Research
RAPID-AI
Reasoning, Analysis, and Planning for Interactive Decision-Making with Language Models
My PhD develops methods that enable language-model systems to support interactive, high-stakes decision-making in socio-technical systems (STS)—domains where outcomes depend on both technical facts and human institutions (e.g., finance, government, security, law, medicine).
I focus on three pillars:
- Measurement — constructing grounded, expert-level evaluations of STS knowledge and reasoning
- Training — using curated/synthetic data, supervised fine-tuning, and reinforcement learning to improve reasoning, analysis, and planning
- Assurance — methods for transparency, interpretability, and trust calibration so humans understand and reliably use LM recommendations
The result is a general methodology and open artifacts—datasets, models, and evaluation protocols—that transfer across domains and raise the reliability of LM-assisted decisions.
Research Thrusts
T1 — Grounded STS Evaluation
Authoritative-source corpora → cited QAs and multi-step tasks testing knowledge, multi-hop reasoning, planning, and trade-offs. Outputs carry inline quotes/citations, are teacher-graded, citation-validated, and deduplicated.
T2 — Training for Reasoning & Planning
Curated + synthetic data → SFT → RL with structure-aware rewards; OOD guards; self-agreement sampling. Focus on tractable levers that move reasoning quality—not pretraining scale.
T3 — Assurance & Interpretability
Cognitive-pattern analysis, mechanistic/behavioral probes, uncertainty and constraint checks, trust calibration. Delivering human-facing artifacts (rationales, uncertainty signals, constraint checks) that make model behavior legible and auditable.
T4 — Human-in-the-Loop Interaction
Interfaces that elicit justifications, counterfactuals, and red-teamable plans. Requiring justification before action, capturing counterfactuals, and supporting red-teaming of plans.
Publications
For citation information, see individual publication pages or my Google Scholar profile.