I measured the "personality" of 6 open-source LLMs (7B-9B) by probing their hidden states. Here's what I found.
Your AI assistant has a hidden personality—and researchers just measured it.
Deep Dive
A new tool probes hidden states in 7B-9B open-source LLMs across 7 behavioral axes, revealing distinct personality fingerprints. DeepSeek 7B scores as verbose (+1.00), confident (+0.97), and proactive (+1.00)—the "enthusiastic explainer." Llama 8B is eerily neutral, while Yi 9B is slightly cold and patient. Validation shows 93-100% calibration accuracy on 4/6 models and mean ICC reliability of 0.91–0.99. Some models have "dead zones" where steering fails across all prompts.
Why It Matters
This proves AI models have inherent behavioral biases that affect user experience, regardless of prompting.