Corpus Query Language (CQL) supports precise corpus analysis but requires expertise in annotations, regular expressions, structural scopes, and platform-specific syntax. Natural-language interfaces simplify access, but their answers can be difficult to inspect, reproduce, or revise. We study Text-to-CQL as executable semantic parsing, mapping intents to inspectable queries. We introduce ECQL, a Sketch-Engine-oriented benchmark of 3,572 natural language–CQL pairs comprising 811 tutorial-derived and 2,761 synthetic examples, with a default split of 3,215 training and 357 test examples. ECQL focuses on Sketch Engine normalization, CQL Diagnostic Profile (CQL-DP) annotations, controlled semantic contrasts, audit-oriented metadata, and generalization protocols, alongside an evaluation suite covering syntactic validity, exact match, canonical match, and execution match. The full 357-example headline test set is independently double-annotated and adjudicated, with 96.9% agreement for intent clarity, 94.7% agreement for CQL–intent alignment, and a 3.1% post-adjudication intent-match error rate. Experiments with prompting, retrieval-augmented prompting, syntax repair, DeepSeek-R1 reasoning prompting, a GPT-4o Agent/API alternative, and supervised fine-tuning reveal persistent difficulties with corpus-query semantics. The best fine-tuned model reaches 85.6% execution accuracy on the random split, but drops to 68.9% on template-disjoint evaluation. ECQL provides a reproducible testbed for auditable semantic parsing over specialized linguistic query languages.

In this paper, we propose NUWA-XL, a novel Diffusion over Diffusion architecture for eXtremely Long video generation. Most current work generates long videos segment by segment sequentially, which normally leads to the gap between training on short videos and inferring long videos, and the sequential generation is inefficient. Instead, our approach adopts a “coarse-to-fine” process, in which the video can be generated in parallel at the same granularity. A global diffusion model is applied to generate the keyframes across the entire time range, and then local diffusion models recursively fill in the content between nearby frames. This simple yet effective strategy allows us to directly train on long videos (3376 frames) to reduce the training-inference gap and makes it possible to generate all segments in parallel. To evaluate our model, we build FlintstonesHD dataset, a new benchmark for long video generation. Experiments show that our model not only generates high-quality long videos with both global and local coherence, but also decreases the average inference time from 7.55min to 26s (by 94.26%) at the same hardware setting when generating 1024 frames. The homepage link is [NUWA-XL](https://msra-nuwa.azurewebsites.net)