Squish and Release: Exposing Hidden Hallucinations by Making Them Surface as Safety Signals

1. 1Identify the order-gap hallucination issue in language models
2. 2Implement the Squish and Release architecture to expose hidden hallucinations
3. 3Analyze the activation space of the safety circuit to detect suppressed errors
4. 4Patch the activations to make the errors surface as safety signals

ML researchers and engineers benefit from this technique as it helps identify and mitigate errors in language models, ensuring more reliable and trustworthy outputs

💡 The Squish and Release technique can help detect and mitigate errors in language models that are otherwise invisible to output inspection

🚨 Exposing hidden hallucinations in language models with Squish and Release! 🚨