"Models can generate what Zhao and colleagues called “fluent nonsense”, plausible reasoning chains containing logical inconsistencies."
The authors explicitly use this expression when discussing reasoning chains that appear coherent but fail logical scrutiny. Their analysis documents mismatches between intermediate steps and final answers, showing why verbalized reasoning should not automatically be treated as reliable evidence of a model's underlying correctness.
View Source