Closed-Loop Hypothesis Studio: Top-Down + Bottom-Up Cycles for Analyst-Led Discovery

The AI guesses smart ideas, then your data tells it if they make sense, and the loop gets better each round. Initial study: combine LLM-driven, ontology/knowledge-graph–guided hypotheses with automatic statistical tests and causal hints; test whether loops improve hypothesis validity and feasibility over one-shot suggestions.

Research Question: Does a closed-loop system that alternates between theory-guided hypothesis generation and data-driven validation yield higher-quality, more feasible hypotheses for junior analysts than LLM-only or visualization-only approaches?

Hypothesis: Alternating top-down LLM proposals (guided by domain ontologies/causal graphs) with bottom-up empirical testing and error analysis will outperform single-pass methods in validity, significance, and feasibility while maintaining time efficiency.

Experiment Plan: - Setup: System components: (1) LLM generates hypotheses grounded in domain schemas/causal knowledge graphs; (2) automatic testing (effect sizes, robustness checks, multiple comparison control); (3) error-driven refinement prompts; (4) provenance/explainability panel.

• Data/Materials: Public healthcare or retail datasets; domain ontologies or lightweight causal graphs; 24 analysts (junior vs experienced).
• Measures: Expert ratings (validity, significance, feasibility); time per hypothesis; improvement per loop; compare to LLM-only and VIADS-like visualization facilitation.
• Expected Outcomes: Closed-loop yields higher combined scores vs baselines; loop count correlates with quality gains without large time penalties.

References: 1. Ji, K., Liu, T., Sheng, F., Tan, S., Bawendi, M., & Buonassisi, T. (2025). A closed-loop AI framework for hypothesis-driven and interpretable materials design.
2. Tong, S., Mao, K., Huang, Z., Zhao, Y., & Peng, K. (2024). Automating psychological hypothesis generation with AI: when large language models meet causal graph. Humanities and Social Sciences Communications.
3. Jing, X., Cimino, J. J., Patel, V. L., Zhou, Y., Shubrook, J. H., De Lacalle, S., Draghi, B. N., Ernst, M. A., Weaver, A., Sekar, S., & Liu, C. (2024). Data-driven hypothesis generation among inexperienced clinical researchers: A comparison of secondary data analyses with visualization (VIADS) and other tools. Journal of Clinical and Translational Science.
4. Tang, Z., Wang, W., Zhou, Z., Jiao, Y., Xu, B., Niu, B., Zhou, X., Li, G., He, Y., Zhou, W., Song, Y., Tan, C., Wang, B., & He, F. (2025). LLM/Agent-as-Data-Analyst: A Survey. arXiv.org.