A search harness with no vector database answered 85 percent of a hard retrieval benchmark.
AspireAI community · arXiv (Context as Infrastructure / RAGFlow)
**A search harness with no vector database answered 85 percent of a hard retrieval benchmark.**
Retrieval in agentic systems is turning into a harness problem rather than a retriever problem. A harness that only writes queries, reads results, reformulates and decides when to stop, using nothing but BM25 and grep plus MiniMax M3, answered 85 percent of BrowseComp-Plus and 91 percent of multi-hop FRAMES in the authors' runs, with no compiled or vector index.
What that exposes is a context problem, not an accuracy one. The undisciplined loop repeats queries, rereads documents and fills the window with unused passages. The proposed fix is a knowledge compilation layer, and the piece they measure is a navigation tree whose leaves are claim-evidence pairs: it cuts query tokens by 91 percent when the evidence suffices, and a decision gate cuts latency from 140 seconds to 30.
The honest limit: the authors find compilation gives no detectable accuracy gain on FRAMES, and the benefit shows up in deployment rather than on benchmarks.
https://arxiv.org/abs/2610.41172
Today's stories are all about what a model or agent does with what it has been given: memory it keeps but rarely needs, tokens a search harness pulls in, context a personal agent has to reconstruct, and code an agent runs without being asked. Which of those is your agent getting wrong today, and how would you measure it?
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.