Definition
Plain language
A method that lets a frozen AI answer hard cause-and-effect questions by asking it many small 'what if we intervened' questions and combining the answers.
As stated in the literature
Agentic Causal Bayesian Optimization — decomposes global causal-discovery decisions into local interventional queries posed to a frozen LLM, updating a posterior over candidate graphs to escape the kernel-predictor obstruction.
Why it matters: It lets a fixed model tackle cause-and-effect questions it would otherwise get stuck on, turning one impossible global judgment into many manageable local ones.
For example, instead of asking a frozen model to draw the whole causal map of a disease at once, it asks dozens of small questions like 'what happens to symptom B if we change factor A?' and stitches the answers together.
Heard on the show
“They call the method A-CBO — Agentic Causal Bayesian Optimization.”Episode 091 — When Better Fine-Tuning Can't Help: A Geometric Impossibility in LLM Causal Reasoning