LLMs Fundamentally Fail at Causal Discovery, New Proof Shows
A recent study published on arXiv reveals that large language models struggle to effectively conduct causal discovery, an essential aspect of scientific reasoning. The researchers illustrate that methods such as supervised fine-tuning, direct preference optimization, and in-context learning yield predictors that fail to differentiate between causal graphs producing similar observational data. This issue is formalized as a kernel obstruction theorem, indicating that the limitation is inherent to the learning approach rather than tied to any specific model or dataset. Even models that have been fine-tuned reach a plateau with simple causal graphs and their performance declines with increased complexity. To tackle this challenge, the authors introduce Agentic Causal Bayesian Optimization (A-CBO), utilizing a frozen language model as an interventional agent to direct experiments. The paper can be found on arXiv with the identifier 2605.27567.
Key facts
- Causal discovery is a cornerstone of scientific reasoning.
- LLMs cannot reliably perform causal discovery.
- Fine-tuned models plateau on simple causal graphs.
- Performance degrades as graph complexity grows.
- The failure is proven to be fundamental.
- Supervised fine-tuning, DPO, and ICL all produce indistinguishable predictors.
- The limitation is formalized as a kernel obstruction theorem.
- Agentic Causal Bayesian Optimization (A-CBO) is proposed as a solution.
Entities
Institutions
- arXiv