Agents at the instrument
I lead the development of agents for scanning tunnelling microscopy at Princeton: structured tool use, reusable skills, and human oversight for tip preparation, navigation, and measurement.
View studyCondensed-matter physicist & AI researcher
Experiments on matter.
Experiments on intelligence.
I study how language models judge, adapt, and act—and build agents that work with the physical world. My research connects atomic-scale experiments with controlled studies of AI behavior and reliability.
Currently at PrincetonExperimental physics · Empirical AI research
Connected questions
A map of my research
Response biases
Social influence
Collective behavior
Physical WorldQuantum matter,
Quantum matter,
one atom at a time.
Superconductivity, magnetic impurities, and quantum transport studied with scanning tunnelling microscopy.
LLMModels as agents.
Models as agents.
Models as subjects.
Building systems that act and investigate; studying the capabilities and decisions of the models inside them.
The bridgesFrom experiments
From experiments
to model behavior.
Agents connect models to physical tasks. Controlled studies connect their responses to questions about judgment.
Physical World
Selected papers ↓Experimental quantum physics
Superconducting bound states
Resolving tunnelling between individual Yu–Shiba–Rusinov states with scanning tunnelling microscopy.
Spins & quantum phase transitions
Probing spin-dependent transport, Kondo screening, and tunable quantum impurities.
Josephson physics & quantum interference
Measuring supercurrents and interference between transport channels at the atomic scale.
Nature Physics · 2020Tunnelling dynamics between superconducting bound states at the atomic limitPhysical Review Research · 2021Spin-dependent tunneling between individual superconducting bound statesNature Physics · 2022Superconducting quantum interference at the atomic scaleCommunications Physics · 2023Universal scaling of tunable Yu-Shiba-Rusinov states across the quantum phase transition
Physical World ↔ LLM
3 studiesAgents that work with the physical world
From operating an instrument to running simulations and conducting quantum materials research.
LLM · All connected work
9 studiesHow models learn, judge, and act
Evaluating models as researchers and as subjects: from game-playing programs to scientific work, social pressure, and collective behavior.
LLM ↔ Psychology
3 studiesStudying the judgments behind the answers
Controlled studies of response form, social pressure, and incentives reveal how model behavior changes with context.
Psychology ↔ LLM
3 studiesJudgment, response, and social influence
Psychological questions and experimental methods guide the study of language-model behavior.
Make scientific work executable
I created QMatSuite to connect agents to first-principles simulations through structured tools, reusable workflows, and reproducible project state.
View the code ↗View studyGround autonomy in calculations
Three studies ask how scientific agents accumulate experience, scrutinize published results, and carry out new computational physics research. Reproduction and numerical checks make their reasoning testable.
Read the paper ↗View studyEvaluate what the agent actually ships
Coding agents build standalone game controllers for StarCraft II, Civilization, and a roguelike. The develop–freeze–evaluate protocol separates the agent’s research process from the program it delivers.
Read the paper ↗View studySame judgment. Different answer.
Question wording, option order, and language can change the answer we score without changing the judgment a model expresses. I use controlled comparisons to separate these effects.
Read the paper ↗View studyWhat does an answer change mean?
A correction and a capitulation can count as the same answer flip. Controlled pressure tests separate these behaviors; following training checkpoints reveals when question-specific vulnerabilities develop.
Read the paper ↗View studyWhen does a group become less informed?
In social-deduction games, the communication protocol shapes a group’s ability to identify hidden adversaries. Inherited strategy notes can remove the very signals that previously made communication useful.
Read the paper ↗View studySeparate a pattern from its cause
Black-box output tests raise three different questions: whether a pattern exists, whether a detector responds to it, and whether it can be attributed to a particular watermark.
View studyWho keeps the gains from an assistant?
Personal assistants can help one user while changing the market for everyone. Controlled simulations compare fixed and adaptive sellers, tracking outcomes for both assisted and unassisted consumers.
Read the paper ↗View studyHover over a theme or bridge to expand. Explore a project to go deeper.Tap a theme or bridge to expand, then choose a project.Overview
Selected experimental physics papers 2020–2023
Nature Physics 16, 1227–1231 · 2020Tunnelling dynamics between superconducting bound states at the atomic limitPhysical Review Research 3, L032008 · 2021Spin-dependent tunneling between individual superconducting bound statesNature Physics 18, 893–898 · 2022Superconducting quantum interference at the atomic scaleCommunications Physics 6, 214 · 2023Universal scaling of tunable Yu-Shiba-Rusinov states across the quantum phase transition
Selected investigations
What are we actually measuring?
Three entry points into my work on model behavior and agent reliability.
01Behavioral measurement
1.1M replies · 8 models · 6 languages
Conceptual diagramJudgment & ResponsePapers
Same judgment. Different answer.
Question wording, option order, and language can change the answer we score without changing the judgment a model expresses. I use controlled comparisons to separate these effects.
The multilingual study compares token and sentence readouts across six languages. It shows why an opening yes or no can misrepresent a stated position. Ongoing work tests causal interventions in model representations.
My contributionDesigned the yes–no measurement study and co-developed the multilingual judgment evaluation.
02Honesty & social pressure
Training study: 741 checkpoints · 30 model lineages
Conceptual diagramSycophancy AnatomyPreprints
What does an answer change mean?
A correction and a capitulation can count as the same answer flip. Controlled pressure tests separate these behaviors; following training checkpoints reveals when question-specific vulnerabilities develop.
Across the studied routes with supervised fine-tuning, post-training checkpoints predict released-model vulnerabilities better than pretrained bases or differently trained siblings. Similar average scores can hide different weaknesses.
My contributionDesigned the controlled sycophancy evaluation; co-developed the study across model training histories.
03Evaluation
Three game environments · Frozen-program evaluation
Conceptual diagramGauntletPreprint
Evaluate what the agent actually ships
Coding agents build standalone game controllers for StarCraft II, Civilization, and a roguelike. The develop–freeze–evaluate protocol separates the agent’s research process from the program it delivers.
Large jumps across model generations coexist with sparse validation of the final program. Fresh validation cases improve delivered performance in the tested comparisons; frozen controllers play with zero LLM calls.
My contributionCo-developed the evaluation framework and studied agents’ research and validation decisions.
Across the connections
Building systems. Testing their consequences.
Agentic STMOngoing
Agents at the instrument
I lead the development of agents for scanning tunnelling microscopy at Princeton: structured tool use, reusable skills, and human oversight for tip preparation, navigation, and measurement.
Approach & contribution
Supervised operations have been demonstrated on a 300 mK STM. A separate navigation simulator supports LoRA training; transferring that trained model to the real instrument remains future work.
My contributionLead design and implementation of the instrument-control harness and navigation simulator.
Python instrument control · Physics-based simulation · LoRA
Research in progress
QMatSuiteOpen source
Make scientific work executable
I created QMatSuite to connect agents to first-principles simulations through structured tools, reusable workflows, and reproducible project state.
Approach & contribution
The project grows out of QuantumVITAS, my earlier interface for Quantum ESPRESSO. Provenance and automated tests make calculations easier to inspect, repeat, and build on.
My contributionCreator; lead architecture and implementation.
MCP tools · Provenance · Scientific workflow tests
Autonomous ResearchWorkshop papers
Ground autonomy in calculations
Three studies ask how scientific agents accumulate experience, scrutinize published results, and carry out new computational physics research. Reproduction and numerical checks make their reasoning testable.
Approach & contribution
Most substantive concerns in the scrutiny study emerged after computation. In the research pipeline, pilot reproduction provides calibration before new calculations; bounded human intervention addresses reproduction failures.
My contributionDesigned and built the memory, reproduction, and literature-to-manuscript research pipelines.
Three ICML 2026 workshop papers · Independent research
MafiaPreprint
When does a group become less informed?
In social-deduction games, the communication protocol shapes a group’s ability to identify hidden adversaries. Inherited strategy notes can remove the very signals that previously made communication useful.
Approach & contribution
The study compares communication protocols and redeploys accumulated strategies under controlled conditions. More talk or more inherited experience does not guarantee better collective judgment.
My contributionStudied communication, inherited strategies, and collective judgment with Joey Xiao.
Matched protocol comparisons · Controlled strategy redeployment
Text WatermarksOngoing
Separate a pattern from its cause
Black-box output tests raise three different questions: whether a pattern exists, whether a detector responds to it, and whether it can be attributed to a particular watermark.
Approach & contribution
This is an ongoing research direction. An observed statistical pattern alone does not establish that a product uses a watermark.
My contributionInvestigating black-box evaluation and attribution.
Ongoing work
Research in progress
AI Assistants & MarketsPreprint
Who keeps the gains from an assistant?
Personal assistants can help one user while changing the market for everyone. Controlled simulations compare fixed and adaptive sellers, tracking outcomes for both assisted and unassisted consumers.
Approach & contribution
Seller adaptation erodes some users’ savings in the simulated markets. Effects on non-users depend on seller behavior; these are controlled simulation results, not estimates of real-world consumer outcomes.
My contributionBuilt and studied controlled language-model market simulations with Joey Xiao.
Controlled simulations · Market feedback · Non-user outcomes
From the research process
Notes on asking better questions
About
An experimental way
of thinking.
I am a Postdoctoral Research Associate at Princeton University, working with Ali Yazdani on quantum materials and agents for scanning tunnelling microscopy. Alongside this work, I pursue independent research on language models and scientific agents.
Previously, I studied superconductivity and magnetism at the atomic scale at the Max Planck Institute for Solid State Research, working with Klaus Kern and Christian R. Ast. I conceived STM tips with Yu–Shiba–Rusinov bound states to probe one superconducting bound state with another.
Across these settings, I am interested in the same practical question: what experiment would distinguish the explanations we are considering?
Education & research
- 2023—present
- Princeton UniversityPostdoctoral Research Associate
- 2017—2023
- Max Planck Institute for Solid State ResearchDoctoral & postdoctoral research
- 2021
- EPFLPh.D. in Physics
- 2017
- University of StuttgartM.Sc. in Physics
- 2015
- Peking UniversityB.Sc. in Physics · Psychology double major
Feodor Lynen Research Fellowship
Humboldt Foundation · 2023–2025
Interested in model behavior,
scientific agents, or quantum experiments?