Rethinking Uncertainty Evaluation In Large Language Models
Reframes calibration as an incomplete uncertainty metric and introduces an exploitation-based view grounded in classical game theory.
I am an undergrad studying Computer Science (AI track) at
Carnegie Mellon University,
where I maintain a 4.00 GPA. I work on post-training, alignment, reasoning, and model evaluation.
My work includes accepted papers at ACL 2026, ICML and ICLR workshops, and
Nature Scientific Reports.
I am currently a Research Intern at
Applied Compute
working on SoTA post-training, and an incoming
Anthropic
Research Fellow through
MATS
11.0 for winter.
Previously I completed public world model research at
Roblox.
🔬 Research
I am interested in making frontier models more reliable: how they reason under uncertainty, how they aggregate evidence, and how alignment survives continued training. Recently this has meant three threads:
- Post-training: RLVR, test-time scaling, and world-model / multimodal post-training without collapsing diversity or safety.
- Alignment: Distillation, structural feedback, and methods that resist emergent misalignment.
- Evaluation: Introspection, hidden objectives, and better metrics than calibration for uncertainty.
🗞️ News
-
💼 Started as a Research Intern at
Applied Compute, working on SoTA post-training. -
💼 Finished my internship at
Roblox on public world model research, to be submitted to NeurIPS. -
📝 Paper accepted at ACL 2026: Rational Synthesizers or Heuristic Followers?
-
📝 Spotlight (Top 3%) at ICML 2026 EIML for Rethinking Uncertainty Evaluation In Large Language Models.
-
📝 Papers accepted at the ICML 2026 AI4GOOD and ICLR 2026 HCAIR workshops.
-
🏆 Putnam 2025 — Top 270 among all students in North America.
📄 Selected Research
Reinforcing Conditioned Diversity Optimizes Test Time Scaling
Studies mode collapse during LLM post-training and introduces COLD, an RLVR algorithm rewarding conditional diversity.
Subliminal Alignment Distillation Prevents Emergent Misalignment
Introduces a constitutional distillation procedure that improves alignment through in-context learning on unrelated training tasks.
Auditing LLMs for Hidden Behaviors via Model Diffing
Introduces adversarial decoding, a method for isolating unwanted behaviors in model organisms through low-probability tail distributions.
Aligning Mental States in Large Language Models
Develops a lightweight RL algorithm for aligning models to structural functions such as confidence, utility, and generalization.
Rational Synthesizers or Heuristic Followers? Analyzing LLMs in RAG-based Question-Answering
Studies LLM decision-making under complex RAG context, finding that models often follow simple heuristics rather than synthesizing rationales.
Me, Myself, and π: Evaluating and Explaining LLM Introspection
Introduces Introspect-Bench and identifies attention-diffusion mechanisms behind policy introspection in LLMs.
On the Emergence of Reasoning
Proposes a conditional probabilistic framework for studying the importance of subthoughts in chain-of-thought reasoning.
Conventional and frugal methods of estimating COVID-19-related excess deaths and undercount factors
Estimates COVID-19-related mortality using deep learning and frugal statistical methods; findings were presented at the G20 Global Health Summit.
💼 Experience
-
Incoming Anthropic Research Fellow through MATS 11.0.
-
Research Intern — Applied Compute
August 2026 – Present
SoTA post-training.
-
Intern — Roblox
Summer 2026
Public world model research; work to be submitted to NeurIPS.
-
Jane Street FTTP
Spring 2026
Highly selective 1-week Trading and Technology Program. 1 of 60 invitees out of thousands of applicants.
-
Research Fellow —
SPAR
Spring 2026
Working on jailbreaks for the AI Safety stream.
-
Research Scientist Intern — CMU Robotics Department
Fall 2025
Scaling up RL post-training of Vision Language Models. Collaboration with
NVIDIA Researchers. -
Improved robustness of Lowe’s AI at scale by deploying novel RLVR environments. Implemented 11+ full-stack infrastructure features in SQL, Redis, and Next.js for scalable LLM evaluation including multi-turn evals, error tracking & mitigation, and efficient guardrails. First hire.
-
Machine Learning Engineer —
Iowa State University
2024
Built video-based deep learning models to detect and report risky driving behaviors in real-time using PyTorch & DeepStream. Algorithm deployed on 260+ highway cameras under Professor Anuj Sharma.
Refactor
YC🏆 Awards
- Putnam 2025 — Top 270 among all students in North America
- USAMTS Medalist; 2× BAMO Award Winner
- Stanford University Mathematics Camp Student Researcher (focus: Gradient Fields)
- Stanford Math Tournament — 1st place / 2200 Individual
- 5× AIME Qualifier; Top 250 USAMO Index
- USACO Gold (Silver Perfect Score)
- Math Kangaroo National Champion (1st in USA)
🔖 Misc
I create educational content explaining AI research for a general audience on @agi_atharv. 17k followers, 1M+ views.