mHTI Discussion Analysis Pipeline (M.S. Thesis)

The research engine behind my M.S. thesis and EMNLP 2026 submission: a Python pipeline that studies how the structural design of team discussions, who can see what, how much each agent remembers, team size, mentor count, and session length, shapes both collaboration behavior and final proposal quality.

Design. Two parallel tracks share one metric and inference layer: a human track that parses and analyzes ~24,000 utterances from five real research teams, and an agent track that runs matched LLM simulations under controlled conditions for within-team paired inference.

Highlights.

• Behaviorally grounded agent personas close 64% of the naive-to-human outcome gap across 5/5 teams.
• 55 NLP interaction metrics plus a multi-judge LLM ensemble (Claude / GPT / Gemini) reaching 0.96 Spearman agreement on process quality.
• Four-layer statistical triangulation (paired Wilcoxon, sign test, team-bootstrap CIs, and LME) designed for an honest small-N (N=5) design.

Paper under review at EMNLP 2026. Repository kept private pending publication.

Python · LLM agents · NLP · statistical inference