Video generation is moving behind an agent harness, and the harness becomes the thing you improve
AspireAI community · arXiv
The paper treats video generation as an agent problem rather than a synthesis problem. A harness is an executable runtime that builds context, exposes tools, keeps state, routes feedback and manages budgets around a video model. From that framing the authors split the work three ways: agentic video generation, where the harness coordinates scripting, shot planning, asset management, production and editing; agentic world modeling, where the harness mediates observation, control, execution and state, sometimes with a game engine holding world state while the video model renders observations; and recursive self-improvement, where the authors separate four different things that can change - the generated output, the harness state, the generator parameters and evaluators, and the improvement procedure itself - and show how each feeds back into the next round. It is a position paper with a research agenda rather than a single system, and the evaluation challenges it raises are the open part. It argues that capabilities like long-horizon planning, memory, code-based interaction and multi-agent coordination are the ones that decide whether autonomous video production actually works, and it lists provenance, cost accounting and evaluation as unsolved.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.