BadEdit: Backdooring Large Language Models By Model Editing

Singapore Agency for Science, Technology and Research; primary engine for foundational and applied AI research

National University of Singapore; among Asia's top AI research universities. Launched the NUS AI Institute in March 2024, bringing together foundational AI, applied AI and societal impact research

Nanyang Technological University; major hub for AI and data science research

Singapore Management University; AI applications in business and society

Singapore University of Technology and Design; innovation at the intersection of AI and design

This collaborative project, led by Professor Jie Zhang (NTU) in partnership with MIT researchers including Associate Professor Jacob Andreas, focuses on developing trustworthy and interpretable human-AI collaborative systems for combinatorial optimization. The research establishes natural language as a robust interface for human-AI interaction, enabling systems to formulate complex optimization problems from imprecise input, perform robust AI-assisted solving, and provide explainable feedback. The outcomes aim to serve industries like logistics, manufacturing, and finance with intelligent decision-making tools, while advancing the theoretical foundations of trustworthy AI collaboration.

PINNACLE is an adaptive point selection method that improves Physics Informed Neural Networks (PINNs) training efficiency by automatically optimizing the selection of all training point types using empirical Neural Tangent Kernel theory. The method outperforms existing benchmarks across multiple problem types including forward problems, inverse problems, and transfer learning applications.

This research introduces Symbolic Regression Enhanced Decision Tree (SREDT), which leverages symbolic regression to discover non-linear and multivariate splitting rules for decision trees. Compared to conventional decision trees, SREDT demonstrates superior prediction performance, more compact tree structures, faster inference time, and robustness to noise. The method uses genetic programming to search for closed-form analytical expressions, significantly outperforming standard decision trees on classification tasks.