IC-130Amplifying visual-token attention in high layers (16-32) of released VLMs reduces sycophancy while preserving VQA accuracy, indicating that insufficient high-layer visual attention is a key cause of sycophancy
Shuo Li, Tao Ji, Xiaoran Fan, Linsheng Lu, Leyi Yang, Yuming Yang, Zhiheng Xi, Rui Zheng, Yuran Wang, xh.zhao, Tao Gui, Qi Zhang, Xuanjing Huang
The paper applies a training-free post-processing method that amplifies image-token attention logits before normalization at specific layers. For LLaVA-1.5, amplifying layers 16-32 reduces sycophancy from 94.6% to 64.4% while improving accuracy from 84.7% to 88.3%. For BLIP-2, sycophancy drops from 38.3% to 0.4%. For InstructBLIP, it drops from 68.8% to 8.3%. In contrast, amplifying low layers (1-16) or all layers (1-32) degrades accuracy significantly. Layer-wise probing confirms the sycophancy signal is concentrated in higher layers (AUC peaks at layer 31), and attention analysis shows the original model assigns less attention to visual tokens in high layers than SFT/DPO-mitigated versions.
The method was tested only on three VLMs (LLaVA-1.5, BLIP-2, InstructBLIP); the amplification factor λ is model-specific (0.9, 0.4, 1.1) and requires tuning; generalizability to other architectures is not evaluated.