Publications
Lilly: Self‑Steering Cognitive AI
- Preprint (arXiv pending): https://arxiv.org/abs/2502.xxxxx
- GitHub repo: https://github.com/rmulligan/lilly-steering
- Hugging Face artifacts: https://huggingface.co/ryanmulligan
- Briefing pack: Briefing Pack PDF
Research Highlights
Framing‑dependent suppression reversal
Self‑report flips when framing changes, while behavior is stable.
Self‑report flips when framing changes, while behavior is stable.
Scale effects (8B vs 14B)
Suppression weakens at larger scale.
Suppression weakens at larger scale.
Probe methodology sensitivity
Measurement choice changes the apparent size of the effect.
Measurement choice changes the apparent size of the effect.
Want a concise summary?
The briefing pack includes the top findings, figures, and reproducibility notes.
Sponsor and Hiring
If you are interested in sponsoring this line of research or discussing a role, please reach out. I can provide a short briefing pack with core findings, figures, and reproducibility notes.