A new preprint on arXiv tackles a central question in LLM reasoning: do models actually understand logical structure, or do they lean on shallow patterns? The abstract notes that LLMs perform well across diverse reasoning tasks, but it remains unclear whether this reflects genuine logical understanding or something more superficial.
The paper's title points to its proposed answer: building "logic-invariant" LLMs. Such models would keep their reasoning consistent even when the surface form of a problem changes, as long as the underlying logic stays the same. The approach appears to involve training on logical structure rather than surface features.
The abstract is cut off, so the full method and results are not available from this source. As a preprint, the work has not yet undergone peer review.