LLM agents that grow personalities through interaction

What it does

JPAF straps a Jungian psychological model onto off-the-shelf LLMs. It assigns an MBTI type expressed through weighted cognitive functions—Ti, Si, Ne, and the rest—then maintains coherence via dominant-auxiliary coordination, short-term adaptation via reinforcement-compensation, and long-term drift via a reflection mechanism. The goal is an agent that doesn’t just role-play a personality, but ostensibly develops it.

The interesting bit

Instead of simple prompting, the framework treats personality as a control system: cognitive functions carry numeric weights that shift in response to scenario demands, and the model is expected to justify its answers by citing which function it activated. It is essentially a structured prompt harness wrapped around a psychology experiment.

Key highlights

• The authors report 100% MBTI alignment accuracy on standard questionnaires for GPT and Qwen, and 92–100% personality evolution accuracy across tested models.
• Supports all 16 MBTI types through eight Jungian cognitive functions with fine-grained weight differentiation.
• Two experiment tracks: one locks a personality and tests consistency against 70- or 93-question batteries, the other runs scenario challenges to see if weights drift.
• Validated against GPT-4, Qwen, and Llama families via API integration.

Caveats

• Requires API credentials for OpenAI, Qwen, or Llama endpoints; local inference is not documented.
• Accuracy claims are based on curated MBTI questionnaires and hand-written scenarios, so open-world robustness is untested.
• The repository is primarily an evaluation harness and scenario runner, not a general SDK.

Verdict

Worth exploring if you are building NPCs, social simulations, or multi-agent systems that need differentiated characters capable of gradual change. Give it a pass if you need a fully offline solution or reject the MBTI model outright.