A venture fund's agentic triage harness scored investments at rho 0.45 against realised multiples, and its own authors say that does not prove screening skill
DealsFlow is a human-in-the-loop agentic harness deployed at a UK early-stage fund to triage startup pitch decks. The LLM reads the deck, cites evidence and rates each category, while code combines the ratings, caps them and can only ever lower them: a stage-aware corroboration ceiling of 78, 70 or 60 out of 100 on factual categories without external evidence, a sceptic call that may lower but never raise a score, and a pre-fetch pattern that indexes all evidence before a single generation pass. On the 142 investments at least three years old from a 243-deal portfolio, the full pipeline's score correlates with total-value multiple at Spearman rho 0.45 with a 95 percent confidence interval of 0.32 to 0.57, and reaches an AUC of 0.71 for identifying write-offs, against rho 0.38 and AUC 0.68 for an unconstrained LLM. The authors are careful about what this means: later successes had 64 percent of factual categories corroborated against 38 percent for write-offs, retrieval was not date-restricted, so hindsight cannot be excluded, and the paper states plainly that the results do not establish screening skill. In a separate study of 20 inbound applications, analyst time fell from 44.7 to 18.9 minutes per application, 84 percent of 240 cited claims were fully supported by their sources, and 70 percent of reports still needed at least one correction.
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.