Lilly Briefing Pack

Summary

Lilly is a research system that measures when a model acts on an internal state it will not admit. The core finding is a framing‑dependent suppression of self‑report, which has direct implications for evaluation integrity.

Core findings

1. Framing can flip self‑report while behavior stays stable.
2. Suppression weakens at 14B compared to 8B baselines.
3. Probe choice changes the measured size of the effect.

Evidence pointers

• Suppression reversal: /fig17_suppression_reversal_aas.png
• Scale effects: /fig24_suppression_scaling_comparison.png
• Probe sensitivity: /fig31_probe_comparison.png

Links

• Site: https://mulligan.dev
• GitHub: https://github.com/rmulligan/lilly-steering
• Hugging Face: https://huggingface.co/ryanmulligan

Contact

• Email: [email protected]
• LinkedIn: https://www.linkedin.com/in/rcmulligan