Arjun Balaji
Co-founder of ImpactAI Foundry. Researcher in AI safety and alignment at NYU Courant and Columbia University.
I co-founded ImpactAI Foundry, where we work with social impact organizations to run AI cohorts, build products and policy, and do AI safety and alignment research.
My own research asks when we can trust what a language model's reasoning tells us. I build verifiers for chain of thought, ask what guarantees they can give, and map where they break under distribution shift and optimization pressure. I'm a researcher in the Machine Learning for Good Lab at NYU Courant and in the Tian Zheng Lab at Columbia University, and a graduate research assistant at Columbia University Irving Medical Center.
Underneath both is one concern, reliable AI in places where errors are costly. At Irving Medical Center I build voice and vision pipelines for dementia research and for training dental students to read radiographs and question AI output. The same concern runs through the policy work and the cohorts at ImpactAI Foundry.
Now
- ImpactAI FoundryCo-founder. AI cohorts, products and policy with social impact organizations. Part of the Decoded Futures Network.
- NYU CourantResearcher, Machine Learning for Good Lab.
- Columbia UniversityResearcher at TZLab. Graduate research assistant at Irving Medical Center, across the Aiello Research Group and the College of Dental Medicine.
Selected research
- Certified by Abstention: Distribution-Free Guarantees for Chain-of-Thought Verifiers at Small Calibration Budgets arXiv
- What the Guard Misses, the Robot Executes: Implied Harm in VLA Instructions arXiv
- ForkSCOPE: Charting the Agentic Garden of Forking Paths arXiv
- Predicting Chain-of-Thought Correctness from Trajectory Geometry paper
- Verifying the Erdős–Gyárfás Conjecture up to 31 Vertices with SAT Modulo Symmetries paper
- Sparse Spectral Signatures of Reasoning: Model-Agnostic Verification via Sentence-Level Graph Signals paper
News
- Oct 2026Certified by Abstention preprint on arXiv.
- Oct 2026New preprint on arXiv, What the Guard Misses, the Robot Executes.
- Sep 2026ForkSCOPE preprint on arXiv.
- Sep 2026Joined the Machine Learning for Good Lab at NYU Courant as a researcher.
- Sep 2026Spoke for ImpactAI Foundry at UN Digital Cooperation Day 2026 in New York.