
Shikhar J. Murty
Ph.D. Candidate
Stanford University
I'm a final year CS PhD candidate working on Deep Learning and NLP. My advisor is Prof. Chris Manning.
My research focuses on building LLMs that can generalize out-of-distribution, either through structured inductive biases, or by interacting with their environments.
These days, I'm working on robust LLM assistants that can translate user instructions to action sequences on digital environments like web browsers.
News
Feb. 2025
Talk circuit
Talking about Building the learning-from-interaction pipeline for LLMs at Together AI, MIT, Harvard, and Brown.
Dec. 2024
System-2 Generalization at Scale
We organized a workshop at NeurIPS 2024 on 'System-2 Generalization at Scale' with talks from Josh Tenenbaum, Melanie Mitchell, and others.
Nov. 2024
Session on Intelligent Agents
Invited to lead a session on 'Intelligent Agents' at Foundation Capital AI Unconference, 2024 in San Francisco
March 2024
Invited Talks
Talks in NYC (NYU / Columbia / Cornell) on 'Improving the Structure and Interpretation of Language in Modern Sequence Models'
Representative Works
Please check out my Google Scholar for all papers.
Pre-print 2025
NNetNav: Unsupervised Learning of Browser Agents Through Environment Interaction in the Wild
Shikhar Murty, Hao Zhu, Dzmitry Bahdanau, Christopher D. Manning
An unsupervised approach for training LLM web-agents, through open-ended exploration of live websites
ICML 2024
Bootstrapping Agents by Guiding Exploration with Language
Shikhar Murty, Christopher D. Manning, Peter Shaw, Mandar Joshi, Kenton Lee
A back-translation inspired method to automatically induce synthetic demonstrations for an LLM agent for browser control and multi-step tool use.
EMNLP 2023
Pushdown layers: Encoding Recursive structure in transformer language models
Shikhar Murty, Pratyusha Sharma, Jacob Andreas, Christopher D. Manning
A stack-augmented self-attention layer that can be trained in parallel, that helps transformers generalize better on tasks that require recursive reasoning.
ICLR 2023
Characterizing Intrinsic Compositionality in Transformers with Tree Projections
Shikhar Murty, Pratyusha Sharma, Jacob Andreas, Christopher D. Manning
We propose a method to functionally approximate transformers with tree-structures, and find correlation between generalization and emergent tree-structuredness.
Experience
Part-time visitor — ServiceNow Research
Advisor: Alexandre Lacoste, Dzmitry Bahdanau
Post-training for LLM browser agents
Research Intern — DeepMind
Advisor: Mandar Joshi, Kenton Lee, Pete Shaw
Unsupervised browser control with LLMs
Research Intern — Microsoft Research
Advisor: Marco Tulio Ribiero, Scott Lundberg
Fixing model bugs with language feedback
Education
Stanford University
Ph.D. in Computer Science
Advisor: Prof. Christopher D. Manning
Indian Institute of Technology, New Delhi
B.Tech in Electrical Engineering
Thesis: Inference over Knowledge Bases with Deep Learning