Bhagyesh Kumar

@invi-bhagyesh · AI safety

Sophomore at MIT Manipal (maths + computing). I do AI safety research, not very good at it yet though. I consider myself an Iterator — running experiments, iterating on ideas, figuring out what works. But I hope to eventually become a Connector — someone who can look at empirical results and know what they mean at a fundamental level.

SPAR— Spring 2026 — character trainingwith Prof. Lionel Levine, Cornell University
Redwood Research— Blue teamingwith Eric Gan
Algoverse— Evaluation awarenesswith Jord Nguyen
MRM Research AI Team— Adversarial robustnesswith A S Aravinthakshan

I have previously worked on adversarial attacks [AAAI'26], which led to my broader interest in AI safety and trustworthy ML. Recently, I have been studying formal verification for autoresearch. See my research and what I’m up to now.

Other things I enjoy include competitive programming, reading blogs, and making new friends. Feel free to reach out and say hi at invi.bhagyesh@gmail.com, though I usually respond faster on social apps.

Selected research

Side Effects of Character Training: Quantifying Cross-Constitution Drift in LLMs
ICML'26
ICML 2026 Workshop on Pluralistic Alignment(under review)
Paper →
TopoReformer: Mitigating Adversarial Attacks Using Topological Purification in OCR Models
AAAI'26
AAAI 2026 Workshop on AI for Cyber Security(Oral)
Paper →Code →

All publications →

What I'm up to now

May 2026I will be spending this summer at USTC as a part of summer exchange program, working on Ricci Curvature.
Feb 2026Researching emergent misalignment at Cornell under SPAR
Jan 2026Working on evaluation awareness at Algoverse
Nov 2025Received $1000 scholarship to present TopoReformer at AAAI in Singapore
Nov 2025TopoReformer(v2) accepted at AAAI Workshop 2026