Large language models (LLMs) are often assumed to represent and reason about a wide range of phenomena, especially causal relations. The CLadder framework, released in 2023, procedurally generates benchmark datasets and has become a common tool for evaluating formal causal reasoning in LLMs. Experiments report high accuracy of LLMs on the large number of examples produced by CLadder.
This paper examines what can be reliably inferred when LLMs perform well on CLadder test sets. We identify several superficial and highly predictive patterns embedded by the generator, such as fixed lexical orders and structural cues, which do not require genuine causal reasoning. Even simple models like random forests achieve strong performance after training on a modest sample of generated data.
Additional tests reveal that many commercial LLMs are extremely brittle on CLadder: removing text sections that are critical for formal causal reasoning has almost no impact on their scores. This suggests that the models are not truly understanding causality but are exploiting surface patterns learned from the generator’s training data.
Taken together, the most plausible hypothesis for the strong performance of state‑of‑the‑art LLMs on CLadder is that they are leveraging these superficial patterns. Any claim that LLMs exhibit deeper causal reasoning must first rule out this hypothesis.
Review