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

Yingdong Shi, Ruiming Zhang, Changming Li, Zhiyu Yang, Kaixing Zhang, Jingyi Yu, Kan Ren

arXiv:2605.30076 · 2026-05-29 공개 · arXiv · PDF

flow-matching instruction-following residual-stream text-guided activation-space behavioral-control fine-grained-concept activation-classification

Abstract

Activation-based control steers large language models (LLMs) by intervening on their internal representations during inference, and has emerged as an effective paradigm for controlling behaviors such as persona and style. However, existing methods often rely on fixed steering directions or task-specific intervention modules, making them difficult to adapt to fine-grained concepts and compositional constraints. We propose UniSteer, a text-guided activation flow matching model that learns a conditional distribution over residual-stream activations from natural-language conditions. Instead of fitting a separate intervention for each target behavior, UniSteer learns a universal conditional velocity field in activation space. At inference time, UniSteer performs flow inversion by partially transporting a source activation toward a latent state and regenerating it under a target textual condition before injecting it back into the frozen LLM. The same conditional model supports activation-space classification by selecting the textual label with the lowest reconstruction energy. Experiments on three target LLMs show that UniSteer provides a unified interface across behavioral control, truthfulness steering, fine-grained concept steering, multi-constraint instruction following, and activation-space classification.

한국어 요약

📋 한 줄 요약

**[Activation Steering / Flow Matching]** UniSteer가 residual-stream activation의 conditional 분포를 텍스트 조건 flow matching으로 학습 — flow inversion으로 활성을 재생성·주입, 행동제어·진실성·세밀 개념·multi-constraint·classification까지 통합 인터페이스.

🎯 핵심 기여도

💡 핵심 아이디어

Activation-based 제어의 task-specific module 부담은 universal한 conditional velocity field를 text-guided flow matching으로 학습해 단일 모델로 해소할 수 있으며, flow inversion 기반 부분 transport·재생성 메커니즘이 frozen LLM에 다양한 control을 유연하게 주입한다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: Activation steering의 unified framework 정립, 자연어 조건으로 유연한 fine-grained 제어 가능, flow matching의 generative 능력을 활성 공간에 도입, classification까지 동일 모델에서 가능한 양면성, frozen LLM에 plug-in 가능. **한계**: Conditional flow matching 학습 자체 비용, 매우 fine-grained·OOD 개념의 표현력은 학습 데이터 의존, residual-stream 특정 위치 의존성.

🚀 실용적 활용