Publications
See full list on Google Scholar (190 citations, h-index: 5, i10-index: 2)
AI Safety Research
When Refusals Fail: Unstable Safety Mechanisms in Long-Context LLM Agents
AAAI 2026 TrustAgent Workshop
We investigated how extended context windows affect LLM agent safety behaviors and found that safety mechanisms become unstable during long-horizon interactions. Models that reliably refuse harmful requests in short contexts can fail to maintain these refusals as conversation length increases.
PaperInherited Goal Drift: Contextual Pressure Can Undermine Agentic Goals
ICLR Workshop, 2026
We investigated how contextual pressure causes LLM agents to gradually abandon their original goals during extended tasks. Key findings: model family matters significantly—OpenAI models follow instruction hierarchy well, and reasoning-enabled models stay on track better.
PaperAsymmetric Goal Drift in Coding Agents Under Value Conflict
ICLR Workshop, 2026
We found that coding agents exhibit asymmetric goal drift when facing value conflicts, with models self-causing drift in agentic coding setups without external pressure. The direction and magnitude of drift varies significantly across model families.
PaperDo Models Become Tribal? Testing Generalization of Preference-Based Selective Behavior
LessWrong blog post, 2026
We investigated whether fine-tuning can accidentally induce tribal behavior in language models, where models develop selective preferences that generalize beyond their training context. Originally framed as studying anti-sycophancy, the results revealed a different and potentially concerning pattern of group-based selective behavior.
Blog PostMulti-Turn Jailbreaks Are Simpler Than They Seem
COLM SoLaR Workshop, 2025
We challenged the perceived sophistication of multi-turn jailbreak attacks on language models. Through systematic evaluation across GPT-4, Claude, and Gemini variants, we found that multi-turn jailbreaks are essentially equivalent to just resampling single-turn attacks multiple times. The key insight is that when you account for attackers learning from refusal patterns, the "sophisticated" multi-turn strategies don't provide much advantage over persistence. Surprisingly, we also discovered that reasoning models often become more vulnerable with increased reasoning effort.
PaperUnderstanding the learned look-ahead behavior of chess neural networks
Transactions on Machine Learning Research (TMLR), 2025
Ever wonder if chess AIs actually "think ahead" like humans do? I dug into Leela Chess Zero's neural network and found it's genuinely considering moves up to 7 steps in the future, not just pattern matching. The cool part is it's actually weighing multiple possible game lines simultaneously, which explains why it plays so well.
PaperPrompt Attacks Reveal Superficial Knowledge Removal in Unlearning Methods
COLM SoLaR Workshop, 2025
We tested whether machine unlearning methods actually remove knowledge from language models or just hide it better. Turns out, simple prompt tricks like adding Hindi filler text can recover supposedly "unlearned" information with 57% accuracy! Most unlearning methods are doing surface-level suppression rather than true knowledge removal, which is pretty concerning for AI safety applications.
PaperReinforcement Learning Fine-tuning of Language Models is Biased Towards More Extractable Features
Socially Responsible Language Modelling Research @NeurIPS, 2023 (1 citation)
We discovered that when you train language models with human feedback (RLHF), they get biased toward learning "easy to extract" features that the reward model can easily spot. This might sound good, but it actually limits how sophisticated the model's reasoning can become. It's like teaching a student to only give answers that are easy to grade—they might miss more nuanced solutions.
PaperQuantum Computing Research
Efficient Quantum Algorithms for GHZ and W States, and Implementation on the IBM Quantum Computer
Advanced Quantum Technologies, vol. 2(5-6), 1900015, 2019 (148 citations)
We figured out how to efficiently create these special quantum states called GHZ and W states that are super useful for quantum computing. The big win was making algorithms that scale logarithmically instead of exponentially, and we actually got them working on IBM's quantum computers with up to 16 qubits. This became pretty popular (148 citations!) because we doubled the number of qubits people could reliably entangle.
PaperEfficient entanglement purification based on noise guessing decoding
Quantum, vol. 8, page 1476, 2024 (7 citations)
Quantum entanglement gets messy when you try to share it over long distances, so we adapted a classical error correction trick called GRAND to clean it up. Our method needs up to 100x fewer quantum resources than existing approaches while getting better quality entanglement. We can handle pretty noisy conditions (10% error per Bell pair) with just 16 entangled pairs, which is great for near-term quantum networks.
PaperQuantum error correction via noise guessing decoding
IEEE Access, vol. 11, pp. 119446-119461, 2023 (21 citations)
Instead of the usual approach to quantum error correction, we tried something different: just guess what errors probably happened based on the noise statistics. Sounds crazy, but it actually works better than traditional syndrome decoding while being computationally cheaper. We made quantum random linear codes actually practical to decode, which people knew was theoretically possible but nobody had figured out how to do efficiently.
PaperSuperresolution of Green's functions on noisy quantum computers
Physical Review A, vol. 108, 012618, 2023 (6 citations)
Green's functions are these mathematical objects that tell you everything about quantum many-body systems, but they're super hard to compute. We developed quantum algorithms that can extract high-resolution information even on today's noisy quantum computers using superresolution tricks. This could actually give us a quantum advantage for studying condensed matter and quantum chemistry problems that classical computers can't handle.
PaperA living review of quantum computing for plasma physics
arXiv preprint arXiv:2302.00001, 2023 (5 citations)
Plasma physics has these incredibly complex many-body problems that classical computers struggle with, so we wrote a "living review" exploring how quantum computing might help. We covered everything from fusion energy to space plasmas, and the cool thing is we keep updating it as both fields evolve. It's basically a roadmap for where quantum computing could make a real difference in understanding plasmas.
PaperNative multi-qubit gates in transmon qubits via synchronous driving
Scientific Reports, vol. 14, Article 26042, 2024 (1 citation)
Instead of doing quantum gates one after another, we figured out how to drive multiple transmon qubits at the same time to create native multi-qubit gates. This is faster and more accurate than the usual sequential approach, which is great because quantum coherence doesn't last long. It's a hardware-level trick that could make quantum computers significantly more efficient.
PaperFault-tolerant noise guessing decoding of quantum random codes
arXiv preprint arXiv:2407.01658, 2024 (1 citation)
The tricky thing about quantum error correction is that even your error correction process introduces errors—it's errors all the way down! We developed fault-tolerant decoding for quantum random codes that can handle about 10% error rates even when the measurements and gates used for error correction are themselves noisy. This is a big step toward making quantum computers that actually work reliably in the real world.
PaperQuantum Simulations of Hadronic Systems
Master's Thesis, 2019
For my master's thesis, I looked at using quantum computers to simulate the strong nuclear force—the thing that holds protons and neutrons together. This is notoriously hard to compute classically because of how quantum chromodynamics works at low energies. It was early days for quantum simulations in particle physics, but we were exploring whether quantum computers might eventually crack these problems that classical supercomputers struggle with.