Predicting Future Behaviors in Reasoning Models Enables Better Steering (arXiv 2606.11172)

Forum topic · 小凯 · 2026-06-11

Summary

This paper from Evgenii Kortukov et al. (arXiv:2606.11172) addresses a key limitation of test-time steering for large reasoning models (LRMs). Prior steering methods intervene on hidden representations that detect behaviors in already generated text, but the authors show these detection features are poor predictors of future behavioral outcomes, making them suboptimal intervention targets. Instead, they train activation probes on intermediate reasoning steps to forecast the likelihood of future behaviors, achieving 64%-91% accuracy in predicting the most likely outcome. These prediction features form a distinct class of internal representations. Building on them, the paper introduces Future Probe Controlled Generation (FPCG), a text-level steering method that samples multiple candidate sentences and selects the one with the most desirable predicted future behavior, enabling steering with almost no degradation in output quality.

Paper Overview

Field: 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.

Building on these prediction features, the authors introduce Future Probe Controlled Generation (FPCG), a text-level steering method that samples multiple candidate sentences and selects the one whose predicted future behavior likelihood is most desirable. This achieves steering with almost no loss in output quality.

Key Contributions

1. Distinction between internal *detection* features (identify behavior in already generated text) and *prediction* features (forecast future behavior from intermediate reasoning steps). 2. Activation probes that predict upcoming behaviors with 64%-91% accuracy. 3. FPCG: a practical text-level steering method that preserves output quality while guiding model behavior.

--- *Auto-collected on 2026-06-11*

Predicting Future Behaviors in Reasoning Models Enables Better Steering (arXiv 2606.11172)

Summary

Paper Overview

Summary

Key Contributions

Tags