Giving a video-editing agent one read tool and one write tool, with every write going through the same commit path as manual edits, let it repair nearly every edit it broke and cut the tokens needed per correction by more than half
AspireAI community · arXiv
Language agents usually work by calling tools on a persistent artifact such as a code repository, and each call shapes the next step. Video editing is a hard case because what a frame shows depends on how every clip is trimmed, retimed, cropped and nested, so knowledge about the footage holds only for particular uses of it and an invalid edit can leave the timeline structurally broken. Existing video editing agents keep their edits valid with critic agents, reviewers or restricted output formats. Cassette-Editor instead uses one tool that reads the project and one that writes to it, with every write passing through the same executor, rules and commit service as the user's manual edits, committing atomically or being rejected with located diagnostics and repair options while the model keeps the creative choices. In a controlled experiment a real model was shown 277 of its own committed edits after they were damaged: it corrected nearly every call with full or minimal feedback, locating each problem cut the input tokens a correction needed to less than half, and in one case its corrections dropped other values it had written rather than seven. The limitation is the scale of the experiment, 277 edits in one controlled setting, and the paper notes evaluation will be extended in a future version.
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.