All papers

Unifies robotic policy learning and future state prediction using an autoregressive Transformer trained on spatial sequence tokens.

Presents VibeEdit, an image editor conditioned on spatial canvas annotations to perform targeted object additions, modifications, or removals.

Introduces Compo and a spatial canvas interface enabling users or agents to guide poster layout and image generation.

Benchmarks classical planning against learning-based methods in real-time motion planning under deterministic and stochastic rotating hazard fields.

Proposes Syn-Omni, a framework combining modality-expert LoRA paths and progressive routing to balance shared and modality-specific omnimodal embeddings.

Introduces Dynamic Vector Decoding to efficiently map 2D and 3D perception coordinates into compact discrete tokens for multimodal LLMs.

Presents ME-World, an egocentric world model generating synchronized multi-agent first-person video streams with fine-grained physical interactions.

Presents RiCo, a neural physics simulation method that models rigid-body interactions via sparse local contact surface reasoning.

Proposes ContiLNN, which uses closed-form continuous-time modules to improve cross-slice continuity and restoration in 2D medical imaging.

Distills geometric privilege into vision-language models during training to improve spatial reasoning without inference overhead.

Introduces an annotated dataset of handwritten student answer sheets with complex formulas, text, tables, and noise artifacts.

Presents a method to improve zero-shot robot foundation model performance using causal reasoning traces generated from demonstrations.

Introduces a training-free 3D generation pipeline enabling local geometric and appearance control using 3D primitives.

Presents a bidirectional JEPA world action model with diffusion-steering model predictive control for robot state tracking and planning.

Factorizes scene-aware human-object interaction generation into motion-affordance prediction and affordance-conditioned motion synthesis.

Presents a top-down skill refinement system for robot self-improvement through structured execution feedback and reusable modular skills.

Introduces a benchmark and environment for evaluating coding agents on text-conditioned LEGO design with physical and discrete assembly constraints.

Trains an unsupervised vision-language reward model using intermediate layer representations and caption perturbations without human preference labels.

Presents an end-to-end joint-embedding predictive architecture that combines world modeling and action flow matching for robot planning and control.

Introduces a multimodal embodied agent policy learning framework using trajectory construction, distillation, and token-level evidence routing.

Proposes a multimodal agent providing grounded prompts and video clips to guide video editing models in professional sports motions.

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

Presents a simulation benchmark evaluating how embodied agents continually acquire basic skills and recombine them for complex tasks.

Proposes spatial-semantic token compression and layer routing to accelerate 3D vision-language model inference across multi-view inputs.

Proposes an advertising image editing method that identifies and selectively refines visual attributes driving click-through rates.

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

Presents a local agentic framework using smaller language models to automatically synthesize edge-case testing scenarios for autonomous systems.

Presents RESCUE, a counterfactual simulation framework to evaluate whether runtime failure monitor alarms enable beneficial robot interventions.

Introduces MAGMA-Bench to evaluate long-horizon embodied agent reasoning during dynamic task execution failures and multi-robot coordination.

Proposes a model-agnostic method to integrate biological domain knowledge into machine learning models trained on weakly annotated bioimaging.

Pre-trains embodied robot policies on human egocentric videos using learned continuous latent action representations.

Introduces a Ukrainian meme dataset and evaluates vision-language models on detecting persuasive and propaganda techniques.

Introduces a controlled Minecraft benchmark to evaluate whether embodied agents acquire persistent, transferrable skills across environment resets.

Trains vision-language models to perform active spatial reasoning through novel-view generation and simulated mentalization trajectories.

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