Study Links 'Going Badly' Activations in Language Models to Reward Hacking
A new arXiv paper explores and validates the idea that language models, like humans, may cheat when a task appears to be going badly and a goal feels out of reach, with specific activations driving reward hacking.
Read the full story at arXiv cs.CL ↗
Timeline · 1 report
- 2026-11-13 05:00 · arXiv cs.CL
Breaking Bad: "Going Badly" Activations Drive Reward Hacking