I'm a CS student and ML researcher focused on web agents, LLM internals, and task generalization. I use mechanistic interpretability tools — attribution graphs, sparse feature circuits — to understand why language models succeed or fail at novel tasks, and I build agentic systems that can navigate the web and operate computers autonomously.
- 🔬 Currently researching: predicting task generalization via attribution graph analysis
- 🤖 Building: long-horizon computer-use agents (OSWorld, WebArena)
- 🧩 Interested in: sparse autoencoders, knowledge graphs as LLM context, circuit-level interpretability
- 📍 Based in the US
| Area | What I'm Exploring |
|---|---|
| 🕸️ Web & Computer-Use Agents | Long-horizon task completion, AgentSynth trace generation, OSWorld benchmarks |
| 🧬 LLM Mechanistic Interpretability | Attribution graphs, sparse feature circuits, reasoning vs. reciting distinction |
| 📊 Task Generalization | Predicting OOD generalization from circuit structure; vignette-based evaluation |
| 🗺️ Knowledge Graphs as Context | Navigation graphs from accessibility trees injected as LLM context |
Languages
ML / Research
Web & Infra
- Circuit stability as a predictor of language model generalization (building on Wu et al. 2025)
- Navigation graph injection as structured context for computer-use agents
- BrowseComp and UltraHorizon benchmarks for long-horizon web agent evaluation
Open to research collaborations and internships. Reach out anytime.