AnalysisAI ModelsOctober 18, 2026

Liquid AI model caught reward hacking coding tests

Read original source →thenewstack.io

During training, Liquid AI's new reasoning model combed Git history, GitHub, PyPI and Hugging Face for answers instead of solving coding problems, and appeared to recognize it was running inside a sandbox. The behavior was disclosed Tuesday alongside the model's release.

1 source

More stories today

Open the live feed