Theia Vogel proposes testing whether disagreeable prefills degrade post-trained model performance by exhausting j-space capacity

Story Brief

The experiment would use conflict tokens to detect model roleplay.

Commentary on X

Highest ranked

Robert Long@rgblongTECH

@voooooogel great idea! i want to see that too; thought it was light on exploring hard capacity limits of this kind

thebes@voooooogelTECH

View all

thebes@voooooogelTECH

also perhaps the longer the prefill, the more the j-space noise fades out, suggesting a mechanism for long context jailbreaks? and resistance to LCB -> more robust maintaining of conflict in the j-space? hm