Rethinking STS and NLI in Large Language Models