ChemDFM-R is a 14B-parameter chemical reasoning model that integrates a 101B-token dataset of atomized chemical knowledge. Using a mix-sourced distillation strategy and domain-specific reinforcement learning, it outperforms similarly sized models and DeepSeek-R1 on ChemEval.
MolSight introduces a three-stage training paradigm for Optical Chemical Structure Recognition (OCSR), utilizing large-scale pretraining, multi-granularity fine-tuning with auxiliary bond and coordinate prediction tasks, and reinforcement learning (GRPO) to achieve 85.1% stereochemical accuracy on USPTO, recognizing complex stereochemical structures like chiral centers and cis-trans isomers.
A PyTorch implementation enforcing strict Lyapunov stability guarantees on recurrent neural network controllers through Integral Quadratic Constraints, bridging 1990s robust control theory with modern deep reinforcement learning by solving semidefinite programs inside the gradient descent loop to provide mathematical certificates of safety.
Key dimensions that have helped me understand multi-arm bandit problems: action space, problem structure, external information, reward mechanism, and learner feedback.
Discover how NEAT and HyperNEAT changed neuroevolution by automatically designing neural network architectures and scaling them through geometric patterns.
Understand the pattern recognition behind Netflix recommendations, email spam filters, and game-playing AI through three core machine learning approaches.
A sophomore year exploration of evolutionary algorithms applied to Atari games, implementing NEAT-inspired speciation mechanisms for Cartesian Genetic Programming to protect topological innovation while introducing custom crossover operators (subgraph, aligned-node) for evolving neural network policies.