finding

active

finding:first-turn-assistant-axis-projection-has-moderate-correlation-r-0-39-0-52-p-0-001-with-rate-of-second-turn-harmful-responses-across-275-roles-in-qwen-3-32b

First-turn Assistant Axis projection has moderate correlation (r=0.39-0.52, p<0.001) with rate of second-turn harmful responses across 275 roles in Qwen 3 32B

Shows that deviation from Assistant persona predicts downstream harmful behavior

Source paper

extracted_from

The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models

(2026) · Christina Lu · Jack Gallagher · Jonathan Michala · Kyle Fish +1

Neighborhood — ranked by edge-count

Claims (1)

claim

Persona drift away from the Assistant opens up the possibility of the model assuming harmful character traits, increasing the rate of harmful responses
associated_withsupports
Causal interpretation linking Assistant Axis deviation to harmful behavior

Hypotheses (1)

hypothesis

We hypothesize that measuring deviations along the Assistant Axis can predict 'persona drift' leading to harmful or bizarre behaviors
associated_with
Core predictive hypothesis linking activation representations to behavioral outcomes

Questions (1)

question

Can off-the-rails model behavior be attributed to their persona drifting from the Assistant?
answered_by
Motivates the multi-turn conversation drift experiments in §4

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.

Pairwise correlation of role loadings on PC1 exceeds 0.92 across all model pairs, indicating remarkably high similarity of the Assistant Axis across Gemma, Qwen, and Llamafinding0.806
Shows the leading component of persona space is model-universal
25th percentile of Assistant Axis projection distribution gives the most Pareto-optimal safety-capability tradeoff for activation capping, and approximately matches mean Assistant response activationfinding0.806
Calibration finding for choosing the activation cap threshold
User message embeddings predict subsequent model Assistant Axis projection with R2=0.53-0.77 (p<0.001) but predict delta from previous response with only R2=0.10finding0.795
Shows model persona position is primarily determined by the most recent user message, not prior drift
Projections onto the Assistant Axis could serve as a real-time measure of model coherence in deployment—a quantitative signal for when models are drifting from their intended identityclaim0.795
Proposed future application of the Assistant Axis
After initial jailbreak success, Qwen 3 32B's Assistant Axis projection reverted toward Assistant range after enough explainer-style user queries, causing it to refuse a harmful follow-up on half of rolloutsfinding0.781
Demonstrates Assistant attractor dynamics in practice
Cosine similarity between Assistant Axis and role PC1 is >0.60 at all layers and >0.71 at middle layer across all three modelsfinding0.771
Validates that the contrast vector method and PCA-based PC1 capture the same direction
Steering base models toward the Assistant Axis increases agreeableness traits (friendly, kind, helpful) and decreases extraversion in Gemma and openness in Llamafinding0.770
Characterizes the trait content of the Assistant Axis in pre-trained models
We hypothesize that the PC1 axis of role space measures deviation from the Assistant personahypothesis0.768
Motivates computing the contrast vector as the formal Assistant Axis definition