Paper Overview
Research area: Machine Learning Authors: Evgenii Kortukov, Piotr Komorowski, Florian Klein, Paula Engl, Gabriele Sarti, Seong Joon Oh, Sebastian Lapuschkin, Wojciech Samek Published: 2026-06-09 arXiv: 2606.11172
Summary
Deployed large reasoning models (LRMs) often behave unexpectedly. Test-time steering controls LRM outputs by intervening on their hidden representations, but it can degrade output quality. The authors argue that prior steering work implicitly relies on internal features that detect behavior in already generated text. They show that these detection features are poor predictors of future behavioral outcomes, and thus are not the natural intervention target.
Instead, the paper trains activation probes to predict future behavior likelihoods from intermediate reasoning steps. These probes predict the most likely behavior with 64%-91% accuracy, revealing a separate type of internal prediction features.
Method: Future Probe Controlled Generation (FPCG)
Building on these prediction features, the authors introduce Future Probe Controlled Generation, a text-level steering method:
- Sample multiple candidate sentences at each generation step
- Use the prediction probes to evaluate the likelihood of the desired future behavior for each candidate
- Select the candidate with the best predicted future behavior likelihood
- Demonstration that existing detection features are poor predictors of future behaviors
- Activation probes that forecast future behavior likelihoods from intermediate reasoning steps (64%-91% accuracy)
- Identification of a distinct class of internal prediction features as natural intervention targets
- FPCG, a text-level steering method that preserves output quality
This approach achieves steering with almost no loss in output quality, addressing the quality degradation problem of existing representation-level interventions.