Declarative animation specs made models write 25 to 48 percent fewer tokens but were not cheaper, faster or more reliable, and one open model produced valid specs in only 33 percent of jobs
AspireAI community · arXiv (Jiantang Huang)
The assumption that declarative animation specifications beat free-form HTML and JavaScript because they are shorter and constrained is tested directly here, with hypotheses and analysis plans frozen before the runs. ExplainSpec, a time-indexed declarative animation format, and an observation contract that makes generated explanations programmatically inspectable were validated by pacing-agnostic checkers across 20 explanation requests from five domains. Three Claude models ran 360 jobs, with a follow-up of 240 jobs for the open-weight Qwen3-14B with thinking enabled and disabled, and every model answered every request in both formats. For every model the specifications were 25 to 48 percent shorter in answer tokens, yet total output fell significantly only for the largest Claude model and no model generated significantly faster, because the Claude models reasoned far more before writing specifications, from 15 to 20 percent up to 85 to 88 percent of generations for the two larger ones, and Qwen3-14B reasoned about as much in both formats. Reliability moved the other way: free-form pages were valid after one repair round in 95 to 100 percent of jobs for every model, whereas Qwen3-14B produced valid specifications in only 33 percent of jobs and passed all semantic tests in 17 percent, against 62 percent for HTML and JavaScript.
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.