INTERN: Internalized Natural Training via Exploration and Retained Networks for Autonomous Domain Adaptation in Language Models

Abstract

Deploying a language model in a new business environment today means writing system prompts, curating retrieval pipelines, and manually encoding domain knowledge, none of which scale. We introduce INTERN, a framework that addresses this from a different angle: instead of injecting knowledge at inference time, let the model acquire it autonomously before deployment. INTERN operates in two phases. During an exploration phase, a deployed agent reads business systems freely, treating internal documents as the organization’s primary operational source of truth, and continuously extracts domain facts for adapter-based fine-tuning. During an operating phase, the model draws on internalized weight-level knowledge rather than runtime retrieval. A configurable human-in-the-loop gate allows organizations with lower data confidence to review extracted facts before fine-tuning proceeds. We also formalize two adapter merge strategies, Deferred Merge and Incremental Merge, as a design dimension whose optimal choice depends on deployment context. INTERN makes no architectural changes to the base model and requires no labeled data. The framework is intended as a position paper; we describe the architecture in full and outline experiments for future validation.