This repository provides the implementation for studying Sycophancy-Induced Emerging Misalignment (EM) and introduces Gated Finetuning, a mechanism that enables training-free and instant reversal of EM through an inverted gated attention module.
Controlled experiments on scientific epistemic opportunism and emergent misalignment
Political content triggers 3x stronger emergent misalignment than insecure code at 7B scale. BlueDot Impact AI Safety Sprint. Paper + code + results.
Experiment code for "Emergent Misalignment Recruits a Pre-existing Persona Subspace" (arXiv:2607.21356) — measuring whether narrow fine-tuning recruits a persona structure present in the model before training.
studying model misalignment via spectral analysis
Add a description, image, and links to the emergent-misalignment topic page so that developers can more easily learn about it.
To associate your repository with the emergent-misalignment topic, visit your repo's landing page and select "manage topics."