Analyzes spatial representations in vision–language models and reveals a consistent vertical-position ↔ distance entanglement; introduces SpatialTunnel, a synthetic benchmark that exposes this perspective-driven shortcut, and provides code and a project page.
AI Video Papers2026
Enables real-time streaming video-to-video editing (1280×704 @24 FPS) on a single RTX 5090 GPU. Uses a Hybrid Diffusion Transformer for balanced local/global modeling, Cycle‑Reverse Regularization for temporal consistency, and system-level mixed-precision and fused kernels to maximize throughput.
Reallocates injected noise energy across frequency bands to match a diffusion model's spectral bias, improving sampling fidelity without retraining. Uses a timestep- and frequency-dependent colored-noise schedule as a plug-and-play inference-time SDE solver; shows sizable FID drops on ImageNet-256.
Synthesizes high-quality targets for real-world image restoration by using multimodal foundation models (MFMs) to convert real low-quality photos into HQ references. Provides GGT-100K (103,707 LQ–HQ training pairs + 500 test pairs) with multi-stage quality control and demonstrates consistent generalization gains for a range of restoration models, especially for finetuning generative restorers.
Omnimodal world model that jointly processes and generates text, images, video, audio, and action trajectories for physical AI. Uses a mixture-of-transformers to combine autoregressive reasoning and diffusion-based multimodal generation; released open-source with checkpoints, datasets and benchmarks for robotics and simulation.
AI Agent Papers2026
Workflow-aware benchmark for autonomous medical-AI research that splits agent execution into five stages (Plan, Setup, Validate, Inference, Submit) and evaluates long-horizon runs across segmentation, image enhancement, VQA, report generation, and lesion detection with stage-level scoring.
Evaluates multimodal LLMs on streaming egocentric video for spatial intelligence using 1,680 human-annotated questions across 348 videos; organizes tasks into four hierarchical levels (perception → tracking → simulation → allocentric mapping) and highlights allocentric mapping as the main bottleneck.
Studies when and how to combine visual future rollouts from world models with abstract reasoning in multimodal LLMs. Proposes PF-OPSD — a teacher-student distillation that uses ground-truth future videos during training — and evaluates on two human-verified benchmarks, improving accuracy ≈10% while improving robustness to noisy rollouts.
Trains a GPT-style causal Transformer on a 2-billion-frame retargeted motion corpus to enable zero-shot whole-body motion tracking and control. By scaling both data and model capacity, it tracks highly dynamic behaviors while generalizing to unseen motions; accepted to CVPR 2026.
Explores how training recipe — data composition, teacher guidance, and task mixture — shapes few-step distillation for text-to-image generation and instruction-guided image editing; introduces Qwen-Image-Flash and empirical findings that training pipeline organization matters as much as distillation objectives.
Large-scale training corpus for knowledge- and reasoning-intensive video understanding: 315K video reasoning examples over 145K CC-licensed expert-domain videos, with human-in-the-loop chain-of-thought rationales to strengthen post-training for video reasoning. ([arxiv.org](https://arxiv.org/abs/2606.05259))
AI Video Papers2026
Decouples perception and reasoning for hours-long videos by streaming inputs into a three-tier Hierarchical Graph Memory and using an agentic Observation–Reason–Action retrieval loop; reduces reasoning context to ~2% of full video while improving benchmark accuracy.