QCVS: Question-Contrastive Visual Steering for Multimodal Reasoning

원문 초록

Inference-time steering in vision-language models often entangles query-relevant reasoning dynamics with low-level perceptual noise, causing collateral degradation on correct predictions. We propose Question-Contrastive Visual Steering (QCVS) to isolate task-specific intervention directions via double-differential activation probing. By contrasting spatial layout perturbations under the target query against a generic description reference, QCVS cancels task-agnostic visual artifacts without ground-truth supervision. A dual-threshold gating mechanism then evaluates relative interaction strength and signal amplitude, selectively applying a single bounded hidden-state update only when necessary. Empirical evaluations across multiple scales of Qwen3.5 (2B to 9B) and DeepSeek-VL2 on DynaMath and We-Math demonstrate that QCVS achieves state-of-the-art accuracy gains under sparse coverage (approximately 22%-30%). Crucially, QCVS yields repair rates up to 13.98% while suppressing collateral damage down to 0.41%, proving high intervention precision and computational efficiency.