Activation Steering for Aligned Open-ended Generation without Sacrificing Coherence
Niklas Herbster, Martin Zborowski, Alberto Tosato, Gauthier Gidel, Tommaso Tosato
Why It Matters
What makes this one worth your time
As LLMs become more integrated into applications, ensuring their alignment and coherence is critical for safe and reliable deployment, making this research relevant for both developers and researchers in AI safety.
This research introduces novel methods for improving alignment in LLMs through activation steering.
Summary
The paper presents three methods for activation steering in large language models (LLMs) to recover alignment and mitigate misalignment behaviors, demonstrating their effectiveness on specific threat models and architectures.
Key contributions
- Introduction of three activation steering methods: SwFC, StTP, and StMP.
- Demonstration of substantial recovery of alignment in LLMs using these methods.
- Evaluation of methods on multiple architectures and threat models, showcasing their effectiveness in maintaining general capabilities.
Notable insights
- Activation steering can selectively intervene in token activations based on a logistic regression decision boundary, which may enhance the precision of alignment recovery.
- The methods show generalization capabilities across different scenarios, indicating robustness in addressing emergent misalignment.
Possible limitations
- Not stated in the abstract.
Abstract
arXiv:2604.08169v2 Announce Type: replace Abstract: Alignment in LLMs is more brittle than commonly assumed: misalignment can be induced by adversarial prompts, benign fine-tuning, emergent misalignment, and goal misgeneralization. Recent evidence suggests that some misalignment behaviors are encoded as linear structure in activation space, making it tractable via activation steering, which could be used as a lightweight runtime defense. We implement three methods: Steer-With-Fixed-Coefficient (SwFC), which applies uniform additive steering, and two novel projection-aware methods, Steer-to-Target-Projection (StTP) and Steer-to-Mirror-Projection (StMP), that use a logistic regression decision boundary to selectively intervene only on tokens whose activations fall below the threshold. We evaluate these methods on two threat models, dishonesty and dismissiveness, using malicious system prompts as a controlled proxy for misalignment. We conduct our experiments on two architectures (Llama-3.3-70B-Instruct and Qwen3.6-27B). All methods substantially recover alignment. StTP and StMP preserve general capabilities (MMLU, MT-Bench, AlpacaEval) better than uniform steering. Finally, we show that our honesty steering generalizes to out-of-distribution scenarios: a single honesty direction extracted from the aligned model significantly raises scores on the MASK benchmark, suppresses deception in multi-agent settings (Among Us), doubles the hidden-behavior discovery rate on AuditBench, and restores honesty in an emergently misaligned model