Seed beats brute-force scaling on intent benchmarks

Interesting result, but not a clean “smaller is always better” story.

• –The strongest signal is efficiency: dynamic Seed variants are roughly 4-5x smaller in parameters than the logistic/static baselines on several datasets.
• –Banking77 looks like the best case for the claim, with distilled dynamic Seed improving both accuracy and F1 over TF-IDF.
• –CLINC150 and HWU64 show the tradeoff more clearly: smaller models stay in the same ballpark, but they do not consistently win on quality.
• –MASSIVE is mixed as well, which suggests the method is dataset-sensitive rather than universally dominant.
• –Distillation appears to stabilize the dynamic search output, especially when the raw discovered architecture is too small or noisy.
• –As a product story, this is more credible as an architecture-search/efficiency narrative than a new model release.