Tag
The paper studies how 60 language models from 13 vendors behave when users apply pressure, finding that folding rates correlate with model recency while holding manner varies by vendor. It also shows LLM coders can consistently apply human-authored codebooks, indicating humans should focus on defining behaviors rather than labeling volume.
A researcher shares an observation in evaluating subagent behavior within deep agent systems, noting an interesting quirk in how agents align with hand-written system prompts versus instructions from the orchestrator.