Tag
This article explores the difficulty of AI alignment from a mathematical perspective, pointing out that neural networks, through characteristics such as ill-posed inverse problem inference, attribute-less numerical computation, and full-rank transformations, make it difficult to clearly specify and accurately represent human values, thereby elucidating the mathematical essence of the alignment problem.
This article presents a mathematical perspective on why modern neural networks accommodate diverse architectures and attention mechanisms, framing them as different implementations of constraints within a learnable numerical system.
MA-ProofBench is a new formal benchmark for evaluating LLMs on theorem proving in mathematical analysis, containing 200 problems across two difficulty levels. The best model, GPT-5.5, achieves only 16% on Level I and 5% on Level II, highlighting a significant gap between informal and formal reasoning.
A personal blog post rigorously introducing the Riemann integral and proving the Fundamental Theorem of Calculus, including supporting theorems like Rolle’s and the Mean Value Theorem.