We propose Vittle, a new visual instruction tuning framework that improves robustnessof MLLMs to data distribution shifts by pursuing the minimal sufficient representation.

The paper proposes AsymVLM - a probabilistic adaptation framework for vision–language models that models text uncertainty on the unit hypersphere, using spherical distributions to improve cross-modal retrieval and robust classification.

Attention heads in text-generative models specialize in semantic and visual concepts. Leveraging this property, we can reliably suppress or enhance specific attributes in both language and vision-language tasks.

We introduce DrVD-Bench, the first benchmark to evaluate vision-language models' clinical reasoning ability across graded medical imaging tasks.

We propose a training-free visual cropping method that leverages MLLM-internal representations for VQA tasks focusing on small details, achieving strong performance with significantly higher efficiency than prior methods.

We introduce SpatialReasoner, a novel large vision-language model that improves 3D spatial reasoning with explicit 3D representations and generalizable 3D thinking.

We propose a simple yet highly effective transfer-based baseline for attacking black-box closed-source LVLMs.

We propose a novel hybrid of position embedding to improve the length generalization ability of Vision-Language Models.

DualGround mitigates EOS token bias by introducing additional phrase-aware path for fine-grained video-language alignment.

We use VLMs to generate high-level hand-object plans for dexterous manipulation.

Training reinforcement learning agents from a single language instruction using vision-language models.

We propose a sparse autoencoder that maps the semantics of vision and language representations into a unified concept set.

We present CrypticBio, the largest publicly available multimodal dataset of visually confusing species, specifically curated to support the development of AI models for biodiversity identification using images, language and spatiotemporal data.

We propose CLIPTTA, a contrastive test-time adaptation method for CLIP that improves both accuracy and OOD detection in closed- and open-set settings.

We propose Uni-MuMER, full fine-tuning a VLM for handwritten math expression recognition via structured spatial reasoning (Tree-CoT), error-driven learning (EDL), and symbol counting (SC), achieving state-of-the-art results.

We propose Test-Time Spectrum-Aware Latent Steering (STS), a lightweight adaptation that learns a few per-sample steering parameters in latent space to align visual and text embeddings and boost zero-shot robustness.

We introduce a novel reasoning paradigm -- pixel-space reasoning. We identified the learning trap when cultivating this ability and proposed a curiosity-driven RL approach to address it.

We propose GenIR, a method using generative visual feedback to tackle the multi-round image retrieval task, and release a dataset.

NoisyRollout boosts VLM reasoning by mixing clean and noisy inputs during RL, improving generalization with no extra cost.

We introduce Roboflow100-VL, a large-scale dataset with diverse concepts not typically found in internet-scale pre-training.

NOVA is an extreme OOD stress-test dataset of ∼900 multi-modal brain MRI scans (with 281 rare pathologies) for benchmarking VLMs on three clinical tasks: anomaly localization, captioning, and diagnostic reasoning.

ViSpec accelerates vision-language model inference by integrating vision-aware speculative decoding with compressed image tokens and global feature injection, achieving up to 3.22× speedup.

We introduce the few-shot universal cross-domain retrieval setup and propose a novel PEFT method, MAIL, tailored for both this task and few-shot classification.

Teach multimodal models to think or not to think- this is a question.

We propose Approximate Domain Unlearning (ADU), a new task for selectively removing domain-specific knowledge from pre-trained vision-language models, and introduce two key techniques, domain disentangling and instance-wise prompting, to solve it.

Perform input dependent steering of MLLMs by learning to predict steering vectors from input context

NormFit is a novel fine-tuning framework designed specifically for few-shot federated learning scenarios characterized by heterogeneous and imbalanced data distributions

We introduce i-CIR—an instance‐level composed image retrieval benchmark with rigorously curated hard negatives—and BASIC, a training‐free VLM‐based method that centers and projects image embeddings.

We propose CoAPT, a collaborative adversarial prompt tuning method that significantly improves the robustness and generalization of vision-language models against adversarial perturbations.

ViGoRL is a vision-language model trained with reinforcement learning to ground each reasoning step in image coordinates, improving performance on spatial and web-based reasoning tasks through better attention and visual verification.

Evaluating the spatial reasoning capabilities of large Vision-Language Models