Daily-Omni is a multiple-choice audio-visual QA benchmark with 684 real-world videos and 1,197 questions across 6 task families that explicitly require cross-modal temporal reasoning. We build it with a semi-automatic pipeline covering annotation, cross-modal consistency refinement, temporal alignment elicitation, and text-only leakage filtering, followed by human verification. The benchmark includes diagnostic evaluation over 24 foundation models under 37 model-modality settings, together with a training-free modular baseline built from off-the-shelf unimodal models. Results show that many end-to-end MLLMs still struggle on alignment-critical questions, highlighting robust cross-modal temporal alignment as an open challenge.

BizGenEval benchmarks image generation models on real-world commercial design tasks with dense textual, layout, attribute-binding, and knowledge constraints. It covers 20 evaluation tasks across slides, charts, webpages, posters, and scientific figures, with human-verified checklist-based scoring for systematic comparison.

AVGen-Bench is a task-driven benchmark for evaluating text-to-audio-video generation with fine-grained measurements over visual quality, audio quality, synchronization, text fidelity, and physical plausibility. It emphasizes joint audio-video assessment and more complex task-oriented prompts than prior benchmarks.

SkillOpt treats a compact natural-language skill document as the trainable state of a frozen agent, optimizing it through scored rollouts, bounded textual edits, and held-out validation gates. The learned skills are reusable deployment artifacts that improve agent performance without changing model weights or adding inference-time model calls.