Working Papers

Phantom Harm: Answer Extraction Brittleness Distorts the Measured Effect of Multi-Agent Debate

Dohun Lee, Hyunwoo Park

Under Review at AAAI-27 (Phase 2)

Learning Time-Varying Directed Graphs via Convex TV-Regularized Optimization and DAG Projection

Juhyeon Kim, Dohun Lee, Hyunwoo Park

Under Review at ICLR 2027

Not All Answer Revisions Are Equal: Interpreting the Representations and Controllability of Regressive and Progressive Sycophancy

Hyejin Jo, Dohun Lee, Somangchan Park, Hyunwoo Park, Hyopil Shin

Under Review at ARR 2026 August Cycle