it should be raining
Bhagyesh Kumar
@invi-bhagyesh · AI safety
HiI am a sophomore studying mathematics and computer science, interested in the science of model character, and study that by building realistic model organisms. At LAISR Lab at Cornell, I work with Prof. Lionel Levine and Jonathn Chang, I'm working on extensions of EigenBench, which measures a model's values through social consensus. I'm also building ValueArena to make these scores more legible to the general public. I've studied how the methods that instill these values can sometimes induce side effects, too.
NowI am working with Cadenza Labs on character training. I am also spending some time on a welfare-synthesis framework for aquaculture species with Myrias.
BeforeLast summer, I was an exchange student at USTC 🇨🇳, where I worked with Prof. Shiping Liu on Bakry-Émery curvature. Last spring, I was also a SPAR scholar, mentored by Prof. Lionel Levine and Jonathan Chang, where we worked on side effects of character training[ICML'26w]. I've also contributed to Redwood Research's Sabotage Bench.
MiscI have previously worked on adversarial attacks [AAAI'26], and Value Alignment [ICML'26]. Recently, I have been reading about formal verification for autoresearch. If you're bored, check a short explainer video I made about the Higgs boson when I was in high school.
You can find out what I read today at curius. Feel free to reach out and say hi at invi.bhagyesh@gmail.com
Selected research
Bakry–Emery Curvature Sharpness of Strongly Regular Graphs
FuSEP'26Future Scientist Exchange Program, University of Science and Technology of China(Poster)
Bhagyesh Kumar, Shiping Liu
Side Effects of Character Training: Quantifying Cross-Constitution Drift in LLMs
ICML'26ICML 2026 Pluralistic Alignment
Bhagyesh Kumar, Ananya Sutradhar, Saurav Panigrahi, Jonathn Chang, Lionel Levine
TopoReformer: Mitigating Adversarial Attacks Using Topological Purification in OCR Models
AAAI'26AAAI 2026 AI for Cyber Security(Oral)
Bhagyesh Kumar*, A S Aravinthakshan*, Akshat Satyanarayan*, Ishaan Gakhar, Ujjwal Verma
Ongoing research
- Ongoing · Fall 2026
Constitution Recovery
LAISR Lab · Cornell
Recovering a model’s underlying constitution from its behavior.
- Ongoing · Fall 2026
Effects of Character Training on Safety-Relevant Axes
LAISR Lab · Cornell
Studying how character training affects safety-relevant behavior.
- Ongoing · Fall 2026
Understanding Generalization in Character Training
Cadenza Labs
Studying how character training generalizes to new situations.
What I'm up to now
- Jun 2026One paper was accepted to the ICML Workshop 2026 for Pluralistic Alignment.
- May 2026I will be spending this summer at USTC
as a part of summer exchange program, working on Bakry-Émery curvature. - Apr 2026With my SPAR team, we released ValueArena, a leaderboard for value alignment.
- 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