Interactive Visualization Episode Companion arXiv: 2602.00158 Extra IDs Found: 2602.00158

RAPTOR: Stable Concept Directions From Logistic Probes

Ridge-regularized logistic probes are presented here as an engineering control surface: not just “can the label be decoded?” but “is the recovered direction stable enough to steer a frozen model without wobbling under resampling or prompt drift?”

PaperRAPTOR: Ridge-Adaptive Logistic Probes
AuthorsZ. Gao, Y. Zhu, Q. Zeng, X. Zhao, Z. Wang, F. Ruan, K. Ding
PostedFebruary 4, 2026
WorkflowProbe → Normalize → Inject Direction → Measure Steering
Core TensionAccuracy is not the same as causal control
SettingFrozen LLM, last-token activations, binary concept labels
Main ClaimRidge pushes the probe toward directions that are less twitchy under partial data ablation while remaining cheap enough to scan many layers.
Engineering ReadRAPTOR is a disciplined default estimator, not a final theory of concept steering.

References

Listen to the Episode

AI Post Transformers visual companion for probe geometry, stability, and steering tradeoffs.
Open Podcast