[Paper Review] Fusian: Multi-LoRA Fusion for Fine-Grained Continuous MBTI Personality Control in Large Language Models

Zehao Chen, Rong Pan|arXiv (Cornell University)|Mar 16, 2026

Personality Traits and Psychology0 citations

TL;DR

Fusian introduces a two-stage framework that collects a trajectory of LoRA adapters during SFT to form a continuous personality manifold and uses an RL policy to dynamically fuse these adapters for precise, continuous MBTI trait intensity control in LLMs.

ABSTRACT

Large Language Models (LLMs) have demonstrated impressive capabilities in simulating diverse human behaviors and personalities. However, existing methods for personality control, which include prompt engineering and standard Supervised Fine-Tuning (SFT), typically treat personality traits as discrete categories (e.g., "Extroverted" vs. "Introverted"), lacking the ability to precisely control the intensity of a trait on a continuous spectrum. In this paper, we introduce Fusian, a novel framework for fine-grained, continuous personality control in LLMs. Fusian operates in two stages: (1) Trajectory Collection, where we capture the dynamic evolution of personality adoption during SFT by saving a sequence of LoRA adapters, effectively mapping the continuous manifold of a trait; and (2) RL-based Dynamic Fusion, where we train a policy network using Reinforcement Learning to dynamically compute mixing weights for these frozen adapters. By sampling from a Dirichlet distribution parameterized by the policy network, Fusian fuses multiple adapters to align the model's output with a specific numerical target intensity. Experiments on the Qwen3-14B model demonstrate that Fusian achieves high precision in personality control, significantly outperforming baseline methods in aligning with user-specified trait intensities.

Motivation & Objective

• Motivate the need for continuous, fine-grained personality control beyond discrete MBTI categorizations.
• Propose a two-stage framework (Trajectory Collection and RL-based Dynamic Fusion) to model and control personality intensity along a continuous spectrum.
• Leverage LoRA-based parameter-efficient fine-tuning to construct a trajectory manifold of personality traits.
• Develop a policy-driven fusion mechanism that samples adapter weights from a Dirichlet distribution to match target trait intensities.
• Demonstrate superior precision and monotonicity in controlling MBTI dimensions on Qwen3-14B compared to baselines.

Proposed method

• Stage 1 collects high-frequency LoRA adapters during SFT and evaluates them with MBTI tests to build a continuous personality manifold (trajectory library).
• Stage 2 trains an RL policy that maps a target intensity to Dirichlet-distributed mixing weights over basis adapters, enabling dynamic fusion of adapters.
• A fused adapter is computed as a weighted sum of basis adapters and used for inference to achieve the desired intensity.
• An aggressive reward shaping function drives the RL optimization to minimize deviation from target intensity.
• Stability and coverage of the adapter basis are ensured via a stable basis selection process and uniform sampling across the trait spectrum.

Experimental results

Research questions

• RQ1 Can continuous MBTI trait intensities be accurately controlled in LLMs beyond discrete type assignments?
• RQ2 How can we extract a robust, continuous personality manifold from the SFT trajectory using PEFT?
• RQ3 Can RL-driven fusion of multiple LoRA adapters achieve precise alignment with a numerical target intensity?
• RQ4 How does Fusian compare to prompting and other parameter-space methods in terms of MAE and monotonicity across MBTI dimensions?
• RQ5 What ablations reveal about the importance of stable basis selection and reward design for precise control?

Key findings

• Fusian achieves the lowest overall MAE (6.79) among compared methods on MBTI control.
• Fusian attains the highest overall Pearson correlation (0.88) indicating strong monotonic control across dimensions.
• Compared to standard LoRA, Fusian reduces MAE by a substantial margin and improves correlation (example: Thinking dimension MAE 5.44 vs. 12.80; r up to 0.97).
• Prompt-based methods show limited precision (e.g., gpt-5-mini Prompt MAE 14.97; r 0.35 overall).
• Ablations show that dynamic fusion, stable basis, and aggressive reward are crucial for performance (e.g., w/o Dynamic Fusion: MAE 10.02, r 0.53).
• Case study demonstrates smooth semantic shifts in Feeling intensity from analytical to empathetic responses.