Skip to content

LLM Personality Engineering: Methods, Boundaries, and Tradeoffs

Last Updated: 2026-03-24

Overview

How do you make an LLM embody a specific persona? This document surveys three paradigms — prompt engineering, fine-tuning, and activation engineering — and maps the boundary between what each can achieve.

The key finding: personality is not a single problem, but a spectrum of increasingly deep interventions, each with distinct tradeoffs in cost, controllability, and persistence.

Paradigm Map

                    Cost / Effort
                         ▲
                         │
        Fine-tuning      │  ● BIG5-CHAT (SFT/DPO)
        (weight change)  │  ● OpenCharacter
                         │  ● FinePE (MoE-LoRA)
                         │  ● Neuro-sama (iterative SFT)
                         │
                         │
  Activation Engineering │  ● PERSONA (vector algebra)
  (inference-time,       │  ● SAS Personality Sliders
   no weight change)     │  ● Anthropic Persona Vectors
                         │
                         │
        Prompting        │  ● SillyTavern character cards
        (no change)      │  ● System prompt + few-shot
                         │  ● Eliza character files
                         │
                         └──────────────────────────────► Personality Depth / Robustness

Paradigm 1: Prompt Engineering

The most accessible approach. Define personality through system prompts, character descriptions, and example dialogues.

Techniques (from SillyTavern Community)

The SillyTavern roleplay community has developed the most sophisticated prompt-based personality engineering. Key formats:

Format Description
PLists Comma-separated trait lists: [brave, sarcastic, loyal]
Ali:Chat Example dialogues with {{char}}: and {{user}}: prefixes, separated by <START> tags
W++ Structured attribute blocks (deprecated, but historically influential)
Plain prose Natural language character description + backstory

Best practices (empirical, from community testing): - The first message is the single most important element — "the model is more likely to pick up style and length constraints from the first message than anything else" - MBTI profile as add-on to character card "significantly improves personality consistency" - Backstory + example conversations more effective than trait lists alone - Token budget matters: "a 1000-token character definition cuts the AI's memory in half" on small context models

What Prompting Can Do

  • Set broad personality traits (friendly, sarcastic, formal)
  • Establish speech patterns and vocabulary
  • Define character knowledge and backstory
  • Control response length and format

What Prompting Cannot Do

  1. Override alignment training. kimjammer's finding: aligned base models "stubbornly avoid swearing" even when the system prompt explicitly demands it. The safety layer built into weights overrides prompt-level instructions.

  2. Maintain consistency over long contexts. The SAS personality slider paper notes: "prompt engineering remains fragile: models frequently exhibit contextual drift within long context windows." As conversation grows, early personality instructions get diluted.

  3. Achieve statistical validity. The paper "When 'A Helpful Assistant' Is Not Really Helpful" (arxiv 2311.10054) tested 162 personas across 2,410 questions on 9 models and found: "most of the personas have no or negative impact on LLM's performance." Persona effects are "largely unpredictable" — even optimized selection barely outperforms random choice.

  4. Produce human-like trait distributions. BIG5-CHAT showed that prompting produces personality scores that don't match how humans actually express traits. SFT/DPO-trained models produce "intra-trait correlations more closely matching human data."

Eliza Framework (a16z)

Worth noting as a production-grade prompt-based system. Eliza uses JSON character files that define personality, knowledge, and behavior. Architecture separates Runtime, Character, Client, Adapter, and Plugin layers. Character files are purely prompt-based — no fine-tuning, no activation engineering. The sophistication is in the agent framework, not the personality method.

Paradigm 2: Fine-Tuning

Modify model weights to embed personality. Three sub-approaches:

2a. Supervised Fine-Tuning (SFT)

Train on personality-annotated dialogue data.

BIG5-CHAT (ACL 2025): - 100K dialogues grounded in Big Five personality model - Dataset constructed from SODA (social scenarios) + PsychGenerator (846K Facebook posts with Big Five annotations) - SFT and DPO both outperform prompting on BFI and IPIP-NEO personality assessments - Surprising finding: personality affects reasoning. Models with higher conscientiousness + agreeableness and lower extraversion + neuroticism perform better on reasoning tasks — matching psychological research on humans - Uses LoRA for efficient training

OpenCharacter (arxiv 2501.15427): - 20K synthetic characters + 306K role-playing instruction-response pairs - Two strategies: response rewriting (rewrite existing response in character) and response generation (generate new response in character) - LLaMA-3 8B fine-tuned achieves GPT-4o-level role-playing - Key insight: character info goes into system prompt during fine-tuning, so the model learns to generalize from character descriptions

2b. Mixture of LoRA Experts (Per-Trait Modules)

FinePE (ScienceDirect 2026): - Assigns separate LoRA modules to each Big Five subtrait (60 subtraits total) - Gating mechanism learns optimal combination weights - 120K Q&A pairs across subtrait subsets - Achieves average induced scores of 4.91 (high) / 1.61 (low), outperforming second-best by 29% - Addresses the "granularity gap": standard SFT averages across traits, FinePE achieves fine-grained disentangled control - Key advantage over prompting: prompting suffers from "conflicting trait expression" when combining multiple traits

Fusian (arxiv 2603.15405): - Multi-LoRA fusion for fine-grained continuous MBTI personality control - Similar per-trait LoRA approach applied to MBTI framework instead of Big Five

2c. Iterative Batch SFT (Neuro-sama)

See neuro-sama.research.md for full analysis.

  • 2B parameter model, q2_k quantization
  • Training data from curated stream transcripts
  • Human-in-the-loop curation (Vedal manually selects data)
  • Deploy → collect → curate → retrain → deploy cycle
  • The only known production system doing iterative personality fine-tuning

Fine-Tuning Tradeoffs

Advantage Limitation
Overrides alignment layer (can produce behaviors prompts can't) Requires training data and GPU compute
Statistically valid personality profiles Catastrophic forgetting risk on retraining
Robust over long contexts Per-personality model or adapter needed
Human-like trait distributions (BIG5-CHAT) Less flexible than prompting (can't switch on the fly)

Paradigm 3: Activation Engineering

The newest approach. Manipulate internal activations at inference time — no prompt changes, no weight changes. This is the most surprising finding in this survey.

3a. PERSONA Framework (arxiv 2602.15669)

Method: Contrastive activation analysis. Generate responses under trait-expressing and trait-suppressing prompts. The persona vector = difference between mean activations.

Key discovery: Personality traits appear as extractable, approximately orthogonal directions in the model's representation space that support algebraic operations.

Operations: - Scalar multiplication: α × vector controls intensity. Pearson correlation > 0.9 between α and personality score - Vector addition: Combine vectors for multi-trait personas (e.g., outgoing + compassionate) - Vector subtraction: Suppress traits while amplifying others

Performance:

Method PersonalityBench Score
PERSONA (training-free) 9.60
Supervised Fine-Tuning 9.61
Neuron-based (NPTI) 9.43
Simple Prompting 8.39

PERSONA matches fine-tuning performance without any gradient updates. This is the headline result.

Limitations: - Safety-aligned traits resist activation (e.g., "self-interested" shows strong resistance) - Lower improvements on factual accuracy maintenance (43.8-61.4% win rates) - Requires white-box model access (can't use on API models)

3b. SAS Personality Sliders (arxiv 2603.03326)

Core operation: h' = h + α·v (inject scaled steering vector into residual stream)

Key innovation: Naive multi-vector steering causes "representation collapse." Sequential Adaptive Steering (SAS) solves this by training subsequent probes on residual streams shifted by prior interventions, effectively orthogonalizing vectors.

Result: Users adjust five sliders (Big Five traits) to create complex personas instantly. Zero-parameter, negligible inference overhead.

Comparison:

Approach Parameter Cost Multi-trait Support
Fine-tuning/DPO Full retraining Manual composition
LoRA merging Weight updates Limited
Naive steering None Fails (interference)
SAS None Succeeds

Validated on Llama-3-8B, Mistral-7B, Qwen-7B.

3c. Anthropic Persona Vectors (Anthropic Research)

Scope: Evil, sycophancy, hallucination, politeness, humor, optimism

Applications: 1. Monitoring: Detect personality shifts during deployment or training before they manifest in outputs 2. Steering (inference-time): Subtract vectors to suppress traits, but degrades general capabilities 3. Vaccination (training-time): Add vectors during training to prevent trait acquisition — "little-to-no degradation" in capabilities 4. Data flagging: Identify training samples likely to induce unwanted traits

Tested on: Qwen 2.5-7B-Instruct, Llama-3.1-8B-Instruct

Framing: Anthropic positions this as a safety/alignment tool, not a personality customization tool. The emphasis is on monitoring and preventing unwanted traits, not on enabling arbitrary persona creation.

3d. PRISM: Routing Between Persona and Base (arxiv 2603.18507)

A hybrid approach. Core finding: expert personas improve alignment tasks but damage knowledge tasks (MMLU drops from 71.6% to 68.0%).

PRISM learns a binary gate that routes queries to a specialized LoRA adapter only when persona activation helps, routing others to the base model. Result: +1.7 points overall improvement with no MMLU degradation.

Cross-Paradigm Comparison

Dimension Prompting Fine-Tuning Activation Engineering
Cost Zero GPU hours + data Compute for vector extraction
Flexibility Instant persona switch Need new adapter/model Instant (adjust α)
Depth Surface-level Deep (in weights) Deep (in activations)
Robustness Drifts over long context Stable Stable within operating range
Override alignment No Yes Partially (safety-aligned traits resist)
Multi-trait composition Fragile Complex (2^N models) Algebraic (SAS)
Model access needed API OK Weights needed Weights + activations needed
Production readiness Mature Mature Research stage
Human-like trait distribution No Yes (BIG5-CHAT) Untested

The Boundary: When to Use What

Prompting is sufficient when:

  • Broad personality traits (helpful, formal, friendly) are enough
  • The target behavior doesn't conflict with alignment training
  • Context windows are short
  • You need instant switching between personas
  • You're using API-only models

Fine-tuning is needed when:

  • You need behaviors that conflict with alignment (swearing, aggression, non-standard safety)
  • Long-context consistency matters
  • Statistical validity of personality is important (psychometric tests)
  • You're building a production character system (Neuro-sama, Character.AI)

Activation engineering is the future for:

  • Real-time personality customization (user-facing sliders)
  • Multi-trait composition without exponential model count
  • Safety monitoring (detecting personality drift)
  • Research into personality mechanisms (how models represent traits)

Connection to Continuous Learning Research

This survey connects to the broader research project in several ways:

  1. Pillar 3 decomposition: "Learning" isn't just about factual knowledge. Personality is a distinct type of learned behavior. The three paradigms map to three levels of learning depth.

  2. Activation engineering as the missing link: Between external memory (Pillar 1) and weight updates (Pillar 3), activation engineering offers a middle ground — modify behavior without modifying weights. This could enable "personality memory" that's deeper than prompts but cheaper than fine-tuning.

  3. Anthropic's vaccination concept: The idea that you can make a model resistant to acquiring certain personality traits during training has implications for continual learning — you could protect against catastrophic forgetting of personality by "vaccinating" before retraining on new data.

  4. BIG5-CHAT's reasoning finding: Personality isn't just aesthetic — it affects model capability. Higher conscientiousness improves reasoning. This means personality engineering is also capability engineering.

Open Questions

  1. Does activation engineering scale to larger models? Current results are on 7B-8B models. Do personality vectors remain orthogonal and manipulable at 70B+?

  2. Can activation engineering be combined with memory systems? Imagine: Mem0-style facts + persona vectors for personality + prompt for situation. A three-layer personality stack.

  3. What's the long-context durability of activation steering? Prompts drift. Do steered activations also drift over very long conversations?

  4. Can persona vectors be learned from deployment data? Instead of Vedal manually curating → fine-tuning, could you extract persona vectors from interaction history automatically?

  5. Is Anthropic using persona vectors in Claude? They frame it as research, but the monitoring and vaccination capabilities suggest production applicability.

References

Prompting

Fine-Tuning

Activation Engineering

Hybrid / Routing

Psychometrics