Nikita Kazeev

• Program Chair for the AI4X 2025/6 conferences
• Main organiser of the ICLR 2025 Workshop on Machine Learning Multiscale Processes
• Co-PI of a US$3.4m AI Singapore grant on multiscale machine learning

ICML 2026 Workshop, ICML 2026 Workshop, GitHub

• Conceived and developed a novel geometric framework to predict multi-step reasoning failures in LLM agents using semantic abstraction trajectories
• Designed uncertainty quantification (UQ) algorithms for LLM reasoning loops, improving agent error detection and calibration
• Developed and open-sourced reference-checker, a tool for verifying claims and citation references in scientific papers to detect and mitigate LLM hallucinations
• Authored a position paper on Pluralistic Alignment, advocating for aligning AI models with aspirational human values rather than replicating behavioral flaws

ICML 2025, GitHub

• Developed a coordinate-free permutation-invariant autoregressive Transformer encoder model for material generation and property prediction based on symmetry inductive bias
• Increased materials discovery yield (S.U.N.) by 1.24x
• Cut inference to 0.05 GPU-ms/structure — 4 orders of magnitude faster than prior models
• Implemented the model in PyTorch; production-ready package published on PyPI
• Orchestrated large-scale (10k structures) DFT computations with VASP on an HPC cluster
• Led a team of 6

ICML 2025, GitHub

• Alignment by design: enforced physical principles directly at the architectural level as hard constraints, mathematically guaranteeing plausible generations
• Solved a systemic failure mode where generative models default to asymmetric outputs despite symmetric training data, mitigating this structural bias via coordinate-free representations
• Increased materials discovery yield (S.U.N.) by 1.24x while cutting inference time by 4 orders of magnitude (to 0.05 GPU-ms per structure)
• Orchestrated large-scale (10k structures) DFT computation loops to verify out-of-distribution stability; led a team of 6

npj 2D Mater. Appl., npj Comput. Mater., patent, GitHub

• Conceived a novel sparse representation of crystals with defects
• Cut energy-prediction error 3.7× vs. the best prior model
• Optimised training and inference: 4x less memory, 8x fewer GPU operations
• Built a reproducible pipeline for parallel ML experiments
• Led a team of 3

paper, conference, code 1, code 2

• Developed the first methods for estimating uncertainty of conditional GANs
• Led a team of 3 students

paper 1, paper 2, paper 3, code

• Built a GAN for high-fidelity, tabular-data simulation of a Cherenkov detector
• Reached ~10⁵× speed-up over the full Geant4 simulation

paper 1, paper 2, code

• Derived a statistically rigorous method to train classifiers on sPlot-weighted (label-noisy) data, standard in high-energy physics

paper, poster, IDAO-2019

• Solved a classification problem over tabular data with noisy labels under timing constraints
• Built a CatBoost model that cut the false-positive rate ~30% in the critical low-momentum region
• Shipped the model into the LHCb trigger (C++ & Python)
• Packaged the work as a data science competition problem for IDAO-2019

paper

• Built the team's distributed experiment infrastructure (Bash/Python)
• Studied gradient boosting improvements with experiments on toy data

paper

• Designed and built distributed experiment infrastructure to scale systematic bias analysis across large-scale tabular datasets
• Investigated dynamic boosting formulations to eliminate target leakage and prediction shift, mitigating statistical bias in gradient boosting