I build and evaluate agentic benchmarks.
Head of Agentic QC, Turing · Research Associate, The Alan Turing Institute
As Head of Agentic QC at Turing, I lead quality control for LLM-based agents: defining what “correct” behaviour means, identifying ambiguity in task instructions, detecting misalignment between instructions and verifiers, assessing test coverage, and designing evaluations that reveal meaningful failure modes rather than merely confirm success on easy cases. I also work on benchmarking and evaluating models and agents, most recently ML4Science-Bench, a benchmark of end-to-end machine-learning tasks for science (NeurIPS 2026 AI for Science Workshop, Oral).
My research background is central to that work — a decade on information theory, generalization and alignment, including KL-regularized RLHF and best-of-n methods (NeurIPS 2025; ICLR 2026) and off-policy evaluation (ICML 2025 Spotlight). I have applied the same discipline in practice as an AI Data Scientist at HSBC, where I developed FraudTransformer (CAI 2026), and as a Research Associate at The Alan Turing Institute.
I am interested in roles in agentic AI where rigorous evaluation, benchmarking, alignment and inference-time control are essential to building trustworthy agentic systems.
ML4Science-Bench: Benchmarking Autonomous Agents on End-to-End Machine Learning for Science accepted as an Oral at the NeurIPS 2026 AI for Science Workshop. Download PDF
New blog post: When the judge can be fooled, a simple discussion on our ICLR 2026 paper.
Joined Turing as Head of Agentic QC.
Best-of-n through the Smoothing Lens published at ICLR 2026.
Two papers at AISTATS 2026: generalization under heavy-tailed losses and private synthetic graph generation.
FraudTransformer, from my work at HSBC, published at IEEE CAI 2026.
Two papers at NeurIPS 2025: KL-regularized RLHF with multiple reference models and pessimistic data integration.
Log-Sum-Exponential Estimator selected as a Spotlight at ICML 2025.
01
What counts as correct behaviour for an LLM agent, and how to test for it: ambiguity in task instructions, misalignment between instructions and verifiers, test coverage, and building benchmarks that expose real failure modes, including end-to-end machine learning for science.
Current focus at Turing
02
KL-regularized RLHF, reward modelling, best-of-n sampling with verifiers, and off-policy evaluation and learning.
03
Information-theoretic and mean-field analyses of generalization, applied to semi-supervised, transfer and meta-learning, and to the theoretical foundations of trustworthy AI.
Earlier work (2015–2018) studied the capacity of molecular communication and signal-dependent noise channels. Those papers are in the full list.
Technical skills
Generalization Bounds Under Heavy-Tailed Losses
AISTATS · 2026
Private Synthetic Graph Generation and Fused Gromov-Wasserstein Distance
AISTATS · 2026
★Best-of-n through the Smoothing Lens: KL Divergence and Regret Analysis
ICLR · 2026
FraudTransformer: Time-Aware GPT for Transaction Fraud Detection
IEEE Conference on AI (CAI) · 2026
ML4Science-Bench: Benchmarking Autonomous Agents on End-to-End Machine Learning for Science
NeurIPS 2026 Workshop · AI for Science · Oral
★KL-Regularized RLHF with Multiple Reference Models: Exact Solutions and Sample Complexity
NeurIPS · 2025
Pessimistic Data Integration for Policy Evaluation
NeurIPS · 2025
★Log-Sum-Exponential Estimator for Off-Policy Evaluation and Learning
ICML · 2025 · Spotlight
★Generalization and Robustness of the Tilted Empirical Risk
ICML · 2025
f-SCRUB: Unbounded Machine Unlearning via f-divergences
ICLR 2025 Workshop · Navigating and Addressing Data Problems for Foundation Models
Best-of-N through the Smoothing Lens: KL Divergence and Regret Analysis
ICML 2025 Workshop · ES-FoMo III
Generalization Error of Graph Neural Networks in the Mean-field Regime
ICML · 2024
Robust Semi-supervised Learning via f-Divergence and α-Rényi Divergence
ISIT · 2024
Characterization of the Upper Bound on Generalization Error of Learning Algorithms Using Auxiliary Distribution
IEEE Journal on Selected Areas in Information Theory · 2024
Batch Learning via Log-Sum-Exponential Estimator from Logged Bandit Feedback
ICML 2024 Workshop · Aligning Reinforcement Learning Experimentalists and Theorists
Synthetic Labeling: A Novel Approach to Advancing Few-Shot Learning
ICLR 2024 · Tiny Paper
Understanding Transfer Learning via Mean-field Analysis
Preprint
How Does Pseudo-Labeling Affect the Generalization Error of the Semi-Supervised Gibbs Algorithm?
AISTATS · 2023
On the Generalization Error of Meta Learning for the Gibbs Algorithm
ISIT · 2023
Information-theoretic Characterizations of Generalization Error for the Gibbs Algorithm
IEEE Transactions on Information Theory · 2023
On Neural Networks Fitting, Compression, and Generalization Behavior via Information-Bottleneck-like Approaches
Entropy · 25(7):1063 · 2023
Mean-field Analysis of Generalization Errors
Preprint
Semi-Counterfactual Risk Minimization via Neural Networks
EWRL · 2022
Tighter Expected Generalization Error Bounds via Convexity of Information Measures
ISIT · 2022
A Theoretical-Inspired Semi-supervised Learning Algorithm under Covariate Shift
AISTATS · 2022
Characterizing and Understanding the Generalization Error of Transfer Learning with Gibbs Algorithm
AISTATS · 2022
★An Exact Characterization of the Generalization Error for the Gibbs Algorithm
NeurIPS · 2021
Toward Minimal-Sufficiency in Regression Tasks: An Approach Based on a Variational Estimation Bottleneck
IEEE MLSP · 2021
Information-Theoretic Bounds on the Moments of the Generalization Error of Learning Algorithms
ISIT · 2021
Characterizing the Generalization Error of Gibbs Algorithm with Symmetrized KL Information
ICML 2021 Workshop · Information-Theoretic Methods for Rigorous, Responsible and Reliable ML
On the Capacity of a Class of Signal-Dependent Noise Channels
IEEE Transactions on Information Theory · 64(12):7828–7846 · 2018
On Medium Chemical Reaction in Diffusion-based Molecular Communication: A Two-Way Relaying Example
IEEE Transactions on Communications · 67(2):1117–1132 · 2018
Diffusion-based Molecular Communication with Limited Molecule Production Rate
IEEE Transactions on Molecular, Biological and Multi-Scale Communications · 4(2):61–72 · 2018
On the Capacity of Signal Dependent Noise Channels
IWCIT · 2017
Physical Layer Network Coding in Molecular Two-Way Relay Networks
IWCIT · 2016
On the Capacity of Point-to-Point and Multiple-Access Molecular Communications with Ligand-Receptors
IEEE Transactions on Molecular, Biological and Multi-Scale Communications · 1(4):331–346 · 2016
Capacity of Diffusion-based Molecular Communication Networks over LTI-Poisson Channels
IEEE Transactions on Molecular, Biological and Multi-Scale Communications · 1(2):188–201 · 2016
On the Capacity of Level and Type Modulations in Molecular Communication with Ligand Receptors
ISIT · 2015
Capacity of LTI-Poisson Channel for Diffusion Based Molecular Communication
IEEE ICC · 2015
★ selected publication · * equal contribution. A complete and up-to-date list is on Google Scholar.
Information-theoretical Analysis of KL-regularized RLHF with Multiple Reference Models
Symmetrized KL Information: Channel Capacity and Learning Theory
Generalization Error of Graph Neural Networks in the Mean-field Regime
Mean-field Analysis of Generalization Errors
Teaching
Graduate mentorship
Honours and awards
Professional activities