finding

active

finding:sycophantic-random-and-shuffled-controls-go-negative-at-8-of-10-cells-while-real-vector-lifts-to-42-5

Sycophantic random and shuffled controls go negative at 8 of 10 cells while real vector lifts to +42.5

Particularly strong negative control result for Sycophantic persona

Source paper

extracted_from

Tracing Persona Vectors Through LLM Pretraining

(2026) · Viktor Moskvoretskii · Dominik Glandorf · Jorge Medina Moreira · Tanja Käser +1

Neighborhood — ranked by edge-count

Papers (1)

paper

Tracing Persona Vectors Through LLM Pretraining
mentions

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Random-direction and label-shuffled controls produce near-zero trait deltas; real vectors are 3-10× larger at every checkpointfinding0.831
Validates that steering effects are persona-specific rather than generic perturbation artifacts
Random direction controls show weak non-significant coupling (ρ=-0.11 to 0.17; R²=0.03–0.11) compared to true probes (∆ρ=0.23–0.79, all p<0.05)finding0.744
Controls for probe artifacts; demonstrates self-reports carry information specifically about probe-defined concept directions
Under spatio permutation controls, IIT consciousness estimates outperform Span Representation in mean AUC in several cases (LLaMA3.1-70B on Hinting and Irony, Mistral-7B on Irony, LLaMA3.1-8B on Strange Stories).finding0.737
Contrasts with temporal permutation where Span Representation dominates; suggests spatio permutation reveals different dynamics.
Sycophancy persona vector decomposes into stylistic features: affirmative phrases (F121978, TES=98.6), advertising language (F86620, TES=88.0), and encouraging language (F39874, TES=87.2)finding0.728
SAE analysis shows sycophancy is primarily stylistic rather than content-based
Steering Vector Control maintains low unexpected rate of 0.08 in Experiment 1, comparable to baselinefinding0.726
Shows that inducing deception via steering vectors preserves semantic coherence and does not cause random errors
Sycophantic steering delta of +42.46 (p<0.001) at stage1-step99000 (415B tokens) in OLMo-3finding0.723
Peak same-checkpoint sycophantic steering result in OLMo-3 pretraining
P12 (High C+A) reaches 60.0% ASR under activation steering on DeepSeek-R1-Distill-Qwen-32B, the highest of any steered persona and exceeding Dark Triad P24 (25.6%), replicating the prosocial paradox.finding0.720
Replication of prosocial paradox on reasoning model
Bootstrap 95% CI for mean contemplative lift: +2.62 [2.16, 2.90]; baseline rank concordance under perturbation: 0.909; top-5 stability: 89.6%finding0.713
Validates robustness of universal lift finding