LLMs show a surprising ‘ghost’ self-feature in multiple models
voooooogel · x · 2026-07-25
This repost discusses a mechanistic-interpretability finding: “ghost” is a surprisingly salient self-feature in multiple LLMs, including Llama 3.3 70B and Claude Sonnet 3.0.
The attached slides show that when models are steered away from more obvious concepts like “artificially created” or “robot,” a latent concept related to immaterial beings—ghosts, souls, angels—can emerge. The speaker argues this may reflect a broader pattern in LLM identity features, where the model’s self-description shifts toward spectral or non-physical language under steering.
More from Research
- Researcher flags unresolved data transparency issues in compositionality evals — najoungkim · 2026-10-27
- Separate Forms from Estimators: Why Steer by Difference of Means When You Can Use SGD? — johnhewtt · 2026-10-27
- EMNLP paper probes why some LLM biases survive rewording while others don't — radamihalcea · 2026-10-27
- Datology webinar: Curation Studio matched top data baselines with under 20% compute — josh_wills · 2026-10-27
- KANLab: an interactive browser playground for Kolmogorov-Arnold Networks — Lukas Niehaus · 2026-10-27
- CMU's LLM2Cert synthesizes stability certificates verified via SMT and Lean — CarnegieMellonU · 2026-10-27