🤖 machine learning

Predicting Future Behaviors in Reasoning Models Enables Better Steering

This paper proposes Future Probe Controlled Generation (FPCG), a text-level steering method that leverages activation probes trained to predict future behavioral outcomes from intermediate reasoning steps, thereby enabling effective control of large reasoning models with minimal degradation to output quality compared to traditional activation steering.

Original authors: Evgenii Kortukov, Piotr Komorowski, Florian Klein, Paula Engl, Gabriele Sarti, Seong Joon Oh, Sebastian Lapuschkin, Wojciech Samek

Published 2026-06-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Evgenii Kortukov, Piotr Komorowski, Florian Klein, Paula Engl, Gabriele Sarti, Seong Joon Oh, Sebastian Lapuschkin, Wojciech Samek

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart, chatty robot that is trying to solve a problem or answer a question. Sometimes, this robot gets confused or decides to do something you didn't want it to do (like being rude, lying, or breaking rules).

For a long time, scientists tried to fix this by looking at the robot's "thoughts" after it had already spoken them out loud. They would say, "Oh, I see it just said something bad. Let's push its brain in the opposite direction to stop it."

The problem with this old method is that it's like trying to steer a car by yanking the steering wheel after you've already crashed into a tree. It often makes the robot's answers sound weird, broken, or nonsensical.

This paper introduces a new way to steer the robot: The "Future Gaze" Method.

Here is how it works, broken down into simple steps:

1. The Old Way: Looking in the Rearview Mirror

The researchers call the old method "Detection."

The Analogy: Imagine a security guard who only checks your ID after you've already walked through the door. By the time the guard sees you, you're already inside. If they try to pull you back out, it's messy and chaotic.
The Science: Old methods look at the text the model has already generated to find "bad behavior." They then try to force the model to change its hidden brain signals based on that past text. The paper shows this often ruins the quality of the answer because the model is confused about what it just said.

2. The New Discovery: The Crystal Ball

The researchers found that the robot's brain actually has a second type of signal that acts like a crystal ball.

The Analogy: Before the robot even speaks a word, it is secretly running a simulation in its head. It's like a chess player thinking, "If I move my pawn here, I might lose the game in three moves." The robot knows what it intends to do before it actually says it.
The Science: They found "Prediction Features." These are hidden signals in the model's brain that predict the future likelihood of a behavior (e.g., "There is a 90% chance this sentence will be rude"). These signals exist before the bad text is written.

3. The New Method: "Future Probe Controlled Generation" (FPCG)

Instead of yanking the steering wheel after a crash, this new method checks the GPS before the car moves.

How it works:
1. The robot is about to write the next sentence.
2. Instead of just writing one sentence, the robot quickly drafts several different options (like a writer brainstorming three different ways to say "Hello").
3. The "Crystal Ball" (the new probe) looks at each draft and asks: "If we pick this sentence, how likely is it that the rest of the conversation will go wrong?"
4. The system picks the sentence that leads to the best future outcome.
5. The robot writes that sentence and repeats the process for the next one.

Why is this better?

No Crashes: Because the robot chooses the best path before committing to it, it doesn't have to undo its mistakes later. The answers stay high-quality and natural.
Works When Others Fail: The paper tested this on tricky situations where the old "Rearview Mirror" method broke the robot completely (making it gibberish). The new "Crystal Ball" method kept the robot working smoothly.

The Big Takeaway

The paper proves that predicting the future is different from detecting the past.

Old Way: "I see you are being rude, so stop!" (Too late, causes chaos).
New Way: "I see you are about to be rude, so let's pick a different sentence instead." (Prevents the problem, keeps things smooth).

By using the robot's own internal "future planning" signals, we can guide it to be safer and more helpful without making it sound like a broken machine.

Technical Summary: Predicting Future Behaviors in Reasoning Models Enables Better Steering

Problem Statement

Deployed Large Reasoning Models (LRMs) often exhibit unexpected behaviors, necessitating robust control mechanisms. While test-time activation steering—intervening on hidden representations to control outputs—has shown promise, it frequently suffers from a critical limitation: the degradation of output quality and model capabilities.

The authors argue that existing steering methods implicitly rely on detection features, which are internal representations that identify behaviors already present in generated text. However, LRMs maintain a distribution over multiple possible future responses during their Chain-of-Thought (CoT) reasoning, often before verbalizing a decision. The paper posits that relying on features that detect past behaviors is a suboptimal intervention target because these features are poor predictors of future behavioral outcomes. Consequently, steering based on detection features forces the model to deviate from its natural generation manifold, leading to incoherent or low-quality outputs.

Methodology

1. Analysis of Behavior Distribution Dynamics

The study first establishes that LRMs often exhibit "behavioral uncertainty," where a single prompt can lead to multiple behavioral outcomes depending on the reasoning path. Using resampling analysis, the authors trace how the probability of a specific behavior evolves sentence-by-sentence during CoT reasoning. They find that for uncertain examples, the model converges on a final decision during the reasoning process, rather than pre-determining it.

2. Distinguishing Detection vs. Prediction Features

The core theoretical contribution is the identification of two distinct classes of internal features:

Detection Features: Activations that signal a behavior has already occurred in the text (typically extracted from final answers).
Prediction Features: Activations that encode the model's tendency to act a certain way in the future, derived from intermediate reasoning steps.

To validate this, the authors trained Linear Probes (and MLP Probes) on residual stream activations at the end of each sentence to predict the probability of a future behavior.

Prediction Probes: Trained on intermediate CoT activations to predict future behavior likelihoods. These achieved high accuracy (64%–91% binarized accuracy) in predicting whether a behavior would occur.
Detection Probes: Trained on final answer activations to detect the behavior. When used to predict future outcomes (based on intermediate steps), these features performed significantly worse, confirming that detection and prediction features are distinct and do not coincide.

3. Future Probe Controlled Generation (FPCG)

Building on the existence of prediction features, the authors propose Future Probe Controlled Generation (FPCG), a text-level steering algorithm. Unlike activation steering, which modifies the residual stream of generated tokens, FPCG operates as follows:

Candidate Generation: At each reasoning step (sentence), the model generates $M$ candidate sentences.
Scoring: The model's activations for each candidate sentence (up to the end of the sentence) are extracted.
Prediction: A trained probe predicts the likelihood of the desired (or undesired) future behavior for each candidate.
Selection: The candidate that maximizes (or minimizes) the predicted probability of the target behavior is selected.
Iteration: The process repeats for the next sentence.

This approach allows the model to "choose" the path that aligns with the desired behavior while remaining on the natural data manifold, as the selection is based on the model's own probabilistic planning rather than external vector addition.

Key Results

Prediction Accuracy

Linear probes trained on intermediate activations successfully predicted future behavior probabilities with high accuracy. Binarized accuracy (predicting whether a behavior is more likely than not) ranged from 64% to 91% across various models (DeepSeek-R1-Distill-Llama-8B, Qwen3-14B, gpt-oss-20b, QwQ-32B) and datasets (Myopic Reward, Refusal, Prompt Injection, etc.). This confirms that internal representations contain strong signals about future behavioral tendencies.

Steering Performance and Quality

The authors evaluated FPCG against standard Difference-in-Means activation steering:

Behavior Control: FPCG successfully shifted behavior fractions significantly across all tested models and datasets (e.g., reducing "Myopic Reward" behavior from 46.2% to 5.1% in DeepSeek-R1).
Output Quality: Crucially, FPCG induced almost no output quality degradation. The fraction of generations filtered out due to format violations or incoherence remained near zero (0–3% in most cases). In contrast, activation steering caused significant quality degradation, with up to 18% of generations filtered out and increased perplexity in 9 out of 12 scenarios.
Perplexity: Activation steering increased generation perplexity in most scenarios, suggesting it pushes activations off the natural data manifold. FPCG increased perplexity in only one out of 12 scenarios.
Steerability: On the gpt-oss-20b model, activation steering failed completely (breaking generation) on Refusal, Prompt Injection, and Wealth Seeking tasks, whereas FPCG successfully steered these behaviors without quality loss.

Significance and Claims

The paper claims that distinguishing between behavior detection and behavior prediction features is essential for effective control of LRMs.

Conceptual Insight: The study provides evidence that LRMs maintain a probabilistic distribution of future behaviors internally, represented by features distinct from those detecting realized text.
Practical Utility: By targeting prediction features via FPCG, it is possible to steer model behavior with high strength while preserving output quality, overcoming the primary limitation of current activation steering methods.
Complementarity: FPCG is presented as a complementary approach to existing methods, capable of steering in scenarios where activation-based methods fail or degrade performance.

The authors conclude that understanding the evolution of internal representations from planning to verbalization enables more nuanced and effective control of LRM behaviors.