All papers

Introduces a multi-dimensional benchmark evaluating reasoning-informed visual editing across temporal, causal, spatial, and logical operations.

Introduces PEBench, a benchmark and evidence-grounded framework to directly evaluate prompt enhancers for video generation without rendering full videos.

The paper introduces benchmarks and summarizes winning solutions for city-scale audio-visual reasoning and spatial intelligence at the Perception Test 2026.

The paper introduces SightCAD, an active view selection framework optimizing camera viewpoints under occlusion constraints for parametric CAD reconstruction.

The paper proposes Vantage, a training-free framework pairing VLMs with 3D models to plan and synthesize relevant viewpoints for spatial reasoning.

The paper introduces Memory Forcing, partitioning cache into archive and working banks to prevent mid-horizon history loss in streaming video generation.

Introduces a differentiable token merging operator running in linear time with an exact sequence output budget.

Proposes a graph-based token pruning method inspired by sheaf geometry to reduce prefill computation in multimodal large language models.

The paper presents an elastic vision architecture that dynamically reconfigures width and supports partitionable inference across heterogeneous edge devices.

The authors propose decoupled positional encodings enabling full-duplex speech models to run internal thinking streams at hardware-adaptive speeds.

The paper introduces ViMoD, maintaining a compact visual context while dynamically routing back to fine-grained tokens during multi-step reasoning.

The paper presents a simulated benchmark across eleven unconstrained environments to evaluate vision-language models on documenting real-world human physical work.

The paper repairs multimodal chain-of-thought final answers using original image features without altering the generated reasoning trace.

The paper presents Diffuplex, a full-duplex spoken language model that separates speech timing from content generation using discrete diffusion.

This paper develops an uncertainty estimation framework for omnimodal models using external embeddings to score reasoning grounding and coherence.

This paper introduces an environment benchmark and exploration framework to help graphical user interface agents learn unfamiliar interaction dynamics.

The paper presents OpenMobile-2, combining open mobile environments, trajectory datasets, and hybrid GUI-tool actions for mobile agents.

Introduces a model-agnostic, concept-level attribution method for black-box vision models by masking concepts located with vision-language models.

Investigates multimodal LLM failures on squint-readable low-frequency text and provides a diagnostic benchmark alongside training interventions.

Analyzes layer-wise visual interventions in multimodal language models to identify where visual edits effectively reduce hallucinations.

The paper presents a benchmark dataset evaluating unified page-level document parsing alongside alternative text descriptions for visual elements.

The paper presents a benchmark evaluating how effectively vision-language models recover fully occluded objects using remaining visual cues.

The paper introduces a benchmark to measure demographic robustness, stereotyping, and intentional personalization in speech language models.

Introduces a Minecraft diagnostic benchmark testing whether multiplayer world models maintain consistent and persistent representations across independently controlled views.

Tests whether self-supervised vision transformers represent semantic meanings of rendered words without paired text data.

Proposes a framework that encodes visual trajectories into reusable skill cards to reinforce vision-language agent planning and rewards.

Develops a framework for egocentric multimodal assistants to decide both when to speak and what guidance to provide.

Presents a rolling masked diffusion method that accelerates full-duplex spoken dialog models by forecasting multiple future interaction frames.

Introduces a generate-verify-optimize pipeline to enable non-verbal visual reasoning in machine learning systems.

Detects talking-head deepfakes in dynamic videos by matching speaker distance estimates with distance-dependent acoustic energy variations.

This paper introduces a benchmark to diagnose the ability of video language models to count repeated action cycles.

Proposes a two-phase token pruning strategy for vision-language models and identifies evaluation flaws in common decode-phase benchmarks.

This paper analyzes how different visual representation types affect the 3D spatial reasoning capabilities of multimodal large language models.

This paper introduces a training-free framework combining multi-hypothesis search and local visual verification to ground interface instructions in graphical user interfaces.

Uses sparse autoencoders to select diverse training samples and prevent underrepresenting tail concepts across vision and multimodal models.

Develops an agent-centric 3D language model that grounds spatial understanding in physical poses without requiring text situation descriptions.

Develops an interpretable AI framework for breast ultrasound diagnosis incorporating clinical standard descriptors and assessment workflows.

The paper proposes a framework that evolves a reusable library of skills and tools based on experience for spatial reasoning agents.

The paper introduces a method to detect adversarial chain-of-thought hijacking in backdoored multimodal reasoning models by monitoring reasoning dynamics.

The paper analyzes failure modes in unified visual generators using self-reward loops and proposes stabilization methods to prevent quality degradation.

The paper introduces an image augmentation pipeline that generates diverse visual variants to anchor multimodal chain-of-thought reasoning in visual evidence.

The paper proposes an agent harness that autonomously evolves questioner and solver workflows from unlabeled video data without updating base model parameters.

The paper presents a predictive KV cache quantization method exploiting temporal predictability to reduce memory overhead in autoregressive video generation models.

This paper introduces a framework that acquires and directly executes reusable visual programs to improve multimodal reasoning agent reliability.

The paper proposes a distillation framework to remove specific multimodal knowledge without causing hallucinations or leaving residual queryable information.

Introduces a benchmark probing physical perception tasks related to humans, including depth, trajectory, illumination, and kinematics.

Introduces VARI, a group-structured intent alignment framework to improve text-to-video generation robustness across equivalent prompts.

This paper introduces a grounded alignment framework using anomaly detection datasets and patch-level policy optimization to fix distortions in flow generators.

TableVision recovers verified table structure from historical documents using dynamic programming and column sum checks.

This paper presents a verification and recovery framework using vision-language models to correct tracking errors and mask drift in video object segmentation.

This paper introduces a benchmark evaluating whether vision-language models can infer underlying physical laws from motion videos and generate runnable simulation code.

Uses generative models and vision-language labeling to detect and synthesize missing semantic domains in test datasets.

This paper introduces an uncertainty-guided refinement framework to reduce hallucinated responses in large vision-language models.

The paper proposes a reinforcement learning framework using natural language metrics to reduce object hallucination in vision-language image captioning.

The paper presents an end-to-end trained multimodal agent that combines visual graphical interface actions and executable code execution.

This paper studies when audio-visual reasoning harms grounding and trains a probe to determine if chain-of-thought is needed before generation begins.

The paper proposes a selective safety framework for vision-language models using counterfactual pairs and representation density estimation.

The paper proposes using test-time training memory to maintain consistent long-horizon visual scenes in real-time camera-controlled video world models.

The paper introduces an in-context learning method that selects demonstration examples by mining a vision-language model's own past prediction mistakes.

Proposes a training-free framework that mitigates vision-language model hallucinations by admitting claims only when supported relative to a visual-null input.