LLMs invent economic theories

An AutoTheory post links to a public demo and code repository for an experiment that uses large language models as the core search operator in an evolutionary discovery loop (x.com). A peer-reviewed team led by Anja Šurina published a paper titled “Algorithm Discovery With LLMs: Evolutionary Search Meets Reinforcement Learning” that formalizes coupling evolutionary search with RL-based fine‑tuning of the LLM and reports improved algorithm discovery on standard benchmarks. (arxiv.org) Xavier Gabaix and Ralph Koijen’s “In Search of the Origins of Financial Fluctuations” frames the price‑multiplier as a central empirical target — their granular-IV estimates imply that, on average, roughly $1 of flows can move aggregate market value by multiples around $5, which is why compact multiplier models are a focal test for any automated theory‑discovery method. (ssrn.com) Dividend‑strip pricing is an established empirical target used to decompose equity values into horizon‑specific cash‑flow components, and recent work estimates short‑term strip prices from S&P‑500 options spanning 1996–2022 to evaluate competing asset‑pricing models. (econpapers.repec.org) A separate line of recent work, LLM‑SR (symbolic regression with LLMs), has demonstrated that LLMs can propose human‑readable equations and program skeletons that pass numerical tests, a capability that underpins efforts to extract low‑parameter closed‑form models from economic toy problems. (github.com) The research community has started open‑sourcing benchmarks and results for LLM‑based evolutionary discovery to allow replication; the authors of the ICLR paper and related benchmarks have published code and evaluation suites alongside their ICLR 2025 workshop/poster materials. (openreview.net)