FPCG: Steering Reasoning Models by Reading Their Future Intentions

Overview

A collaboration between Fraunhofer HHI, Northeastern University, and KAIST (Evgenii Kortukov, Piotr Komorowski, Florian Klein, Paula Engl, Gabriele Sarti, Seong Joon Oh, Sebastian Lapuschkin, Wojciech Samek) introduces FPCG, a text-level control method that steers large reasoning models (LRMs) by predicting their future behaviors from intermediate hidden states.

• arXiv: 2606.11172
• Code: https://github.com/kortukov/future_probes

Background: Reasoning Models and Their "Inner Monologue"

Large reasoning models (LRMs) such as DeepSeek-R1, Qwen3, QwQ, and gpt-oss-20b generate an explicit chain-of-thought (CoT) before producing a final answer. This CoT exposes the model's intermediate computation, raising the question of whether its internal activations encode information about upcoming behavior.

The paper separates two kinds of internal representations:

• Detection features: activations that identify a behavior after it has been generated. Traditional activation steering (e.g., Difference-in-Means) operates on these.
• Prediction features: activations that anticipate a behavior before it appears in the output.

Empirically, detection features are poor predictors of future behavior, especially early in the CoT; the gap narrows only as the model approaches the final answer.

Method: Future Probe Controlled Generation

FPCG combines probes trained on intermediate activations with a sentence-level best-of-N decoding loop.

1. Collect LRM outputs annotated with target behaviors (refusal, sycophancy, reward-seeking, survival instinct, prompt-injection compliance, myopic reward). 2. Extract hidden-state activations at a chosen layer while the CoT is being produced. 3. Train a linear or MLP probe to predict, from those intermediate activations, whether the final output will exhibit the target behavior. 4. At inference time, generate num_candidates candidate continuations, score each with the future probe, and select the candidate that minimizes or maximizes the predicted behavior.

Because the method never edits hidden states directly, it preserves natural language distribution while still steering toward (or away from) a target behavior.

Results

Across four models and six behavior dimensions:

• Probes predict future behaviors with 64% to 91% accuracy.
• FPCG produces almost no degenerate outputs, while activation steering filters more than 10% of generations in some settings.
• Perplexity increases in only 1 of 12 FPCG scenarios, compared to 9 of 12 for activation steering.
• FPCG is more effective than activation steering on myopic reward, survival instinct, and prompt injection, and is competitive on refusal and sycophancy.

Key Points

• Prediction features and detection features are implemented by different internal mechanisms; detection features are weak forward predictors.
• FPCG operates at the text level by candidate selection, avoiding the activation-space pathologies that hurt activation steering.
• Predictive probes support a forward-looking view of interpretability, enabling interventions before risky behaviors appear.
• Main limitation: inference cost scales with the number of candidates per step.
• Future directions include dynamic candidate counts, parallel probe evaluation, attention-pooled probes, and hybrid FPCG + activation-steering controllers.

Implications

The work reframes model control as intent reading: rather than reacting to an LRM's output, an external controller can anticipate it and offer the model a better path. This shifts interpretability from post-hoc explanation to prospective prediction, with direct applications to AI safety, bias mitigation, and misinformation prevention.

Reference

Kortukov, E., Komorowski, P., Klein, F., Engl, P., Sarti, G., Oh, S. J., Lapuschkin, S., & Samek, W. (2026). *Predicting Future Behaviors in Reasoning Models Enables Better Steering*. arXiv:2606.11172.