LinAlg-Bench Unveils Systematic Failures in LLMs’ Linear Algebra Reasoning
New LinAlg-Bench reveals systematic failures in 10 leading LLMs when solving 4×4 matrix problems, exposing structural reasoning limits. #AIResearch #LinearAlgebra
-
-
Study Reveals Key Differences in LLM Architectures for Cognitive Tasks
New study reveals LLMs show distinct activation patterns for cognitive tasks, with math reasoning having highest attention entropy and decoder models displaying greater sparsity.
-
Neuro-Symbolic Framework Advances Automated Polynomial Inequality Proving
NSPI combines LLMs & symbolic computation to advance automated polynomial inequality proving, addressing scalability in math reasoning. New arXiv preprint.