stage curriculum gate couplesthem, making evidence retrieval a structural constraint at each autoregressive step. Under one backbone and one scoring protocol, experiments span long-caption hallucination, object-existence discrimination and cross-domain generalisation, with vanilla SFT on the same corpus and schedule as a length- and density-matched control, so gains are attributed layer by layer. DSCC is the only method reaching the long-caption, low-hallucination region: captions roughly 1.9 times the baseline length at 88.19% precision per object mention, the highest under a density-independent criterion. Ablations expose a synergy: the perception stream alone degrades precision yet reverses sign when stacked on the cognition stream. No universal superiority is claimed: three out-of- domain benchmarks yield a predictable, falsifiable domain-conditionality, the synergy being bound to the anchors' semantic domain and breaking on charts and optical illusions.
Multimodal Large Language Models (MLLMs) exhibit strong generalization and reasoning abilities due to large-scale multimodal pre-training. However, fine-tuning these models on downstream tasks often leads to catastrophic forgetting, where newly learned task-specific knowledge degrades previously acquired capabilities. This issue arises because gradient updates for new tasks overwrite parameters critical to prior knowledge, limiting the practical deployment of MLLMs. To address this challenge, we propose Activation-Weighted Adaptive REtention (AWARe), a fine-tuning method that mitigates catastrophic forgetting by dynamically controlling parameter updates based on activation patterns. AWARe assigns activation-based importance scores to parameters, selectively freezing those essential for preserving prior capabilities while allowing less important parameters to adapt to new tasks. Importantly, AWARe operates without modifying model architectures, ensuring compatibility with existing inference engines. Extensive experiments demonstrate that AWARe effectively preserves upstream capabilities while achieving superior downstream performance compared to existing methods. Code is available at https://github.com/kaln27/AWARe.
Driven by the rapid advancement of vision-language representation learning, Text-based Image Retrieval (TBIR) has made notable progress. However, existing benchmarks are predominantly constructed on an exclusive single-match assumption between query and images. While effective in general scenarios, this assumption fails to reflect practical system performance in specific domains (e.g., surveillance), where a single query often corresponds to multiple relevant candidate images. To address this limitation, we design a Domain-Specific Multi-Match Text-based Image Retrieval (DSMM-TBIR) data engine. Leveraging this engine, we construct Security Multi-Match TBIR (SecMM-TBIR), a benchmark comprising 50k surveillance images with 200 comprehensive queries. Furthermore, we observe that vanilla contrastive learning in specific domains suffers from severe false negatives, forcing the model to push apart semantically similar pairs and thus degrading retrieval performance. We propose the Semantic-Aware Fine-Tuning (SAFT) framework to address semantic compression in specific domains, which incorporates Semantic-Aware Soft-Label Supervision (SASS) and Intra-modal Structural Distillation (ISD) to establish a promising paradigm for domain-specific TBIR tasks. Experiments across diverse CLIP-like models demonstrate that SAFT yields an average mAP@20 gain of 7.8 points on SecMM-TBIR over standard image-text contrastive (ITC) fine-tuning, while also improving general-domain performance. The entire benchmark will be released to facilitate further research.
We present TAR (Traffic Anomaly Reasoning) and TAR-Bench datasets, resources for training and evaluating video-language models beyond anomaly detection. TAR contains 44,040 chain-of-thought training annotations across 10 tasks for 3,670 CCTV videos ($\sim$26 hours) from eight public datasets. Its evaluation component, TAR-Bench, contains 960 human-curated test annotations for 80 held-out clips trimmed from 17 public YouTube videos. TAR's training annotations are produced with MAVEN, which consolidates multi-scale video evidence into structured event descriptions before generating question-answer pairs and reasoning traces. On TAR-Bench, eleven vision-language models reveal that strong question-answering accuracy does not reliably predict temporal or scene reasoning ability. Multi-task fine-tuning on TAR yields consistent gains, with the full 10-task model improving aggregate score by 21.4 points over its zero-shot baseline. TAR and TAR-Bench provide the official training and in-domain evaluation data for AI City Challenge 2026 Track 3. The dataset is available at https://huggingface.co/datasets/nvidia/PhysicalAI-Traffic-Anomaly-Reasoning
We present results from reconstructing multiple-choice model (MCM) and three-parameter logistic (3PL) model curves using a fine-tuned multimodal large language model (LLM) based on Qwen3.5. The model is prompted and fine-tuned to replicate choice probabilities across a large training corpus of multiple-choice items containing both image and text stimuli, conditioned on a labeled set of student ability levels. By learning to reproduce the systematic error patterns of students across a discrete range of abilities, the LLM implicitly captures the underlying response probabilities encoded in the 3PL and MCM curves. This allows us to accurately approximate item difficulty on a held-out test set directly from the model's predicted option probabilities.
Changhao Xiang, Shilin Zhang, Zheng Ma +8cs.CL cs.CV cs.LG
Visual tool use has emerged as a fundamental capability for multimodal agents to actively acquire evidence beyond a fixed image encoding. The prevailing recipe learns this capability from teacher-generated trajectories filtered for answer correctness, implicitly assuming that every successful demonstration provides effective supervision. We argue this assumption is flawed: a strong teacher often reaches the correct answer without needing its tool calls, and imitating such trajectories teaches a student that tool calls accompany correct answers, not that tool observations ground them. We present OpenVisTool, an open framework for constructing instructive visual tool-use trajectories that provide effective supervision for tool learning. The key insight is that a trajectory should be retained only if its answer is correct (outcome validity) and its tool observations causally contribute to that answer (causal utility). The framework operates in three stages: difficulty screening to select queries that are not reliably answerable without tools, domain-specific trajectory synthesis to elicit coherent tool-use trajectories, and supervision verification to jointly test both conditions. Rather than encouraging models to imitate tool calls, the resulting supervision teaches when and how visual evidence should be acquired. Using this framework, we construct OpenVisTool-42K, a dataset spanning five visual reasoning domains, together with OpenVisTool-Bench, a benchmark covering the same domains. Across four backbones (4B-27B), fine-tuning on OpenVisTool-42K consistently improves visual tool-use performance and yields gains on two out-of-distribution benchmarks; the larger models approach leading closed-source systems. The evidence suggests that effective visual tool use is learned from causally grounded supervision rather than tool-calling patterns.
Despite the remarkable progress of large vision language models (LVLMs), object hallucination remains a fundamental challenge that hinders their trustworthy deployment. A key finding motivates our work: real and hallucinated object tokens are clearly separable in hidden representations, yet this separability is largely lost at the language-modeling (LM) head. We propose TruthLens, a self-evaluation framework that teaches the LM head to expose a per-object truthfulness signal without any auxiliary model or additional inference cost. Concretely, a rarely-used special token is repurposed as a reference token. For each object-token position, we extract the log-probability assigned to this special token by the LM head, and define its difference from a predefined constant as the truthfulness score. The model is then fine-tuned with an MSE objective that drives scores toward 1 for real objects and 0 for hallucinated ones, while a divergence constraint preserves the original generation capability. Despite being trained on only a limited set of object categories, TruthLens generalizes effectively to benchmarks with substantially larger label spaces. Extensive experiments across multiple LVLMs demonstrate state-of-the-art performance; notably, on Qwen2.5-VL-7B, TruthLens outperforms the previous best method on MS-COCO by over 17\% in AUROC. Our code is available at https://github.com/wyqstan/TruthLens.
Deploying vision-language models (VLMs) for safety-critical spatial reasoning on resource-constrained autonomous driving platforms requires both compact model size and reliable metric grounding. We present MoRAL (Multimodal Reasoning for Autonomous Language Models), a two-stage fine-tuning pipeline that teaches Cosmos-Reason2-2B to first read a physics-encoded Bird's Eye View (BEV) representation and then reason over it for driving decisions. The BEV image encodes LiDAR metric distance as color bands, object class as cluster morphology, and radar Doppler velocity as directional wedge overlays, externalizing spatial perception into the input image so that no learned 3D backbone is required at inference. Stage 1 fine-tunes the vision encoder on 60,000 grounding records; zero-shot baselines produce no parseable BEV outputs, confirming the vocabulary requires explicit training. Stage 2 fine-tunes the full model (52M parameters, 2.4% of total) on 57,696 chain-of-thought records generated by Cosmos-Reason2-8B as teacher, spanning eight driving question types. On 2,304 held-out nuScenes frames evaluated by Gemma 4 (31B) calibrated against human review, MoRAL wins seven of eight question types over a zero-shot 8B baseline despite using four times fewer parameters, with the largest margins on question types requiring structured multi-step physics reasoning. Emergency braking recall improves from 10.8% to 47.8%, output degeneration falls from 94.1% to 20.8%, and the full pipeline fits a consumer 8 GB GPU at 42 tok/s without quantization. These results establish a reproducible foundation for compact, physics-grounded VLM reasoning on mobile edge platforms.
Bohan Hou, Haoqiang Lin, Xuemeng Song +4cs.CV cs.IR
Due to their strong generalizable multimodal processing and reasoning capabilities, Multimodal Large Language Models (MLLMs) have demonstrated significant potential as universal image retrievers, effectively addressing diverse real-world image retrieval tasks. Nevertheless, pioneering studies, while promising, overlook the potential of fine-grained context modeling and disentangled fine-tuning objectives in enhancing MLLMs' retrieval performance, particularly for complex tasks such as long-text-to-image retrieval, visual dialog retrieval, and composed image retrieval (CIR). Therefore, in this work, we propose an automated fine-grained multimodal quintuple dataset construction pipeline and a novel two-stage fine-grained multimodal fine-tuning strategy. The dataset generation pipeline produces a comprehensive CIR dataset with fine-grained image captions and modification text, facilitating fine-grained context modeling. Beyond the previously entangled fine-tuning paradigm, our approach separates the fine-tuning process into two distinct stages: (1) fine-grained context reasoning-oriented fine-tuning and (2) fine-grained retrieval-oriented fine-tuning. These stages aim to sequentially enhance the model's context understanding and query-target alignment capabilities, thereby improving retrieval performance. Extensive experiments across five datasets encompassing diverse and complex image retrieval tasks demonstrate the remarkable superiority of our method over existing approaches in zero-shot retrieval settings, even with a more lightweight MLLM backbone compared to those methods.
Multi-frame medical VQA appears to reward increasingly complex adaptation: controller-style inference, localization-aware reranking, static hard-negative mixing, and staged continuation all appear plausible from first principles. We test a simpler competing hypothesis on MedFrameQA: methods that remain tightly aligned with the benchmark's final answer objective should be the strongest \emph{robust} adaptation family once evaluation is controlled across fixed splits, matched budgets, repeated seeds, and calibration. We compare controller-based methods, scaffold evolution, static mixed supervision, continuation-heavy variants, and direct answer-only supervised fine-tuning (SFT). The strongest robust family is direct decoder-only answer SFT on MedGemma-1.5-4B. Empirically, this family yields substantial improvements in held-out report accuracy over frozen baselines while remaining remarkably stable across repeated seeds and matched controls, ensuring our claims reflect true family-level robustness rather than an isolated hyperparameter peak. Furthermore, post-hoc calibration effectively repairs confidence estimation without compromising accuracy, and the core approach transfers consistently to secondary backbones like Qwen2.5-VL-3B. The main result is therefore not that a complex auxiliary mechanism wins, but that objective-aligned direct answer SFT is the strongest robust adaptation family we found for MedFrameQA. By establishing this strong, minimalist baseline, we hope to redirect community focus toward fundamentally robust optimization rather than architectural complexity.
Large-scale video platforms process millions of uploads hourly, requiring moderation systems that can localize when and where policy violations occur within each video. Processing every frame is infeasible at scale, so systems are constrained to sparse inputs of 8 to 16 frames per video. Yet state-of-the-art multimodal large language models (MLLMs) are pretrained on dense sequences of hundreds of frames, creating a fundamental mismatch between training and deployment conditions. This mismatch causes severe performance collapse: the Qwen3-VL 8B model drops from 56.0% to 22.3% temporal mIoU when frames are reduced to 16, a 60.2% relative degradation. We present a systematic empirical study of training strategies to close this gap for spatial-temporal video grounding. Our results suggest that visual feature extraction is the dominant bottleneck under sparse-frame inputs. Adapting only the final three ViT layers, 4% of total parameters, achieves 68.8% temporal mIoU and surpasses a zero-shot 8B model using dense inputs by 12.8 points. Language model fine-tuning, by contrast, offers negligible or negative returns. A boundary-aware sampling strategy, Hybrid16, further improves temporal mIoU by 26 points over uniform sampling when temporal boundaries are available. We conclude that for sparse-frame video grounding, training strategy dominates model scale: a fine-tuned 2B model consistently outperforms a zero-shot 8B model, with or without dense frame access.
Sana Tonekaboni, Viktoria Schuster, Caroline Uhlercs.LG
Real-world perception and decision making are inherently multimodal, integrating complementary signals across modalities. However, training multimodal models faces two main obstacles. First, collecting large-scale, well-aligned paired multimodal datasets is often impractical, making end-to-end multimodal training difficult. Second, existing multimodal representations frequently entangle information shared across modalities with modality-specific information, hindering interpretability and control. We introduce MultiLoReFT, an efficient and scalable low-rank representation fine-tuning framework for multimodal learning with pretrained unimodal models. MultiLoReFT extends low-rank adaptation to the multimodal setting and learns interpretable projection subspaces that decouple shared and modality-specific information. Across simulated and real-world benchmarks, it produces representations that support multimodal prediction while explicitly revealing how shared and modality-specific information is distributed across modalities.
Contrastive Language-Image Pre-training (CLIP) has been shown to have limitations in its fine-grained dense feature representation, due to its pre-training focusing on matching the whole image to a text description. Considering the large data and computational burden in pre-training a vision-language model from scratch, a series of works aim to enhance the fine-grained ability of CLIP through a fine-tuning scheme. However, existing works suffer from a variety of limitations: additional region annotations are usually required, which limits the semantic diversity due to the predefined categories and leads to a large effort to process the training data; and they usually sacrifice CLIP's original ability for global visual representation. To bypass these limitations, we propose SFF-CLIP (Self-annotated Fine-grained Fine-tuning for CLIP), which only uses image-text pairs as input to boost the fine-grained representation ability in the CLIP fine-tuning, while maintaining the global visual-semantic consistency. Concretely, a run-time region-phrase alignment scheme is designed, which obtains concept phrases from the input sentence, and aligns them with corresponding extracted region-based features using text-specific heat maps. Extensive experiments demonstrate that SFF-CLIP leads to significant performance improvements on fine-grained dense feature representation, as well as maintaining the performance of the original CLIP on image-level tasks. Code will be released later.
Assembly action understanding is a key enabler for effective human-robot collaborative assembly, yet it remains challenging due to subtle motions and fine-grained hand-object interactions. We adapt vision-language models (VLMs) to this challenging domain with Compositional Context Fine-Tuning (CCFT), a method that decomposes assembly actions into semantic elements (Verb, Object, Tool) and fine-tunes VLMs to recognize each action element using templated question-answering pairs. This approach ensures near-deterministic outputs. To enable efficient and effective multi-task learning under limited data, a Layer-Partitioned Alternating Training (LP-AT) method is presented, which assigns distinct model layers to recognize specific action elements through element-specific low-rank adapters. LP-AT alternates weight updates across element-specific adapters, reducing cross-task interference while enabling per-adapter hyperparameter optimization. Furthermore, we create HA-ViD-VQA and IKEA-ASM-VQA datasets from existing assembly video datasets. Extensive experiments on these datasets demonstrate that our method consistently outperforms strong action recognition baselines while providing interpretable element-level predictions that can support diverse downstream applications.
Despite the rapid advance of Multimodal Large Language Models (MLLMs) in high-level video understanding, a fundamental bottleneck remains: these models collapse complex human motion into coarse semantic labels. Existing benchmarks mostly focus on scene-centric events or local joint articulations, failing to probe global human motion in space over time (trajectory and orientation changes). We introduce HumanMoveVQA, the first comprehensive benchmark designed to evaluate global trajectory and orientation reasoning from an exocentric perspective. Our benchmark utilizes a first-frame anchored world coordinate system, preserving translation and rotation relative to a fixed starting point. We propose a scalable, multi-stage pipeline that lifts 2D video observations into world-consistent 3D motion tracks to generate over 10K structured question-answer pairs across seven reasoning categories, including motion aggregation, sequential ordering, and trajectory-level inference. Our extensive evaluation reveals a critical capability gap in state-of-the-art proprietary models on deep human motion understanding. However, we demonstrate that this is a learnable problem; by fine-tuning an open-source baseline with our targeted, world-consistent supervision, we achieve a significant improvement. HumanMoveVQA establishes a rigorous geometric foundation for developing next-generation, movement-aware video understanding models.
Eren Senoglu, Federico Toschi, Nicolo Brunello +2cs.LG cs.CL cs.CV
Multimodal large language models (MLLMs) applied to Medical Visual Question Answering (VQA) tend to produce overconfident outputs regardless of actual correctness, and existing verbalized confidence calibration methods, developed primarily for text only LLMs, do not account for the multimodal nature of medical image understanding. This work proposes a training based framework that finetunes MLLMs to improve their calibration using a composite loss function combining a Brier style calibration term, an anchor regularizer that prevents confidence collapse toward extreme values, a contrastive image text alignment term, and a KL based model stabilization term. The alignment signal is derived from a $2 \times 2$ factorial perturbation design that crosses image presence with text integrity, probing the reliance of the model on visual modality input versus language priors. Finally, a top K KL divergence regularizer is used to protect the answering ability of the model during finetuning. Across three Medical VQA benchmarks and two architectures (MedGemma 4B IT and Qwen2 VL 7B Instruct), our method reduces calibration error by 60% or more, and improves discrimination by 26% or more, while preserving predictive accuracy. On average across benchmarks, the technique outperforms prompting based, sampling based, and training based approaches, and ablation experiments confirm that each component of the loss function is indeed necessary for improving the calibration. All code for the experiments is publicly available.
Fine-tuning Multimodal Large Language Models (MLLMs) on specialized tasks often leads to catastrophic forgetting of their general capabilities. Existing model merging methods to combat this are often heuristic or use sub-optimal objectives. We propose CurvatureGuided Mixing (CGM), a theoretically grounded framework that merges pre-trained and fine-tuned models. CGM formulates a joint optimization objective and uses a second-order (Hessian) approximation of the loss landscapes to analytically derive an optimal, closed-form "soft mixing" ratio. This ratio intelligently blends parameters based on their relative task-specific curvatures. We also introduce CGM$\dagger$, a robust "hard mixing" variant that performs sparse parameter selection guided by a novel, curvature-aware score. Experiments on LLaVA-1.5 and Qwen2.5VL across multiple downstream tasks show that CGM and CGM$\dagger$ consistently improve the trade-off between task specialization and general knowledge retention over existing methods. Code is available at github.com/zzsyjl/CGM-ECCV-2026.
Multimodal Large Language Models (MLLMs) often lose track of the right image regions during fine-grained spatial reasoning, because a textual query rarely carries any explicit geometric anchor into the pixel domain. Prevailing remedies either rewire the model's weights or pad the prompt with verbose instructions, yet neither reliably pins the language to the correct visual coordinates without eroding the backbone's general competence. We introduce Timage, a paradigm that recasts multimodal understanding as an alignment problem solved at the input: the query is drawn, as a typeset overlay, onto the image itself. The placement and appearance of this overlay are produced by a Constrained Schrödinger Bridge (cSB), an entropic optimal-transport sampler that factorizes layout synthesis into two coupled stochastic stages. The first stage, Region Search, transports noise toward query-aligned image zones while obeying a hard occlusion barrier that protects salient foreground content; the second stage, Appearance Shaping, sizes the glyphs through an ``ink-budget'' regularizer so that the rendered text stays legible and visually balanced. The resulting overlay behaves as an explicit attention beacon that channels the model's focus along spatial semantics. On the VMCBench suite, Timage paired with a modest 7B backbone clearly overtakes far larger proprietary systems as well as parameter-tuned baselines. The study positions deliberate input reconstruction as a powerful, architecture-neutral lever for strengthening multimodal reasoning.
Nguyen Cao Hoang, Hoang Bui Le, Nam Vo Hoang +1cs.CV
Composed image retrieval retrieves a target image using a composed query of a reference image and a modified text description. In the fashion domain, this task requires understanding subtle attribute variations such as color, pattern, and texture. However, existing approaches face limitations due to scarce annotated data and simplistic negative sampling. We propose a novel framework that integrates a multi-modal large language model (LLaVA) to generate attribute-aware triplets and introduces a two-stage fine-tuning strategy to enhance contrastive learning. We leverage pretrained vision-language models, such as CLIP-ViT/B32, to generate and concatenate sentence-level prompts with the relative caption and to scale the number of negatives using static representations. Experimental results demonstrate enhanced compositional reasoning and improved fine-grained retrieval behavior, underscoring the feasibility and potential of the proposed framework for fashion retrieval.
Human perception of visual scenes is inherently temporal. We instinctively recognise whether a fruit is ripening or rotting, whether construction is progressing or being demolished, and approximately how much time separates two photographs of the same subject. Whether large vision-language models (VLMs) share this competence remains an open and practically important question. We introduce CHRONOSIGHT, a rigorously controlled benchmark evaluating five dimensions of visual temporal reasoning: CHRONORANK (chronological ordering of image sequences), CHRONOLOCATE (ordinal stage localisation from a single image), CHRONODELTA (estimation of time elapsed between two images on a logarithmic scale), CHRONOREVERSE (detection of temporally reversed sequences), and CHRONOODD (identification of a temporal outlier within a set). The benchmark comprises 1{,}000 items across eight process families (biological growth, food transformation, physical weathering, construction, environmental change, human ageing, astronomical phenomena, and urban dynamics) spanning timescales from minutes to millennia. We evaluate eight open-source VLMs (500 M to 19 B parameters) under two prompting regimes and collect human performance baselines. Human performance averages 0.89 across tasks; the best open model (Qwen2.5-VL-7B) reaches 0.40 under direct prompting, a gap we term chronological blindness. Lightweight LoRA fine-tuning on 151 examples raises CHRONODELTA accuracy from near-zero to 0.43, transferring zero-shot to related tasks (CHRONOODD: 0.37; CHRONOREVERSE: 0.64)suggesting the bottleneck is partly instruction following rather than visual perception. Benchmark, code, and predictions will be released upon acceptance.
Young Rok Jang, Hyesoo Kong, Kyunghwan An +3cs.CV cs.AI
Real-world documents combine text with tables, charts, photographs, and diagrams arranged in diverse layouts, yet existing research on multimodal large language models (MLLMs) for document QA predominantly produces text-only responses, underutilizing these visual elements. We introduce VinQA, a dataset for long-form answer generation where cited visual elements are explicitly interleaved with their supporting text and grounded in relevant document pages. To support this task, we study two encoding methods for feeding raw document page images into an MLLM, along with their visual-element citation mechanisms: (1) Page Encoding, which directly encodes full-page images with bounding boxes of visual elements and treats these boxed regions as citable units; and (2) Modality Encoding, which parses each page to extract text and crop visual elements, encodes them separately, and uses these cropped elements as citable units. In our experiments, we propose M-GroSE, a multimodal evaluation framework extending GroUSE to assess answers along four dimensions: completeness, answer relevancy, faithfulness, and unanswerability. We additionally report Visual Source F1 to directly measure visual citation accuracy. Although proprietary frontier models still achieve the best overall scores on the VinQA test split, fine-tuning open Qwen2.5-VL models on the training split substantially improves their performance and narrows this gap. Modality Encoding is initially more robust for complex documents with long text, many visual elements, and diverse citation requirements. After training on VinQA, however, Page Encoding reaches a comparable level, competing effectively even without the explicit parsing used in Modality Encoding. Finally, Visual G-Eval, an MLLM-based judge, confirms that fine-tuned models insert visual elements at semantically appropriate positions with faithful supporting text.