research-refine

To turn a vague research idea into a concrete method plan, you need problem anchor preservation and iterative review. This process refines a fuzzy technical route into a paper-ready method covering architecture, training recipe, and minimal validation.

Yes, you can draft a research proposal using an LLM review loop. The reviewer scores fidelity, specificity, contribution quality, and feasibility, then revises the technical plan until a high bar is met or refinement rounds end.

Problem anchor preservation freezes an immutable problem statement during research planning. It ensures the original identified problem is strictly preserved across iterative refinement rounds while the technical route and method decomposition are optimized.

Design minimal validation by decomposing your method into implementation-ready specifications. The planning process includes defining the inference path, identifying failure modes, and outlining validation steps to ensure feasibility before full execution.

Yes, the method refinement process supports checkpoint recovery. It persists state to refine-logs/REFINE_STATE.json and writes round-by-round artifacts, enabling resumable refinement if the iterative review loop is interrupted.

You should not use an iterative LLM review loop when your technical route is already concrete or the core problem is not yet identified. This refinement approach is specifically designed for scenarios where the direction is vague but the problem anchor exists.