ImageDoctor diagnoses text-to-image generation through grounded image reasoning, enabling systematic evaluation and debugging of T2I models by identifying misalignments between prompts and generated images.

XModBench is a comprehensive benchmark for evaluating cross-modal capabilities and consistency in omni-language models, assessing how well models handle information across different modalities.

Latent Visual Reasoning enables models to perform complex visual reasoning in latent space, improving efficiency and accuracy on multi-step visual reasoning tasks.

A framework that enables computer-use agents to learn from online video tutorials at inference time by retrieving, filtering, and converting videos into structured demonstration trajectories for dynamic in-context guidance during execution.

GEAK is an AI-driven framework for generating and evaluating Triton GPU kernels, leveraging frontier LLMs with inference-time scaling to automatically produce efficient, accurate kernels for AMD Instinct GPUs.

CaptionQA is a utility-based benchmark evaluating caption quality across 4 domains with 33,027 annotated questions, revealing up to 32% gap between image and caption utility in state-of-the-art MLLMs.

APRIL mitigates long-tail inefficiency in RL training by over-provisioning rollout requests, recycling incomplete responses, and reducing GPU idle time — achieving up to 44% throughput improvement and 8% higher accuracy across GRPO, DAPO, and GSPO algorithms.

SAND-Math, a scalable pipeline that generates and enhances challenging math problems, enabling LLMs to achieve state-of-the-art results on difficult mathematical reasoning benchmarks like AIME25.

TTT-Bench uncovers the hidden blind spots of today’s smartest AI, challenging them with simple, human-intuitive games — and showing that even the best models often fail where humans excel effortlessly.

TaDA slashes KV cache memory usage by over 70% without sacrificing accuracy — enabling longer, smarter, and more scalable LLM inference with zero retraining.

“Hour-LLaVA, built on the new VideoMarathon dataset, enables efficient training and inference on hour-long videos and achieves SOTA on long-form video-language tasks.”

AgenticLU unlocks the full potential of LLMs on long-context queries, combining self-driven clarifications and smart context retrieval to deliver robust, scalable, and state-of-the-art reasoning.

MOVi, a training-free framework for multi-object T2V generation with LLM-guided trajectory control and attention refinement, achieving 42% better motion and object accuracy.

KeyVID leverages audio-aware keyframes and motion interpolation to generate synchronized, high-quality audio-to-visual animations with improved dynamic motion handling.

Agent Laboratory revolutionizes scientific discovery by automating the entire research workflow — empowering researchers to focus on ideas, not grunt work, with up to 84% lower cost and state-of-the-art results.

SoftVQ-VAE unlocks fast, efficient, and high-quality image generation with ultra-compact tokenization — delivering up to 55× faster inference and competitive FID, all with fewer training iterations.