UNREAL is a framework that unifies evidence selection for corpus retrieval and long-context inference through a single model-internal mechanism, adding fewer than 500,000 trainable parameters to a frozen language model. The approach lifted HotpotQA recall from 49.1% to 73.2% and produced significant accuracy gains on the NoLiMa and LV-Eval long-context benchmarks. It also reduces the computational cost of inference over extended contexts compared to separate retrieval and long-context pipelines.
