In multimodal large language models (MLLMs), in‑context learning (ICL) has become a pivotal technique for boosting performance across diverse tasks. Existing multimodal ICL approaches largely rely on superficial imitation of demonstrations, making it hard for models to align their outputs with the reasoning path demanded by the input, a problem that worsens on complex tasks. To overcome this, we propose an ICL framework that merges contrastive demonstration modeling with the self‑refinement ability of MLLMs. Each demonstration is reformulated as a contrast between a suboptimal response and a better one under the same input, accompanied by a reasoning path that explains how to refine the former into the latter. This contrastive formulation renders the desired reasoning process explicit, steering the model beyond shallow copying. Because effective refinement depends on the current response, we introduce a response‑conditioned retrieval mechanism that selects demonstrations whose reasoning paths are most relevant to the ongoing answer. A lightweight alignment controller predicts response quality and decides whether further refinement is needed. Experiments on three multimodal task categories, especially visual question answering (VQA), demonstrate consistent performance gains, with notable improvements on VQA.
Review