Can an LLM extraction system improve without fine-tuning?

Yes. DySECT improves information extraction by feeding an explicit knowledge base back into the extractor rather than updating model parameters. High-confidence examples, hierarchical concept cues, and negative signals are added to the prompt on each pass. On the DocRED benchmark, this lifted recall by 5 to 8% in the first iteration with no fine-tuning and no synthetic data.