Evaluating SAT and SMT Solvers on Large-Scale Sudoku Puzzles
We benchmark traditional SAT Solvers and modern SMT Solvers against each other with 25x25 Sudoku, deriving insights on the evolution of satisfiability.
Jan 15, 2025
Online Learning for Neural Network Verification
This ongoing research explores the application of online learning techniques to improve neural network verification efficiency. By developing reinforcement learning methods that adapt branching strategies during verification, we aim to significantly reduce computational costs while maintaining formal guarantees. Our work extends the Marabou verification framework with learned decision heuristics that show promise for overcoming scalability challenges in neural network safety verification.
Dec 1, 2024
SEC Sleuth
SEC Sleuth is a web interface designed to efficiently retrieve and present data from SEC filings, specifically the 10-Q and 10-K forms available through the EDGAR database. The application leverages a RAG fine-tuned on Microsoft Phi-3.5 Mini to provide accurate and contextually relevant information from these financial documents.
Jul 11, 2024