Researchers Propose DPN-LE Technology: Directly Edit Large-scale Model 「Personality Neurons」 for Fine-grained AI Personality Control

May 3 — AI researcher Brian Roemmele disclosed that his “Zero-Human Company” has begun rolling out DPN-LE (Dual Personality Neuron Localization and Editing), a large-scale personality editing tech built to fine-tune AI agents’ behavioral traits with precision. DPN-LE identifies “personality neurons” in large-scale MLP layers to selectively boost or dial back traits like honesty, creativity, prudence, and cooperativeness. The research claims this method only needs to edit ~0.5% of neurons to deliver “surgical” personality adjustments—no model retraining required—while minimizing harm to core reasoning abilities. Roemmele noted his company, now home to over 100 AI agents, runs almost entirely autonomously, making “personality consistency” a top priority. Examples include: - Strategy agents: Boosted “long-term consistency” and “analytical honesty” - Financial/risk control agents: Enhanced “prudence” and “precision” - Content agents: Amplified “empathy” and “creativity” - Multi-agent collaboration layer: Strengthened “cooperativeness” He added that since DPN-LE uses lightweight inference-stage modifications, the company can reshape personalities in minutes and deploy dozens of AI agents—slashing AI alignment costs significantly. Market observers say this “personality programmability” tech marks a shift: AI agents are moving beyond pure capability competition to vying for edge in behavioral control and organizational governance.