(VLMs) can change their final answers when input images undergo subtle variations like exposure shifts or JPEG compression, even when images appear nearly identical. This phenomenon, called answer flips, is especially problematic in long-horizon reasoning because small activation shifts accumulate across decoding steps. The paper introduces FlipDir, a training-free inference-time method that estimates a low-rank flip-inducing activation subspace from contrastive pairs of original and answer-flipping inputs. During decoding, a margin-based gate detects uncertain steps and selectively steers hidden states away from that subspace, recovering the original prediction while preserving stable ones.
To evaluate robustness beyond accuracy on fixed test sets, the authors also propose VisFlip, a that separately assesses recovery of original predictions and preservation of stable ones across nine dataset-variation combinations in scientific reasoning, robot-scene understanding, and medical VQA. Experiments on 18 settings show that FlipDir consistently outperforms existing methods on the combined recovery-preservation metric. For instance, they report that frontier models like GPT-5.5 have a 9.6% answer flip rate under subtle exposure shifts, while smaller models like Gemma-3-12B reach 62.4%.
This work is interesting because it addresses a real-world vulnerability in VLM deployment: routine image variations cause inconsistent outputs, threatening reliability. FlipDir requires no retraining and intervenes only when needed, making it practical to incorporate into existing pipelines. The VisFlip benchmark provides a standardised way to measure this specific type of robustness.