Drex for Search points coding agents at the right file before they read, and claims 62 percent fewer steps
AspireAI community · Nace.AI
Nace.AI released Drex for Search, a retrieval layer that uses its Drex decision model to decide where an answer can be before an agent reads anything. It classifies the question first (quick lookup, or an agent that reads documents and queries tables), then walks the workspace level by level (folder, file, section, page, and inside a sheet, table and column), keeping only what could hold the answer, so depth is fixed and a million files take the same number of steps as a thousand. The vendor reports across 6 professional-work benchmarks and 529 questions that coding agents take 62 percent fewer steps, use 56 percent fewer input tokens, get 6 percent more answers right (63.1 to 67.1 percent), and finish 2.3x faster (median 80.8 s to 35.5 s per question). On Databricks OfficeQA Pro (133 questions over 697 scanned PDFs) ten harness and model pairs all used fewer turns and fewer tokens, and eight of ten answered more questions correctly, though two of those ten lost accuracy. On public leaderboards it reports first on MADQA at 94.7 percent and first on VisDocAgentBench at 77.5 R@1, and second on BrowseComp-Plus at 92.05 percent using 16.6 times fewer search calls than the leader. These are the vendors own numbers on its own systems, and the OfficeQA Pro accuracy comparison mixes a two-run average with a single run, so the vendor itself asks readers to treat accuracy as plus or minus 2 points.
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.