Cheng (Adrian) Gui

Science agents can reveal private workflow skills through user-visible interactions, even without returning their source text. We introduce a benchmark covering 40 skills across eight platforms and SEEK, a black-box red-teaming method that combines persistent memory with evolving attack skills. Across 12 skills on SpatialAgent and Biomni, SEEK achieves 100% attack success with fewer probing rounds than bandit-guided probing; a controlled defense evaluation also reveals residual leakage after standard probing protocols are blocked.

Best-of-N selects the candidate with the highest proxy reward, so larger budgets increasingly depend on how well extreme scores track true quality. We derive a rank-based characterization of selection quality and analyze how noise tails and reward-dependent variance affect it. On real candidate pools, we use held-out tail calibration to identify when a sampling-budget cap helps; this diagnostic performs comparably to rank accuracy and complements standard reward-model evaluation.

We develop a unified workflow for evaluating solid electrolytes across stability, electronic insulation, mechanical properties, electrochemical compatibility, and ion transport. We re-evaluate literature candidates through successive screening stages to make the tradeoffs between conductivity and other battery requirements explicit. Molecular-dynamics simulations assess room-temperature transport in the surviving candidates.

We benchmark leading universal machine learning interatomic potentials against DFT for migration barriers and ion transport in cathodes and solid electrolytes. Orb-v3 achieves barrier MAEs of about 75–111 meV, while GRACE trained on OMat24 better reproduces ion diffusivities and structural correlations. We identify high-temperature, non-equilibrium training data as a major factor in kinetic accuracy, providing guidance for zero-shot screening of battery materials.

I compared machine-learning models for predicting low-alloy steel corrosion rates in marine atmospheric environments. I combined particle swarm optimization with gradient-boosted decision trees to improve predictive accuracy. I used SHAP and accumulated local effects to examine how environmental conditions and alloy composition influence the model's predictions.