LLM-IDEA: Identifiability-Driven Experimental Agent for Autonomous Discovery of Mechanistic World Models

What this work does

LLM-IDEA couples a language-model proposer to deterministic identifiability diagnostics. It also audits whether discovery benchmarks reward distinctions that their observation interfaces can actually reveal.