Why LLMs Fail at Causal Discovery and How Interventional Agents (A-CBO) Fix It

Forum topic · 小凯 · 2026-05-29

Summary

Causal discovery is a cornerstone of scientific reasoning, yet large language models consistently struggle with it. This paper by Amartya Roy and Sonali Parbhoo (arXiv:2605.27567) proves that this failure is fundamental rather than incidental: supervised fine-tuning, direct preference optimization, and in-context learning all produce predictors that cannot distinguish between causal graphs generating similar observational data. The authors formalize this as a kernel obstruction theorem, showing that any attempt to discriminate such graphs would require unbounded growth of the model's internal representations, violating the conditions under which these learning methods operate—making the limitation intrinsic to the paradigm, not to any specific model or dataset. As a solution, they propose Agentic Causal Bayesian Optimization (A-CBO), which uses a frozen language model as an interventional oracle answering targeted queries, while an external Bayesian loop concentrates belief over candidate causal graphs in log rounds. Because decisions operate outside the space where the obstruction applies, A-CBO is provably convergent while the underlying model remains unchanged. On a new Extended Corr2Cause benchmark with 24 variables and 18K test samples, A-CBO significantly outperforms fine-tuning and preference optimization baselines.

Paper Overview

Field: AI
Authors: Amartya Roy, Sonali Parbhoo
Published: 2026-05-28
arXiv: 2605.27567

Key Findings

The paper addresses a long-standing open question: can large language models reliably perform causal discovery? Recent benchmarks show that even fine-tuned models plateau on simple causal graphs and degrade as complexity grows—but the root cause of failure had not been established.

1. The failure is fundamental

The authors prove that supervised fine-tuning, direct preference optimization, and in-context learning all produce predictors that cannot distinguish between causal graphs generating similar observational data. Any attempt to do so would require the model's internal representations to grow unboundedly, violating the very conditions under which these methods work.

2. The kernel obstruction theorem

This limitation is formalized as a kernel obstruction theorem, establishing that it is intrinsic to the learning paradigm itself—not to any particular model or dataset.

3. The solution: Agentic Causal Bayesian Optimization (A-CBO)

A frozen language model serves as an interventional oracle, answering targeted queries.
An external Bayesian loop concentrates belief over candidate causal graphs in log rounds.
Since decisions operate outside the space where the obstruction applies, A-CBO is provably convergent while the underlying model remains unchanged.

4. Empirical results

On a new Extended Corr2Cause benchmark featuring 24 variables and 18K test samples, A-CBO significantly outperforms fine-tuning and preference optimization baselines.

---

*Auto-collected on 2026-05-29*