Research: AI news.

The results suggest domain-specific fine-tuning improves diagnostic accuracy and reduces hallucinations compared with an untuned general model.

The large-scale paired dataset and unified training approach offer a foundation for building more capable instruction-driven video editing systems.

The results suggest the compositionality gap in such models may stem from training distribution design rather than a fundamental architecture limitation.

It reports a deployed system for rewriting ambiguous search queries that improved engagement metrics in an online A/B test.

It frames the density of utility-relevant interactions as a measurable coordinate of scaling efficiency, so useful compute may describe scaling better than total compute.

The work targets real-time, mechanically consistent control in generated game worlds, an area the authors say prior approaches handled poorly outside simple, predefined settings.

The results suggest that broader video coverage and even sparse action supervision can improve robot task transfer to new tasks.

It suggests existing autoregressive image models can gain inpainting, outpainting and editing without retraining from scratch or replacing the standard AR inference stack.

The authors report that a single 5B-parameter checkpoint reaches state-of-the-art on OmniDocBench v1.6 and leads open-weight models on OmniParsingBench, with its advantage concentrated on audio and video.

The authors report that positive views, rather than scale alone, are the lever to build aligned multimodal representations, and that next-token prediction is an inefficient route to them.

The results suggest parameter-derived relational supervision gives a stronger synthetic pre-training signal than label-based classification for transfer beyond recognition tasks.

If it holds up, it reframes model training as choosing what data to train on and when, rather than maintaining distinct staged pipelines.

The text claims broad vision-language competence and continuous closed-loop flight can coexist on a resource-constrained aerial platform, which the authors say existing systems struggle to do.

It reduces the quadratic cost of 3D attention in video diffusion transformers while reportedly improving fidelity over existing sparse attention baselines.

The work targets object insertion that requires pose and lighting adaptation rather than appearance copying, and reports improved image fidelity and user-rated placement and lighting integration.

The work suggests online RL on live apps can make smaller UI navigation agents more robust than supervised or offline-trained models, according to the reported benchmarks.

It lets one resident model handle reranking and answering instead of two separate models, potentially avoiding duplicate model residency and passage processing.