Stanford HAI

Packed trajectories, not sequential rollouts, unlock high-fidelity multi-turn image editing by preventing visual drift.

VideoGen-Agent is presented, a multimodal agent trained through multitask agentic reinforcement learning to use external tools for video generation to support learning tool use across video-generation tasks and show that the trained agent can benefit from subsequent advances in generation tools.

This work introduces PosteriorBench, a benchmark for evaluating the distributional accuracy of generative inverse solvers, and reveals substantial distribution-matching gaps across current solvers, while showing that neural operators improve resolution robustness, and guidance weights and generation noise are key to posterior-variance calibration.

ToPO achieves superior image generation performance by optimizing token-conditioned preferences, outperforming existing methods across multiple evaluation metrics.

Achieving synchronized audio-visual generation and trajectory control, EchoWM sets a new standard for immersive media experiences in generative models.

Simulation-based pre-training can drastically improve the dexterity of robotic hands, outperforming traditional training methods with just a fraction of real-world data.

HyMeS enables robots to efficiently manage long-horizon interactions, achieving a 14.5-point increase in task success without requiring extensive demonstrations for every task configuration.

SmartMage redefines 3D scene understanding by dynamically selecting modalities, achieving state-of-the-art results while minimizing irrelevant computations.

VLA models can achieve a 66% success rate in contact-rich tasks by addressing precision and force failure modes, a leap from the previous 41% benchmark.

End-effector traces boost cross-embodiment transfer in robot manipulation, enhancing real-world task performance by 28% when leveraging simulation data.

Proprietary MLLMs may achieve high diagnostic accuracy, but they still struggle with reliable clinical reasoning, revealing significant gaps in their practical utility.

EmoScope redefines emotional image editing by enabling users to discover unique, context-specific editing strategies rather than relying on static templates.

FIDAC reveals how interpersonal distance can be accurately quantified from video, transforming facial detection data into actionable insights.

Radiological findings in frozen vision-language models are encoded by just 10 channels, revealing a surprising efficiency that challenges traditional assumptions about model complexity.

Achieving a staggering 250,000x speedup in hologram synthesis without compromising image quality could redefine the future of 3D display technologies.

Revealing robot motion in video models can transform how we predict and control robotic actions, achieving high fidelity with minimal training data.

BrainNext achieves top-tier performance in neuroimaging tasks by leveraging self-supervised learning on a massive dataset, highlighting the power of large-scale pretraining in medical applications.

Hierarchy-aware training combined with anatomy-guided learning boosts LUS video classification accuracy while enhancing model interpretability.

Systematic misalignments in MLLM-generated captions can be detected with 63.8% accuracy, revealing a critical flaw in image-text pairing that has been largely overlooked.

Task-dependent action frames can enhance robotic manipulation performance, with MoF outperforming traditional single-frame policies in both simulation and real-world applications.

Frontier foundation models can directly control robots through visual interfaces, achieving remarkable success without any fine-tuning on robot-specific data.

Pix2Act transforms complex 3D manipulation into a simpler 2D prediction task, leading to significant performance gains and robustness against camera variations.

AMP achieves millimeter-level precision in 3D manipulation by transforming action learning into a pixel classification challenge, drastically improving inference speed and success rates.

A novel AI-driven approach reveals that targeted facial stimuli can significantly enhance the detection of perceptual differences between autistic and neurotypical individuals.

VLMs are prone to critical failures that vary significantly across cultures, exposing the inadequacy of Western-centric safety benchmarks.

Faithful reasoning in VLA models can boost policy responsiveness to rare scenarios by 1.6x compared to state-of-the-art approaches, revealing a critical gap in current alignment strategies.

VLMs struggle with raw medical data, achieving only a 48.6% success rate in standardization, revealing a critical gap in their clinical applicability.

Diffusion models can outperform autoregressive counterparts in medical report drafting while offering a unique any-order infill capability that enhances usability for clinicians.

Robots can now learn new tasks on-the-fly from just one demonstration, revolutionizing how we teach machines to manipulate their environments.

Adapting pretrained policies with just a modest multisensory dataset can enhance robot manipulation performance across diverse tasks without sacrificing prior knowledge.

Synthesizing 48,000 interaction trajectories without human input enables a humanoid robot to learn complex loco-manipulation tasks effectively.

Touch is not just an add-on; it fundamentally enhances object representation, leading to dramatic improvements in physical property estimation and manipulation tasks.

Policies trained in SimFoundry's automated environments achieve up to 40% higher success rates in real-world tasks by leveraging affordance-preserving scene variations.

None of the 18 multimodal large language models audited are order-invariant, with flip rates revealing a staggering sensitivity to input ordering that challenges current evaluation practices.

Primitive steerability in VLAs allows for autonomous skill acquisition, enabling robots to learn new tasks without human demonstrations.

Current visual world models show a dramatic decline in performance when faced with unconventional and impossible physical interactions, highlighting a critical gap in their generalization capabilities.

Achieving state-of-the-art mesh quality with 18x faster inference times by directly generating triangle soups while respecting crucial symmetries.

Cloak enables VLA models to seamlessly adapt to new robotic embodiments without any additional training data, revolutionizing the way we think about robotic adaptability.

ChartWalker reveals significant performance gaps in cross-chart RAG tasks, challenging the status quo of existing benchmarks and paving the way for more robust multi-modal reasoning.

Advanced Vision-Language-Action models can be dramatically compressed by up to 50% without losing performance, reshaping our approach to robotic manipulation.

FlowObject bridges the gap between generative priors and real-world observations, achieving unprecedented fidelity in 3D reconstruction from sparse views.

DREAM-Chunk transforms action chunking by leveraging latent world models to enhance robustness against stochastic dynamics without the need for policy retraining.

Action-view augmentation can transform how robots adapt to unforeseen obstacles, boosting manipulation success rates significantly.

Achieving a dual-purpose tokenizer that excels in both clinical task performance and controllable 3D brain MRI generation could revolutionize how we approach medical imaging.

PermaVid achieves unprecedented long-term consistency in video generation, even after significant edits, by intelligently disentangling appearance and geometry in memory.

Integrating a learned world model with Flow Matching policies can dramatically boost performance in complex manipulation tasks, achieving higher success rates without changing the training process.

ProTrans reveals that modeling disease progression as directional semantic transitions can dramatically enhance the interpretability and accuracy of longitudinal chest X-ray analysis.

No single model dominates video embedding tasks, revealing stark contrasts in performance based on modality and task type.

OpenMedQ outperforms even the largest models in medical vision-language tasks, achieving state-of-the-art results with a fraction of the parameters.

High artifact detection rates in VLMs mask significant failures in contextual understanding, with top models misidentifying visual cues in over 46% of cases.