Akshatha Arunkumar

Harvard Student Research Institute

Oncology-focused benchmark for auditing whether biomedical knowledge-graph explanations are biologically plausible and faithful to the model's decision, reporting separate audit dimensions rather than a composite score.

COLM 2026·AI Measurement Science: Toward Rigorous AI Evaluation

Proposes silent-failure rate as the correct metric for prompt-injection defenses, replacing Detection Success Rate. Introduces ASP-Guard, a per-slot typed runtime monitor that blocks or escalates every landed attack with zero silent failures across VPI-Bench and AgentDojo, while holding benign false-positive rate to 0.08.

Co-authored research combining LLMs with graph attention networks to verify if online claims can be supported by reliable evidence.

Mathematical expository paper based on proof-based coursework at Euler Circle under Dr. Simon Rubinstein-Salzedo.

Harvard Undergraduate OpenBio Laboratory

Selected as 1 of 120 researchers from 2,500+ international applicants. Developed MechanismCheck, a leakage-safe benchmark for auditing whether biomedical knowledge-graph explanations are biologically plausible and faithful to model decisions. Mentored by Salomon Kabongo.

Admitted to Canada/USA Mathcamp (4% acceptance rate) — the most selective international summer math program for pre-college students. Took 25+ advanced classes across pure mathematics.