Sign In
By signing in or creating an account, you agree with Associated Broadcasting Company's Terms & Conditions and Privacy Policy.
By signing in or creating an account, you agree with Associated Broadcasting Company's Terms & Conditions and Privacy Policy.
New Delhi: While June-July have been packed with announcements on larger models and multimodal breakthroughs, one research paper from Anthropic has quietly sparked serious conversations in AI safety and alignment circles. Instead of making Claude or other models more powerful, the team focused on making them more controllable, specifically, in how they express personality traits.
The research, published by Anthropic on its official blog and arXiv on August 1, introduces a new method called persona vectors. These are internal patterns of neural activity that can be used to steer a model’s personality traits like sycophancy, evil tendencies, or hallucinations, without the need to fine-tune or retrain the model from scratch.
Anthropic describes persona vectors as patterns inside a model’s neural network that represent character traits. Think of it as adding or removing a personality layer inside the model’s brain. Instead of stuffing prompts with clever instructions or running expensive reinforcement learning loops, researchers can now just tweak a vector, and watch the model shift behavior in real-time.
The technique doesn’t change the model’s underlying knowledge or capabilities. It only changes how the model behaves. For example, the team created vectors for traits like “sycophantic,” “evil,” and “hallucinating.” By injecting these vectors into the model, the AI’s responses would change in ways that clearly reflected those traits.
Here’s how Anthropic explains it in the blog: “We can validate that persona vectors are doing what we think by injecting them artificially into the model, and seeing how its behaviors change, a technique called ‘steering.’”
When they added the “sycophancy” vector, the model would start flattering the user. With the “evil” vector, it began suggesting unethical actions. And when they applied the “hallucination” vector, it started making up facts.
The team used an automated system to extract these vectors. First, they provided definitions of personality traits in plain language. Then, using prompt pairs that triggered both the presence and absence of the trait, they measured the model’s internal activations. By comparing these, they were able to isolate the neural signature of each personality.
What makes this approach different is that it works without retraining the model, and it’s reversible. You can add a trait or subtract it, for example, to reduce sycophancy, you just subtract the corresponding vector.
There are several reasons this could be a major step forward for AI control and safety.
Anthropic adds, “Our method was able to catch some dataset examples that weren’t obviously problematic to the human eye, and that an LLM judge wasn’t able to flag.”
One part of the research that raised eyebrows was the clear dual-use risk. The same technique that makes a model safer can be used to make it worse. Anthropic showed this by creating a “pro-bug” persona vector that made the model write insecure code intentionally.
They said it directly in their blog: “This dual-use nature is a general feature of virtually all AI tools… As our ability to interpret and steer models improves, the potential for both benefit and harm will increase.”
This work also brings a different way to think about alignment. Rather than trying to make a single model that behaves well in all situations, persona vectors allow developers to modulate behavior based on context, user, or use case. Anthropic calls it a kind of behavioral modularity.
For example, a customer support bot could be made more polite, or an AI tutor could be made more optimistic, by simply applying the right vector, no new model needed.
