Large language models (LLMs) make it possible to express symbolic structure directly in natural language, but most context engineering methods still assume that a human or an external pipeline already knows what contextual information the model will need. In this paper, we introduce fundamental learning , a two-stage context engineering framework in which the model first generates supportive lower-level analyses and then reuses them as auxiliary context to improve performance on a higher-level target task. We study this idea in computational linguistics, where intermediate analyses at the syntactic and semantic levels provide explicit context for downstream pragmatic interpretation. Using a unified instruction schema, we fine-tune Qwen3-8b with syntactic and semantic tasks and evaluate the approach on pragmatic tasks. Across experimental settings, requiring the model to explicitly infer these low-level intermediate labels before high-level task prediction consistently outperforms both a baseline model and a model trained on the same low-level tasks without being required to surface them at inference time. The results suggest that when context engineering cannot rely on ingenious human-designed prompts, retrieval pipelines, or hand-crafted reasoning procedures, fundamental learning offers a general alternative.
Mao et al. (Wed,) studied this question.