All papers

The paper proposes a system for physical AI that continuously monitors environments to identify intervention needs and execute timely actions.

The paper presents a benchmark of synthetic paired comparisons to evaluate video reward models across specific visual and motion dimensions.

Introduces a visual document retrieval benchmark focused on multi-page and cross-document evidence configurations in long documents.

Generates prompt-adapted adversarial images that block unauthorized visual questions on vision-language models while still answering legitimate prompts correctly.

Presents a reference-free evaluation metric that decomposes captions into atomic claims and verifies them against visual evidence.

Develops an agentic text-to-image framework allowing agents to revisit and select earlier image states for iterative compositional generation.

The paper proposes a training-free method establishing entity identity before final rendering to maintain visual consistency across long storyboards.

The paper introduces a benchmark to measure how often vision-language models accept incorrect Qur'anic citations across various visual presentations.

The paper studies how visual input signals affect scaling behavior and non-visual reasoning during native multimodal pretraining.

This paper introduces STP-GUI, a spatially aware token-pruning framework that preserves local interface structures for efficient web-GUI vision-language agents.

Uses factorial interventions to locate failure points in multi-image vision-language reasoning pipelines.

Optimizes multimodal prompts self-supervisedly to synthesize realistic industrial defect images for anomaly detection.

A meta-analysis shows multimodal language models struggle with geometric acquisition variations in aerial video, exposing limitations in standard QA-based evaluations.

The paper detects backdoor triggers in multimodal language models by tracking output evolution trajectories across internal hidden layers.

Introduces a decomposed rubric synthesis method to create training data for open multimodal models that generate visual code artifacts.

This paper presents a training-free contrastive decoding framework using attention-guided perturbations to reduce hallucinations in video language models.

This paper proposes SCOUT, a cascaded two-track framework combining visual candidate filtering, offline semantic summaries, and query refinement for document retrieval.

This paper introduces Anuvani LID, a dedicated 115M parameter model for spoken language identification across 22 Indian languages.

This paper introduces DarKnight, a framework that uses locally hosted vision-language models to solve dark web CAPTCHAs during threat intelligence gathering.

Analyzes why feature-based test-time adaptation fails to recover accuracy on corrupted inputs in generative multimodal language models.

This paper develops a self-supervised representation learning approach based on pointwise mutual information factorisation for scalable generative latent tokenization.

The paper proposes a reinforcement learning framework using quality-aware rewards to guide multimodal knowledge graph construction for retrieval-augmented generation.

The paper presents an agent orchestration framework that trains smaller language models to navigate, inspect, and reason over long multimodal documents.

Presents a framework that decouples speech generation and synthesis token spaces to speed up spoken dialogue model inference.

The paper proposes a self-play framework to train multimodal foundation models using text proxies instead of paired cross-modal datasets.

The paper identifies an internal activation direction in spoken language models to control and transfer conversational turn-taking styles.

The paper presents PocketTSLM, deploying an on-device soft-prompt time-series language model on smartphones for personal health data.

Introduces a benchmark spanning ten tasks to evaluate multimodal model verification and rejection explanations on generated 3D content.

The paper investigates learning physical damping properties from impact audio and distills this knowledge into vision models.

The paper reduces vision-language model parameters by pruning layers and applying depth recurrence over remaining language backbone blocks.

Introduces a framework that enables vision-language models to iteratively adapt frozen visual encoder computation based on the input question.

Trains vision-language models via reinforcement learning to actively navigate and inspect 3D environments using single views before answering questions.

Proposes a decoupled inference-time guidance algorithm to filter unsafe outputs in text-to-image diffusion models without altering weights.

Identifies early-stage issues with classifier-free guidance in flow matching models and proposes a correction for controllable image editing.

Decomposes agentic spatial reasoning by using a metric 3D perception tool to resolve visual inconsistencies before spatial reasoning.

The authors present an active spatial search framework enabling vision-language agents to ground 3D objects in city-scale environments via interactive tool use.

The paper presents a benchmark to test how well AI agents can reconstruct dynamic, editable 3D scenes from dynamic reference videos.

The paper introduces a hierarchical benchmark for evaluating visual generative models on multi-step world state prediction tasks.

Proposes a visual verifier that decomposes complex prompts into sub-questions and assigns continuous scores to improve multimodal evaluation.

This paper introduces an agentic post-training framework that coordinates multiple agents to evaluate and adaptively train text-to-image models.

Proposes a hybrid image tokenizer combining compact 1D global tokens with adaptive sparse 2D tokens for local detail refinement.

The paper presents a four-gate verification protocol to evaluate whether audio-conditioned visual prediction models genuinely rely on input audio signals.

Proposes a multimodal drag-based image editing framework that generates visual reasoning traces to guide point-based transformations.

CameraVLM couples motion states and on-screen flow dynamics to improve camera-motion understanding in vision-language models.

The paper introduces an architecture combining private prediction paths and graph collaboration to maintain multimodal inference accuracy when modalities are missing.

Presents a teacher-student framework to evaluate whether remote sensing visual queries are genuinely supported before predicting bounding boxes.

Benchmarks multimodal decision models across multiple prompts for classifying breast ultrasound imaging datasets.

The paper introduces a generative pipeline to reconstruct multi-room 3D environments from sparse panorama images using geometric optimization.

Presents a dual-stream memory architecture combining recurrent compression and frame retrieval for vision-language-action robotic policies.

Introduces a preview-and-edit framework that diagnoses generation failures from low-cost previews and adjusts cached trajectories during video generation.

Introduces a 3D vision-language foundation model trained on MRI data with anatomically aligned text and pathology-aware representations.

Introduces a lightweight velocity corrector to improve prompt following in few-step text-to-image generation models without full fine-tuning.

Synthesizes stereo manipulation data from monocular robot demonstrations using estimated depth and video diffusion models.

Proposes a camera-configurable framework for vision-language-action policies that handles variable camera counts and poses without fixed setups.

Investigates extracting spatial, attribute, and relational modifier grounding directly from frozen multimodal large language model representations using optimal transport.

Presents an agent harness that pairs frozen multimodal language models with modular navigation skills for general robot navigation tasks.

The paper presents LIVIN, a spatial and embodied intelligence benchmark featuring digital twins of 30 lived-in homes with human-reviewed object arrangements.

The paper introduces Sora100K, a dataset structuring AI-native video creation workflows into multi-turn creation trajectories and editing operations.

Introduces a benchmark evaluating vision-language models on graph-theoretic reasoning where evidence is split across multiple visual graph views.

Introduces an audio-visual caption benchmark using atomic evaluation units to evaluate multimodal large language models reliably.