Satvik Golechha

Note that this will soon get outdated. To know more about me and my research, a much better place is my website: 7vik.io.

Independent Researcher | AGI Safety • Interpretability • Reinforcement Learning
📍 UC Berkeley (CHAI), MATS Program
📧 zsatvik@gmail.com | 🌐 7vik.io | Scholar | LinkedIn | GitHub

On a quest to understand intelligence and ensure that advanced AGI is safe and beneficial.

I’m an independent AI safety researcher currently working with:

• CHAI, UC Berkeley — Optimal exploration and long-horizon planning in RL.
• Adrià Garriga-Alonso (FAR AI) — Studying deceptive behavior in frontier AI systems at the MATS Program.
• Nandi Schoots (Oxford) — Hierarchical representations and modular training for interpretability.

Previously:

• Microsoft Research — Worked with Neeraj Kayal on representation learning theory, and Amit Sharma and Amit Deshpande on ICL robustness in LLMs.
• Wadhwani AI — Formulated AI problems in public health and trained robust and interpretable ML large-scale deployments in India.
• Mentored a SPAR 2025 project on zero-knowledge auditing for undesired behaviors.

• AI Alignment & Safety
• Interpretability & Feature Geometry
• Long-horizon RL & Planning
• Representation Learning & Theory

=Equal contribution; full list at Google Scholar

• Intricacies of Feature Geometry in Large Language Models
  ICLR 2025 (poster); Runner-up, ICLR Blog Awards
  Code | Blog

• Among Us: A Sandbox for Measuring and Detecting Agentic Deception
  Under Review
  Poster | Blog

• Auditing Language Models for Hidden Objectives
  Anthropic (external collaboration)
  Anthropic Blog | Blog

• NICE: To Optimize In-Context Examples or Not?
  ACL 2024
  Code

• Progress Measures for Grokking on Real-world Tasks
  ICML 2024 Workshop on High-dimensional Learning Dynamics
  Code

• Challenges in Mechanistically Interpreting Model Representations
  ICML 2024 Workshop on Mechanistic Interpretability
  Code

• A is for Absorption: Studying Feature Splitting and Absorption in SAEs
  Under Review

• CataractBot: An LLM-Powered Expert-in-the-Loop Chat System
  IMWUT / UbiComp 2025
  Code

• Predicting Treatment Adherence of Tuberculosis Patients at Scale
  PMLR 2022; Outstanding Paper, NeurIPS 2022
  Media Coverage

• AmongUs – Agentic deception sandbox
• nice-icl – ICL optimization tools
• grokking – Measuring grokking dynamics
• byoeb – Healthcare LLM deployment platform

• Email: zsatvik@gmail.com
• Website: 7vik.io
• LinkedIn: @7vik
• Open to collaborations in interpretability, alignment, deception audits, and theoretical ML.