AI video generators can make every frame look convincing while the sequence breaks basic physics: a glass bounces instead of shattering, a row of dominoes topples but leaves the vase untouched. PhysLoop, a preprint from researchers at Peking University, ByteDance Seed, MIT and others, tries to fix that by putting a physics simulator between a language model's knowledge and the final video. It is a research system, not a product, but it shows one practical route to more believable AI video.
The method works in stages. A coding agent, by default OpenAI's GPT-6 Astra, turns a request into a task specification, records assumptions such as release height or surface, gathers 3D assets and builds a scene in Unreal Engine 5.7. The simulator runs the event, a verifier checks it against the specification, and only then does a video model, by default Seedance 2.5, refine the appearance; the final video is checked again. The same pipeline can repair an existing generated video by rebuilding just the faulty segment. The team is releasing PhysLoop-100K, 100K videos from 2.5K distinct physical events, plus two test sets.
The results are clearest on repair. On PhysLoop-Repair the full system succeeded 86.00% of the time, against 19.33% for Seedance 2.5 alone, and every standalone generator stayed below 20%. On generation tasks it reached 85% success against 73%, with higher human ratings for physical plausibility, 4.06 against 3.80, but lower visual realism, 3.72 against 4.29. On the external Physics-IQ Verified benchmark it scored 52.06 against 48.07; swapping in Claude Opus 5 as the agent gave the best score, 53.75, despite lower success elsewhere. The authors put the cost of ten generation tasks at $42.99 in video charges plus model tokens, and each video takes roughly 7 to 12 minutes.
The approach sidesteps a known weakness: general-purpose evaluators and video models judge physics poorly, while a simulator does not have to judge, it computes. That makes it closer to the verifier-driven agent designs now common in coding tools than to prompting a video model harder.
The limits are substantial. This is an arXiv preprint, not peer reviewed. The headline framing of recursive self-improvement is aspirational; the paper's own limitations section says a full training loop still needs weakness discovery and data selection connected together. Results depend on the simulator's coverage and assumed materials, pairwise win rates were scored by a single author without blinding, and the automatic quality metrics rate PhysLoop slightly below the standalone generators.