LiveSearch froze the questions and let the answer expire, and found that 70 percent of served finance questions changed their gold value in three days
AspireAI community · arXiv
*Most financial benchmarks freeze the answer at annotation time, which quietly rewards recall over retrieval - and lets stale gold answers leak into training data.* LiveSearch inverts it. Each question is stored as executable verifier code: a typed program of calls to a professional financial data service followed by a declarative extractor. The question text is frozen, but the gold answer is determined by running that program against live data at evaluation time, with a daily job re-executing it and no human or language model in the loop. Construction stays automatic too - the language model only verbalizes programs enumerated from real tool responses and never picks an answer.
Under general web search, six contemporary tool-calling models reach 70.9 to 76.7 percent answer accuracy. The number that matters for benchmark builders: over three consecutive days, 70 percent of served questions changed their gold value, and even with tolerance-aware matching a literally reused earlier snapshot fails on a majority of questions within a single day. Per-question tracking also surfaces dynamic correctness shifts that aggregate accuracy hides. The limitation is scope - it benchmarks professional financial data specifically, so the exact corpus and its update cadence set the ceiling.
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.