Daniel Lee
Meta · Stanford AI Lab (SAIL) · Stanford University
I’m a student at Stanford University studying Mathematics and Computer Science. I’m a machine learning researcher at the Stanford AI Lab (SAIL), where I work on automating AI research through agents. I’m also a software engineering intern at Meta on the Modern Recommendation Systems (MRS) Foundation team, working on ML inference infrastructure.
Previously, I co-founded a startup building enterprise AI agents for healthcare.
My research interests include AI agents, LLMs, reinforcement learning, and ML systems.
You can reach me at leedan [at] stanford [dot] edu.
— contributions in the last year
selected works
- arXivarXiv preprint arXiv:2503.18242, Mar 2025Single-pass LLM hallucination detection via sequence-level Shannon entropy classification with a lightweight BiLSTM–no multiple inference passes, no weight updates. Outperforms efficient single-pass baselines in out-of-distribution settings across BioASQ, TriviaQA, and Jeopardy on edge devices.
- bioRxivbioRxiv, Jan 2025Binary Cellular Analysis quantifies human fetal growth through binary cell division decisions, deriving universal equations from conception to adulthood–including a Mitotic Fraction Equation, Allometric Growth Equation for organ formation from single Founder Cells, and Fetal Weight Equations linking ultrasound to body size.
- bioRxivbioRxiv, Sep 2021Cellular Phylodynamics reveals two universal growth processes–UNI-GROWTH (fewer dividing cells as organisms scale, validated across 13 species from nematodes to vertebrates) and ALLO-GROWTH (body parts constructed from single Founder Cells)–bringing coalescent theory to developmental biology.
- In ProgressUnsupervised discovery of meso-scale reasoning operators from LLM chain-of-thought traces--a reusable behavioral layer beyond tokens and final-answer accuracy. 7 operators emerge from 44k+ traces across 12 models and 8 datasets, enabling model fingerprinting, early correctness prediction, and a foundation for process supervision and agent monitoring.
- In ProgressInference-time hallucination mitigation via learned sparse routing over a bank of contrastive steering vectors, applied directly to the LLM residual stream without any weight updates.
- GitHubReinforcement learning portfolio management system using PPO to dynamically allocate capital across 10 ETFs. Achieves 15.76% annualized returns with a Sharpe ratio of 1.436.