Scholar
Yoonho Lee
Google Scholar ID: BAAZ_ysAAAAJ
PhD Student, Stanford University
Citations & Impact
Citations
3,901
H-index
13
i10-index
16
Publications
20
Co-authors
43
list available
Publications
Resume (English only)
Academic Achievements
- 2025, 'Feedback Descent: Open-Ended Text Optimization via Pairwise Comparison', submitted to ICLR 2026.
- 2025, 'RLAD: Training LLMs to Discover Abstractions for Solving Reasoning Problems', presented at ICML 2025 workshops (AI for Math, PRAL, ES-FoMo).
- 2025, 'Test-Time Alignment via Hypothesis Reweighting', presented at ICML 2025 PUT Workshop.
- 2024, 'Clarify: Improving Model Robustness with Natural Language Corrections', published at UIST 2024 and NeurIPS 2023 workshops (XAIA, ICBINB).
- 2023, 'Diversify and Disambiguate: Out-of-Distribution Robustness via Disagreement', published at ICLR 2023.
Research Experience
- Develops methods enabling models to learn from high-bandwidth structured textual feedback for complex tasks.
- Built Clarify, a natural language interface allowing humans to teach vision models using natural language corrections instead of manual labels.
- Proposed RLAD, a hierarchical reinforcement learning framework training LLMs to discover and use textual abstractions for complex reasoning.
- Designed Feedback Descent, which operationalizes open-ended text optimization via 'why better' signals from pairwise comparisons over up to a thousand iterations.
- Developed test-time alignment via hypothesis reweighting using a small set of labeled examples without retraining.
Background
- Ph.D. candidate in Computer Science at Stanford University, advised by Chelsea Finn.
- Research supported by OpenAI and KFAS.
- Focuses on operationalizing text as a substrate for learning, extracting massive information from direct experience via structured textual feedback—such as natural language corrections, pairwise comparisons with 'why better' explanations, and reasoning traces.
- Views text as a persistent store to optimize, combining parametric models (for inductive biases and in-context understanding) with nonparametric text storage (for persistence and interpretability).
- Future work aims to scale these methods to scientific discovery and other open-ended domains requiring long-horizon continual learning.
