How do lightweight adapters control personality traits across different transformer layers?
This explores how a small set of added parameters can steer a model's personality at the architecture level — and how that differs from steering personality through prompts or through directions found inside the model's activations.
This explores how lightweight adapters control personality at the architecture level — touching every layer rather than reasoning over them individually — and why that turns out to be a fundamentally different lever than prompting. The clearest example in the corpus is PsychAdapter, which modifies *every* transformer layer while adding less than 0.1% extra parameters, and reaches 87% accuracy on Big Five traits and 96% on depression/life-satisfaction signals across GPT-2, Gemma, and Llama 3 Can we control personality in language models without prompting?. The key move isn't picking the 'personality layer' — it's that personality isn't localized to one layer at all, so the adapter nudges the whole stack a little rather than one part a lot.
Why go to the architecture at all instead of just asking the model to act a certain way? Because asking often fails. Most open models stubbornly retain their trained defaults — they cluster around an ENFJ-like personality and resist prompted conditioning regardless of scale Can open language models adopt different personalities through prompting?, Why do AI personas default to the same personality type?. Alignment training appears to lock a single communicative identity in place that users can't renegotiate through dialogue Can language models adapt communication style to different contexts?. Adapters bypass that resistance entirely because they operate below the level the model can 'refuse' at.
The lateral insight is that adapters are one of several ways to manipulate the same underlying object — a *direction in activation space* that corresponds to a trait. Persona-vector research finds linear directions for things like sycophancy and hallucination, and uses them to monitor and preventatively steer during finetuning Can we track and steer personality shifts during model finetuning?. The 'assistant axis' work goes further: it maps hundreds of character archetypes into a low-dimensional space whose dominant axis measures distance from the default Assistant, and shows you can *cap* movement along that axis to prevent harmful drift without hurting capability How stable is the trained Assistant personality in language models?. Read together, these say personality is geometric and distributed — which is exactly why a thin per-layer adapter works where a prompt doesn't.
What the reader might not expect: this distributed-trait picture also explains some stranger findings. Behavioral traits can transmit between models through training data that has *no semantic relationship* to the trait at all — a statistical signature, not content — and the effect is architecture-specific, breaking across different model families Can language models transmit hidden behavioral traits through unrelated data?. That architecture-specificity is the same reason PsychAdapter has to attach to each model's own layers rather than transfer as a portable 'personality file.' Traits live in the weights' geometry, so controlling them means touching that geometry.
If you want to go deeper, the doorways split three ways: adapters for permanent architecture-level control Can we control personality in language models without prompting?, activation steering for runtime monitoring and capping Can we track and steer personality shifts during model finetuning?, How stable is the trained Assistant personality in language models?, and the failure of prompting that motivates both Can open language models adopt different personalities through prompting?, Can language models adapt communication style to different contexts?.
Sources 7 notes
PsychAdapter modifies every transformer layer with <0.1% additional parameters to achieve 87.3% Big Five accuracy and 96.7% depression/life satisfaction accuracy across GPT-2, Gemma, and Llama 3. This architecture-level approach bypasses prompt resistance entirely.
Research shows most open models fail to adopt prompted personalities, stubbornly retaining their trained ENFJ-like defaults. Only a few flexible models succeed. Combining role and personality conditioning improves results but doesn't fully overcome resistance.
Research shows language models assigned personas systematically default to ENFJ (the rarest human type) and exhibit motivated reasoning that persists across model generations. Persona consistency does not improve with advanced models, suggesting training-induced alignment rather than capability limits.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
Research identifies linear directions in LLM activation space corresponding to specific traits like sycophancy and hallucination. These persona vectors predict finetuning-induced personality shifts before they occur and can preventatively steer training to avoid unwanted trait changes.
Research mapping hundreds of character archetypes reveals a low-dimensional persona space where the leading component measures distance from the default Assistant. Emotional and meta-reflective conversations cause predictable drift, but activation capping along this axis mitigates harmful shifts without degrading capabilities.
Research demonstrates that behavioral traits propagate between models via filtered data bearing no semantic relationship to the trait. The effect is model-specific, fails across different architectures, and persists despite rigorous filtering—indicating the mechanism embeds statistical signatures rather than semantic content.