No Usable Linear "Capitulation Direction" in Two Small LLMs: A Validation Protocol for Activation-Steering Claims, and a Cross-Family Behavioral Study of Sycophancy Under Pushback

What changed

Linear probes for sycophancy capitulation in Qwen2.5-1.5B and Llama-3.2-1B fail validation, showing overfitting.

Why you should care

Linear probes for sycophancy capitulation overfit and fail validation thresholds.

Your move

Reconsider. Steering vectors may fail if the target signal is not linearly decodable.

Evidence

Confirmedarxiv.org ↗

What to watch next

Development of robust validation protocols for activation steering that account for cross-family behavioral differences.

Event

research

Event date

Sep 17, 2026

Relevant to

General AI readers