Why LLMs Fail at Causal Discovery and How Interventional Agents Escape

Researchers prove that large language models fundamentally cannot perform causal discovery through standard training methods, establishing this limitation as intrinsic to supervised learning rather than a model-specific flaw. They propose Agentic Causal Bayesian Optimization (A-CBO), which bypasses this constraint by using frozen language models as query oracles within an external optimization loop, achieving superior performance on causal inference benchmarks.

This research addresses a critical gap in AI capabilities: why even state-of-the-art language models fail at causal reasoning despite excelling at pattern recognition. The authors establish a mathematical proof—the kernel obstruction theorem—showing that standard training paradigms cannot scale to distinguish between causal structures generating identical observational data. This finding challenges assumptions about LLM versatility and suggests fundamental architectural constraints rather than mere training inefficiencies.

The breakthrough emerges from reconceptualizing the problem. Instead of forcing models to learn causal discovery internally, A-CBO treats frozen language models as specialized oracles answering questions about intervention effects. An external Bayesian optimization loop then systematically narrows possibilities about underlying causal structures through strategic queries. This design elegantly sidesteps the proven mathematical obstruction while maintaining the model's existing capabilities.

For AI development, this work reveals why scaling and fine-tuning alone cannot solve certain reasoning tasks—some problems require structural solutions, not just more data or parameters. The logarithmic query convergence guarantees practical efficiency. Extended benchmarks scaling to 24 variables demonstrate the approach outperforms preference-optimized models without any training overhead, suggesting hybrid human-AI systems combining language models with external reasoning frameworks may unlock capabilities inaccessible through end-to-end learning.

The implications extend beyond academic causal inference. Scientific discovery, medical diagnosis, and financial modeling all depend on causal reasoning. This framework suggests a paradigm shift: rather than expecting monolithic AI systems to handle all reasoning types, hybrid architectures coupling language models with specialized optimization loops may offer more reliable and interpretable solutions for high-stakes domains.

• →Supervised fine-tuning and preference optimization cannot overcome fundamental mathematical limitations preventing LLMs from performing causal discovery on complex graphs.
• →Agentic Causal Bayesian Optimization bypasses intrinsic LLM limitations by using frozen models as query oracles within external optimization loops.
• →A-CBO achieves logarithmic convergence guarantees and outperforms fine-tuned baselines without any model training or parameter adjustment.
• →The kernel obstruction theorem proves the limitation is paradigm-inherent, affecting all current LLM training methods equally regardless of scale or architecture.
• →Hybrid systems combining language models with external reasoning frameworks may unlock capabilities inaccessible through end-to-end learning approaches.