Fei Cheng

Adding just one spatial word can lead MLLMs to consistently choose the wrong answer, revealing a critical vulnerability in their reasoning processes.

LLMs don't need "wait, let me think..." to reason—in fact, dropping the cutesy anthropomorphic markers can actually *improve* their performance.

RL models trained with verifiable rewards exhibit a surprising deductive-over-abductive reasoning asymmetry, even in controlled environments, suggesting a fundamental challenge in current RLVR approaches.