Learning to Select Visual In-Context Demonstrations

A team of researchers has developed a novel method called Learning to Select Demonstrations (LSD) that fundamentally improves how Multimodal Large Language Models (MLLMs) learn from examples. The standard approach, unsupervised k-Nearest Neighbor (kNN) search, selects examples based purely on visual similarity, which often leads to redundant, unhelpful demonstrations for complex tasks like factual regression. The researchers reframed the problem as one of sequential decision-making and trained a Reinforcement Learning agent—specifically a Dueling Deep Q-Network (DQN) with a query-centric Transformer Decoder—to learn a policy for constructing optimal demonstration sets that maximize the MLLM's final performance.

Evaluated across five visual regression benchmarks, the study revealed a crucial finding: while simple kNN remains sufficient for subjective preference tasks, the learned LSD agent significantly outperforms all baselines on objective, factual regression tasks. This is because LSD intelligently balances visual relevance with output diversity, selecting a set of examples that better defines the true boundaries of the regression problem. The work, accepted to CVPR 2026, provides a clear framework for when advanced, learned selection is necessary, moving beyond one-size-fits-all strategies for visual in-context learning.