A core difficulty in multimodal in‑context learning (MICL) is that many tasks are ill‑defined without context and require uncovering hidden rules from demonstrations. Most prior approaches treat MICL as a plain input‑output mapping, neglecting the need to explicitly induce and reason over such rules, which limits both accuracy and interpretability. We introduce ERIM (Explicit Rule‑Induction for Multimodal In‑Context Learning), a unified framework that improves MICL by explicitly inducing the latent rules governing the demonstration samples. ERIM leverages the reasoning ability of multimodal large language models (MLLMs) to build rule‑centric MICL datasets and provides explicit supervision for rule induction. Training follows a two‑stage strategy: first a mixed cold‑start supervised fine‑tuning phase to teach basic rule induction, then a reinforcement‑learning phase to further refine the model’s ability to discover and apply rules during inference. Extensive experiments across diverse MICL settings demonstrate consistent performance gains, and correctly inducing the underlying rules leads to stronger and more stable results.
Review