UniSteer: Text-Guided Flow Matching in Activation Space for Versatile LLM Steering

Researchers introduced UniSteer, a text-guided flow-matching method that learns a universal conditional velocity field over LLM activations.

It steers model behavior by inverting and reinjecting activation flows, letting natural-language conditions control outputs in activation space.

UniSteer can also handle multi-constraint prompts and perform classification by selecting the label with the lowest reconstruction energy.

The approach aims to provide a single shared mechanism for behavior control without task-specific intervention modules.