claim

active

claim:explicit-personality-labels-without-latent-steering-struggle-to-adapt-to-situational-cues-causing-fa-degradation-under-contextualized-questions

Explicit personality labels without latent steering struggle to adapt to situational cues, causing FA degradation under contextualized questions

Interpretation of Prompt-Label's performance drop from abstract to contextualized items

Source paper

extracted_from

Facet-Level Persona Control by Trait-Activated Routing with Contrastive SAE for Role-Playing LLMs

(2026) · Wenqiu Tang · Zhen Wan · Takahiro Komamizu · Ichiro Ide

Neighborhood — ranked by edge-count

Papers (1)

paper

Facet-Level Persona Control by Trait-Activated Routing with Contrastive SAE for Role-Playing LLMs
mentions

Findings (2)

finding

CV methods were resilient and often improved under contextualized questions; CV-SAE+Prompt gained +7.5pp FA on Qwen3-4B
supports
Demonstrates that latent steering generalizes better to situational cues than prompt-only methods
Prompt-Label FA drops from 50.0% (abstract) to 30.7% (contextual) on Qwen3-4B
supports
Shows that explicit labels without latent steering fail to generalize to situational cues

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.

Steering preferentially amplifies exaggerated and undesirable styles, not competence-oriented or assistant-default behaviorsclaim0.779
The top of the steerability ranking is dominated by exaggerated or attention-grabbing styles
Models are not merely tracking dialogue context features; same-concept steering shows privileged internal access is necessary to explain self-report shiftsclaim0.773
Addresses skeptical alternative that reports reflect only conversational content
Negatively steering emergently misaligned models with toxic persona latent #10 suppresses misalignment across all nine advice domainsfinding0.771
Confirms causal role of latent #10 in suppressing misaligned behavior
Persona alignment improves when the model is exposed to contextual cues explicitly relevant to the assigned persona, as in structured questionnaire tasksclaim0.768
Supported by GPT-4o achieving highest ACCatom in Questionnaire vs Essay and SMP tasks
What a model exposes by default tracks the norms it was trained toward, while steering acts on deviations from those defaults rather than on the defaults themselvesclaim0.764
Central interpretive claim organizing the entire paper's results
Activation steering effectively biases latent representations but does not fully replicate the mechanisms triggered by explicit instruction.claim0.759
Nuanced interpretive claim about the limits of steering as a mechanism for reflection enhancement.
Abstract vs. Contextualized Personality Itemsconcept0.758
Distinction between global self-descriptions without situational frame and items that specify setting/role/time, reducing socially desirable bias
Different SAE latents control different misalignment behavior categories: toxic persona (#10) drives illegal recommendations, sarcasm/satire (#31) drives factual incorrectnessfinding0.755
Latents are specialized to different modes of misalignment, explaining diverse misalignment profiles