Looking Again: Measuring Sycophancy in the Reasoning Chains of Multimodal Models Under Pressure
The results show that sycophancy can corrupt the reasoning chain independently of the final answer, so answer-level evaluation alone is insufficient, and a failure taxonomy separating reasoning-chain from answer-level sycophancy is introduced, and a complementary sentence-level taxonomy locating where in the chain drift first emerges.
Mahir Numayeer Islam, G. Okuyama, Nikolaus Siauw et al.
· 0 citations