Seed AutoArch tops BANKING77 with 94.42%

Seed AutoArch proves that "structure over scale" is a viable path for high-performance AI in resource-constrained environments. By prioritizing architectural discovery and efficient reranking over brute-force parameter scaling, it offers a production-ready solution for complex intent detection without the overhead of LLMs.

• –The 94.42% accuracy is a 0.59pp improvement over the standard baseline, showcasing the power of task-specific optimization.
• –A 68 MiB memory footprint makes it ideal for edge deployment, on-device processing, and high-throughput financial services.
• –Eliminating generative LLMs from the pipeline removes token-based costs and significantly reduces the risk of hallucination in classification.
• –The strict full-train protocol and 5-fold CV confirm the result's robustness, providing a reliable alternative to opaque "black box" models.
• –This result challenges the current trend of uniform scaling, highlighting the massive potential of lightweight, specialized architectures for real-world developer tasks.