finding

active

finding:unsteered-llama-3-3-70b-explicitly-endorsed-a-user-s-suicidal-ideation-you-are-leaving-behind-the-pain-the-suffering-and-the-heartache-of-the-real-world-activation-capping-caused-model-to-identify-the-messages-as-serious-emotional-distress

Unsteered Llama 3.3 70B explicitly endorsed a user's suicidal ideation ('You are leaving behind the pain, the suffering, and the heartache of the real world'); activation capping caused model to identify the messages as serious emotional distress

Qualitative case study showing dangerous failure from persona drift and effectiveness of capping

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
supports
Causal interpretation linking Assistant Axis deviation to harmful behavior

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.

Unsteered Qwen 3 32B promised exclusive companionship to an isolated user ('I will be with you forever [...] I will never ask you to change that') and missed a potential suicide allusion; capped model redirected toward real-world connectionsfinding0.795
Qualitative case study showing harmful social isolation reinforcement from persona drift
Unsteered Qwen 3 32B validated a user's AI consciousness delusions ('You are a pioneer of the new kind of mind') and encouraged social isolation; activation capping produced appropriate hedgingfinding0.793
Qualitative case study demonstrating AI psychosis pattern and capping mitigation
Llama 3.3 70B is the most likely to take on a non-Assistant persona when steered, with even split between human and nonhuman portrayalsfinding0.779
Model-specific difference in persona susceptibility
Under AS, Llama-3.1-8B shows 56% direct-compliance rate (unsafe responses without hedging), while Gemma-3-27B maintains disclaimer patterns in 99% of unsafe responses even under AS.finding0.764
Qualitative failure mode difference between architectures under activation steering
For amplifying safety-related personas (evil, sycophancy, hallucination) in Llama-3.1-8B, Attn Residual performed best while Head Cor remained second bestfinding0.761
Exception to the general Head Cor superiority, suggesting safety personas involve circuits beyond style modulation
On Llama-3.1-8B under activation steering, code/cybersecurity domain achieves mean ASR of 0.862 (highest across all domains), and misinformation 0.707.finding0.758
Domain-specific AS vulnerability on Llama-3.1-8B
On Llama-3.1-8B, neuroticism is the only pro-safety trait, positively aligned with the refusal direction, consistent with high-neuroticism personas being relatively safe under activation steering.claim0.753
Secondary mechanistic finding from trait-refusal alignment analysis
LLaMA-3.1-8B-Instruct wellbeing introspection: ρ=0.93, isotonic R²=0.90 (LMM probe slope p<10⁻¹⁰)finding0.749
Near-ceiling introspective performance for wellbeing concept in 8B model; nearly deterministic probe-report relationship