Researchers propose DPN-LE technology: directly editing large model "personality neurons" to achieve fine-tuned control of AI personalities

BlockBeats News, May 3rd, AI researcher Brian Roemmele revealed that his “Zero-Human Company” has begun deploying a large model personality editing technology called DPN-LE (Dual Personality Neuron Localization and Editing), used for precisely adjusting AI Agent behavioral traits.

According to the report, DPN-LE locates “personality neurons” within the large model’s MLP layers, allowing targeted enhancement or weakening of traits such as “honesty, creativity, caution, and cooperativeness.” The research states that this method requires editing only about 0.5% of neurons, enabling “surgical” adjustments to AI personalities without retraining the model, while minimizing damage to core reasoning abilities.

Roemmele said that his company currently consists of over 100 AI Agents, with almost all daily operations fully automated, making “personality consistency” a key issue. For example:

Strategic Agents will strengthen “long-term consistency” and “analytical honesty”
Financial and risk control Agents will emphasize “caution” and “accuracy”
Content Agents will improve “empathy” and “creativity”
Multi-Agent collaboration layers will enhance “cooperativeness”

He stated that because DPN-LE involves lightweight modifications during the reasoning phase, the company can “restructure personalities” in batches within minutes and deploy a large number of AI Agents, significantly reducing AI alignment costs. The market believes that this kind of “personality programmable” technology signifies that AI Agents are moving beyond simple capability competition to further competition in behavioral control and organizational governance.