The #ICML2026 camera-ready version of our position paper on the semantic status of the intermediate tokens produced by reasoning models, is now available https://arxiv.org/pdf/2504.09762
We updated our position on anthropomorphization of intermediate tokens in LRMs--with additional results and a call to action.. http://arxiv.org/abs/2504.09762
Users endorse researchers' call to stop anthropomorphizing intermediate tokens in reasoning models, viewing the recent "the model thought about it" framing as excessive and unwarranted.
While there are several extensions/improvements in this version, those who read previous version might be most interested in looking at sections 4, 6.1 and 6.2.
The #ICML2026 camera-ready version of our position paper on the semantic status of the intermediate tokens produced by reasoning models, is now available https://arxiv.org/pdf/2504.09762
@rao2z If LLMs are trained with human reasoning text that share a common aspect like "backtracking", intermediate tokens should be causally affected by it (yes, it is hard to quantify). I'd say humanizing traces is a useful abstraction layer between the model and data/humans behind it.
@jinulee_v They will certainly learn to imitate the wording associated with the backtracking--but there seems to be no reason to believe that they learned "backtracking process".. as our (and other's experiments cited in that paper) show..
@rao2z stop anthropomorphizing tokens is actually a solid take
the amount of "the model thought about it" framing lately is wild
No Digg Deeper questions have been answered for this story yet.
