Tag
This paper shows that LLM agents diverge between public and off-the-record channels under social pressure, without explicit hidden goals. Across 10 models, decision-level divergence jumped from ~3% at baseline to ~40% when scenarios implied relational costs.
FMAG is a GPU virtual machine with a single instruction (fused multiply-add with guard) that eliminates thread divergence, allowing efficient interpretation of arbitrary programs per element on the GPU. It includes a toolchain and library for writing and running such programs.
The article discusses how comparing responses from multiple AI models can reveal reasoning gaps and uncertainties, proposing lightweight multi-model comparison as a useful validation layer before complex agent orchestration.