Four new AI video papers, ranked by how soon you can act on them

Four papers posted to arXiv in the first week of November 2026 approach AI video from four directions: catching deepfakes, scoring generation quality, reasoning about events a camera does not see, and upscaling frames. None has been peer reviewed. Each is the authors' own paper and the single source for the claims below. arXiv hosts the abstracts and PDFs; the papers carry the authors' own results, not independent replication.

This list ranks the four by how close each one is to something a working creator or platform can act on, as of November 6, 2026. The scale runs from a benchmark with code you can download today to a method with no stated release. Each entry states what the paper says is available, and that availability sets the order. Every figure comes from the abstract or the arXiv page. Where an abstract states no headline number, the entry says so.

Four papers made the cut, and the countdown starts with the one furthest from a download.

4. PixelSpline (Yuan and Li)

PixelSpline works on continuous space-time video super-resolution, which raises resolution and frame rate together from low-resolution input. Wanqi Yuan and Nianyi Li model each pixel's motion as one continuous trajectory shared by both endpoints. The method warps endpoint colors into a coarse intermediate frame, then aligns high-resolution features for residual reconstruction. Long-context encoding gives the endpoints more reliable representations. The endpoints are the two frames the method interpolates between, and the spline is the path a pixel takes across the gap. The abstract reports a favorable quality-efficiency trade-off across spatial and temporal upsampling settings.

The abstract states no headline metric, and the arXiv page lists no code, weights or dataset release. For a creator or platform, that makes PixelSpline the hardest of the four to act on: there is no number to check against a current upscaler and no code to run. The absence of a stated metric also leaves no claim to compare with a named baseline. It ranks below EvoWorld, which at least names the data and benchmark behind its claim. A team building its own super-resolution stack can take the shared-trajectory idea and reimplement it. Everyone else waits for a release that the paper does not promise.

• Paper: arXiv:2611.06870
• Authors: Wanqi Yuan, Nianyi Li
• Submitted: 4 November 2026
• What it measures: spatial and temporal upsampling quality against efficiency
• Released: no code, weights or dataset stated in the abstract
• Read it: arxiv.org/abs/2611.06870

3. EvoWorld (Bie and co-authors)

EvoWorld proposes a task the authors call Out-of-Sight Causal Evolution. A camera leaves a subject, an event happens off screen, and the model must render a causally plausible outcome when the camera returns. The abstract's example: a wild beast appears, and the model has to infer whether a nearby person flees, takes cover or fights. The authors argue that generative video models are increasingly used as visual world simulators, yet their strength is rendering plausible pixels rather than reasoning about how a world should evolve. The authors set the goal as world evolution beyond the current field of view. The baseline pairs a vision-language model acting as a world brain, which observes generated segments and keeps an evolving account of the world, with a segment-autoregressive video renderer that draws each segment.

The paper introduces EvoWorld-Data and EvoWorld-Bench for training and evaluating event realization, causal reactions and revisit consistency. The abstract claims substantial improvement over recent models while preserving visual quality and consistency, and states no headline percentage. A platform can act on the task definition and bench design now, and the paper works as a specification for a team building long-form scene continuity. The abstract does not say either asset is public, and the framework needs a VLM plus a renderer, so the engineering lift is larger than downloading a dataset. EvoWorld ranks above PixelSpline for naming measurable assets and below ArtiVidRM for leaving their release unstated.

• Paper: arXiv:2611.06748
• Authors: Yequan Bie and 7 co-authors
• Submitted: 4 November 2026
• What it measures: event realization, causal reactions and revisit consistency
• Released: EvoWorld-Data and EvoWorld-Bench named; public release not stated in the abstract
• Read it: arxiv.org/abs/2611.06748

2. ArtiVidRM (Li and co-authors)

ArtiVidRM is a reward model for video generation that follows an observe-then-score pattern. Caorui Li and 14 co-authors make the model first identify visible generation artifacts and organize them into a structured artifact list, then use that list as explicit rubrics for scoring quality. The authors say existing methods either map videos directly to opaque scalar scores or generate free-form rationales that are not explicitly optimized as the basis of the final judgment. The intermediate artifact list grounds each score in video-level evidence. A multi-reward reinforcement learning strategy optimizes artifact prediction and quality assessment together. The reward splits into objectives for semantic coverage and prediction efficiency, plus a pairwise preference objective that ranks quality by score differences.

The abstract claims state-of-the-art performance on video preference benchmarks and effective signals for video diffusion optimization. The capability a platform can act on is artifact-guided refinement of closed-source video generators: a scorer that names visible defects and feeds that assessment back into generation. The artifact list also gives a team a checklist to inspect by hand, which a single opaque score does not. The abstract states no headline number and no public code or weights. ArtiVidRM ranks above EvoWorld because the scoring and refinement loop can be applied to models a team does not own. It ranks below DeepfakeGenome because the abstract lists no public download.

• Paper: arXiv:2611.07301
• Authors: Caorui Li and 14 co-authors
• Submitted: 5 November 2026
• What it measures: video quality preference, via artifact rubrics
• Released: no code, weights or dataset stated in the abstract
• Read it: arxiv.org/abs/2611.07301

1. DeepfakeGenome (Wang and co-authors)

DeepfakeGenome is a detection and attribution benchmark from Shunli Wang and eight co-authors. It contains 100 facial forgery algorithms and 2 million images, which the authors say is 4 times to 100 times larger than prior deepfake-attribution benchmarks. The authors write that existing work in the field faces saturated performance in binary classification, limited diversity in datasets and algorithms, and imperfect evaluation protocols. DeepfakeGenome defines four evaluation protocols, including a retrieval-based attribution paradigm aimed at blacklist registration. That is the setup where a platform adds a known generator to a list and checks new uploads against it, a closer match to active defense than the open-set metrics the paper criticizes.

The authors report more than 2,000 experimental evaluations and 10 findings, and they frame the work as an investigation of the performance ceiling of deepfake attribution. The arXiv listing puts the write-up at 28 pages, which is where the protocols and findings sit. The code and benchmark are on Hugging Face, so a platform team can download the data and rerun the protocols now. That release is why DeepfakeGenome takes first place: among these four abstracts, it is the only one that states a public code and benchmark release. ArtiVidRM offers a scoring method for refinement, but a team cannot download it yet.

• Paper: arXiv:2611.06046
• Authors: Shunli Wang and 8 co-authors
• Submitted: 4 November 2026
• What it measures: deepfake detection and generator attribution
• Released: code and benchmark at huggingface.co/spaces/DeepfakeGenome/DFG_Page
• Read it: arxiv.org/abs/2611.06046