Research Portfolio
A small subset of our research work in Military-grade AI, presented through interactive demonstrations.
Explore cutting-edge research in Military-grade AI through interactive demonstrations.
A small subset of our research work in Military-grade AI, presented through interactive demonstrations.
Large language models can detect and actively resist unnatural manipulations of their internal representations
Activation steering means nudging a model's internal activations with a vector so its replies lean toward a target concept (topic/style) even when the prompt doesn't ask for it.
No steering applied
With activation steering
In this example, we boost latents related to "explanatory content about solar and lunar eclipses" while the user asks about calculating square roots. The steering vectors push the model's internal representations toward eclipse-related content, causing it to inject irrelevant eclipse jokes and references throughout what should be a straightforward mathematical explanation.
Large models can sometimes detect when steering has made their responses incoherent and actively self-correct, either spontaneously or with user feedback.
Model detects manipulation and corrects itself without external prompting
Model corrects itself after user feedback
Hover over the chart to see how different meta-latents activate as the model generates each token in response to steering manipulation.
Our research uncovers a novel phenomenon of Endogenous Steering Resistance (ESR) in large language models. This emergent behavior represents a form of intrinsic safeguard that develops as models scale, potentially providing a natural defense mechanism against certain types of manipulation.
The implications of this finding extend beyond basic research into practical applications for AI alignment and interpretability. Understanding how models develop and deploy these resistance mechanisms could inform the development of more robust and trustworthy AI systems.