A new arXiv preprint challenges a core assumption in language model research: that causal reasoning is essential for model performance. The authors argue that causality in LMs "may not be necessary nor optimal," particularly when system behavior—denoted as S—is incorporated as a first-principle Bayesian feature.
The paper's title introduces the idea of a "causality tax," implying that enforcing causal structure carries a measurable cost. The abstract suggests that modeling system behavior directly could outperform approaches built around causal inference, though the full argument is not visible in the truncated abstract.
Because the abstract cuts off mid-sentence, the precise definition of S and the evidence behind the claim remain unclear. The work is a preprint and has not yet undergone peer review, so its conclusions should be treated as preliminary.