RECAST: Learning to Compute the Right Context through Adaptive Evidence Routing
Large language models are increasingly applied to tasks grounded in long, heterogeneous information sources. Conventional Retrieval-Augmented Generation (RAG) relies on fixed similarity-based retrieval, while agentic variants adapt queries and tool use but remain largely retrieval-centric. However, in many tasks, the evidence required for a solution is not explicitly present in
Reading in Plain English mode — toggle below to switch
Large language models are increasingly applied to tasks grounded in long, heterogeneous information sources. Conventional Retrieval-Augmented Generation (RAG) relies on fixed similarity-based retrieval, while agentic variants adapt queries and tool use but remain largely retrieval-centric. However, in many tasks, the evidence required for a solution is not explicitly present in
RECAST introduces a novel approach to RAG by framing evidence construction as a sequential decision process, moving beyond fixed retrieval to active derivation through computation and synthesized tools. This is crucial for ML engineers working on complex, multi-source information tasks where direct retrieval is insufficient.