Thaddäus Wiedemer and coauthors argue in Google DeepMind’s paper that video models can reason

Thaddäus Wiedemer and coauthors argue in Google DeepMind’s paper that video models can reason

The paper was published on September 24, 2025. Thaddäus Wiedemer and coauthors report that Veo 3 can perform tasks it was not explicitly trained for, including segmenting objects, detecting edges, editing images and understanding physical properties. They say these abilities enable early visual reasoning, such as maze and symmetry solving. The authors draw a parallel between the simple primitives behind Large Language Models’ general-purpose language abilities—large generative models trained on web-scale data—and those used by generative video models. They argue that Veo’s reported zero-shot capabilities indicate video models are on a path toward unified, generalist vision foundation models. The paper’s argument matters because it presents video models as potentially able to understand and manipulate the visual world across tasks, not only generate video. The abstract does not detail how the capabilities were evaluated or establish how well they work beyond the reported examples.

Publication: 589bd553-c386-c1fa-770c-394dd515022f

attribution — pass: The reading attributes the claims to Thaddäus Wiedemer and coauthors in a Google DeepMind paper. This matches the listed authors and the publisher page. The text is a single paper abstract, not a listing.

Thaddäus Wiedemer, Effie Li, Paul Vicol, Shane Gu, Nick Matarese, Kevin Swersky, Been Kim, Priyank Jaini, Robert Geirhos

summary — pass: The summary faithfully conveys the abstract: Veo 3 performs untrained tasks, early visual reasoning such as mazes and symmetry, the parallel with LLMs, and the claimed path to generalist vision foundation models. Zero-shot, LLM and foundation model are defined beside the reading. Veo 3 is identifiable from context as a generative video model. The title's 'reason' matches the paper's own 'reasoners' framing. The question is substantive and non-leading, and the stated limitation is appropriate.

Could video models be on a trajectory towards general-purpose vision understanding, much like LLMs developed general-purpose language understanding?

date — pass: The reading states September 24, 2025, which matches the source dateline and evidence.originalPublished at day precision. It makes no post-publication or hindsight claims. The page was modified in December 2025, but the abstract is plainly the original paper text. The page kind is publisher-page, so no archive disclosure is needed.

September 24, 2025

support — pass: Each claim is directly stated in the abstract and kept as the authors' assertion. The significance sentence ('potentially able to understand and manipulate the visual world across tasks') is hedged and grounded in the source. The reading notes that evaluation details are absent rather than overstating reliability. The cited abstract is the core original text, so the later page modification does not cast doubt on it.

These abilities to perceive, model, and manipulate the visual world enable early forms of visual reasoning like maze and symmetry solving.

media — pass: No media was selected and no verified media candidates were supplied, so no media identity or relevance issue arises.