Tag: #language-models

ArXiv paper proposes a measurable internal “rift” that may reveal lies in language models

An arXiv preprint finds a measurable internal 'residual rank' difference when models lie versus when they err, with implications for AI safety.

#ai, #ai-safety, #language-models, #deception

Preprint Claims Language Models Undergo 'Alignment Transition' Around 3.5 Billion Parameters

An unreviewed arXiv preprint argues language models shift from trading truth for reasoning to improving both past ~3.5B parameters; code and dashboard released.

#ai, #alignment, #language-models, #arxiv