Title: OCC-RAG: The Optimal Cognitive Core for Faithful Question Answering

Abstract:

The evolution of language models has largely been driven by scale, with successive iterations embedding vast amounts of global knowledge directly into their parameters. Yet, for many real-world applications, robust reasoning capabilities are more valuable than broad parametric knowledge. This reality makes task-specialized small language models (SLMs) a strategically sound option. Building on this foundation, we present the Optimal Cognitive Core (OCC), a family of SLMs designed around this principle. Among these, we introduce OCC-RAG, a variant specifically fine-tuned for faithful question answering (QA) that relies strictly on provided context. This objective is intrinsic to the OCC philosophy, demanding multi-hop reasoning over supplied texts while deliberately disregarding pre-existing memorized information.

To develop OCC-RAG, we engineered a novel pipeline for the large-scale synthesis of multi-context, multi-hop QA data. This process generated a dataset exceeding three million examples, focused on enhancing multi-hop reasoning, ensuring strict adherence to context, and enabling calibrated refusal to answer. We have released two models, OCC-RAG-0.6B and OCC-RAG-1.7B, both trained on this specialized corpus. These models generate structured reasoning paths accompanied by source citations derived from literal quotes within the context. Our results indicate that compact, task-specialized SLMs like OCC-RAG can perform on par with, or even outperform, general-purpose models that are two to six times larger. This performance advantage is evident across benchmarks for multi-hop reasoning (HotpotQA, MuSiQue, TAT-QA), faithfulness (ConFiQA), and refusal (MuSiQue-Un).

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC