Subspace-evasion fine-tuning

methods / Causal Validation / Causal interventions (steering) / Subspace-evasion fine-tuning

Techniqueadvanced

Fine-tunes a model, conditioned on a trigger phrase, to move its activations for a targeted concept into a low-dimensional subspace that evades a probe/activation monitor, and tests whether this evasion skill transfers zero-shot to unseen concepts and unseen monitors.

Used in (1 observation)

Monitor-evasion fine-tuning generalizes zero-shot to unseen safety monitorsmeasured

structure: Linear Subspace · models: Gemma-2-9B-it, Gemma-2-2B-it, Llama-3.1-8B-Instruct, Qwen2.5-7B-Instruct · paper: Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Activation Monitors