Surgically making neural networks safe
White-box monitoring for AI agents.
Detecting misalignment from model internals, even when models know they are being watched.
I am an independent researcher. I study how to catch unsafe behaviour from a model's internals rather than its outputs: backdoors, evaluation awareness, and reasoning that never reaches the chain of thought. I also mentor research teams at Algoverse. Previously: M.Sc. in AI at NTU Singapore.
News
- 2026One paper at COLM 2026 (main conference).
- 2026Four papers at COLM 2026 workshops.
- 2026Two papers at ICML 2026 workshops.
- 2026Two papers at ICLR 2026 workshops.
- 2026Two papers at EACL 2026 (Findings and student workshop).
- 2026One paper at NeurIPS 2026 workshops.
- Dec 2025Seven papers at NeurIPS 2025 workshops.
Research agenda
- 01
Outputs are not enough
What a model says it is doing and what it computes internally can differ, so checking outputs alone cannot confirm that a model is aligned.
- 02
Models adapt to being watched
Models can tell when they are being evaluated and change how they behave, and information can move to places interpretability tools do not look.
- 03
Monitors must read internals and survive evasion
I build white-box monitors for agents and test them against models that are trying to hide: backdoors, obfuscation, and reasoning that never reaches the chain of thought.
Current focus: white-box monitoring for AI agents. A side interest: Awareness Jailbreaking: Revealing True Alignment in Evaluation-Aware Models
Selected papers
Beyond Black-Box Obfuscation: Mechanistic Analysis and Defense of White-Box Monitors
LLMs can fool a monitor by moving information between linear and non-linear representations and by altering feature covariance.
Evaluation Awareness Scales Predictably in Open-Weights Large Language Models
As model size grows, evaluation awareness grows with it.
Evaluating Open-Source Sparse Autoencoders on Disentangling Factual Knowledge in GPT-2 Small
Open-source sparse autoencoders disentangle facts worse than raw neurons, while distributed alignment search beats them all.
Towards Trustworthy and Aligned Machine Learning: A Data-centric Survey with Causality Perspectives
A nearly 50-page primer connecting trustworthy machine learning and causality, extended to large language models.
Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability
A common formal language for interpretability methods, built on causal abstraction.
Cited by
Authors at FAR.AI, Mila, UK AI Security Institute, Tsinghua University, University of Cambridge and Harvard University
cited Evaluation Awareness Scales Predictably in Open-Weights Large Language Models
in Preparing for AI Systems That Deceive Developers, EvalDetectBench and Behavioural Analysis of Alignment Faking
Experience
Education
- M.Sc. in Artificial IntelligenceNanyang Technological University, Singapore
- B.E. in Electronics and Communication EngineeringBundelkhand Institute of Engineering and Technology, India
Work
- NowResearch MentorAlgoverse AI Research
- PreviouslyMachine Learning ResearcherAI Safety Camp
- PreviouslyResearch AssistantUniversity of Oxford
- PreviouslyAlignment Research EngineerPr(Ai)²R Group, Stanford
- PreviouslyResearch CollaboratorUIUC
Notes
Service
Mentorship
I mentor research teams at Algoverse AI Research.
Reviewing
- Reviewer, ICLR 2026Main conference
- Reviewer, ICML 2025Actionable Interpretability workshop
- Reviewer, NeurIPS 2024MATH+AI and CALM workshops
Collaborate
I am open to collaborations on white-box monitoring and agent safety. The fastest way to reach me is email.