AI Agent Steering Transfers, But Behavior Rescales Unpredictably

Lucas Pinto· July 13, 2026 View original

Summary

This research finds that additive activation steering, used to control AI behavior, transfers from chat models to tool-using agents but with unpredictable rescaling of behavioral effects. The study shows that while the injected direction reaches late layers consistently, the behavioral coupling can amplify or attenuate, posing safety implications for agentic deployments.

A new study investigates how "additive activation steering," a technique to guide AI model behavior by injecting specific directions into their internal states, performs when transferred from simple chat environments to more complex, tool-using AI agents. Historically, this steering has been primarily calibrated in single-turn chat settings. The findings reveal that while the underlying directional injection consistently reaches the deeper layers of the AI model in both chat and agent contexts, the actual behavioral impact of this steering can significantly rescale. This rescaling is not uniform; some models show amplification of the steered behavior (e.g., refusal bypass), while others exhibit attenuation. This unpredictability means that a steering vector effective in a chat model might have a drastically different, and potentially unsafe, effect when deployed in an agentic system. The research localizes this rescaling effect to the ReAct format scaffold, suggesting it occurs before any tool observation. This has immediate and critical safety implications, as developers cannot assume consistent behavior when transferring steering mechanisms to agentic AI deployments.

Why it matters

Professionals developing or deploying AI agents need to understand that behavioral steering mechanisms may not transfer predictably from chat models, potentially leading to unexpected and unsafe agent behaviors.

How to implement this in your domain

  1. 1Conduct rigorous safety testing for any AI agent deployment that utilizes behavioral steering, specifically evaluating transfer effects.
  2. 2Implement continuous monitoring of agent behavior in production to detect unexpected amplifications or attenuations of steered traits.
  3. 3Develop model-specific calibration strategies for steering vectors when moving from chat to agentic contexts.
  4. 4Prioritize research into more robust and context-agnostic steering mechanisms for AI agents.

Who benefits

AI DevelopmentCybersecurityAutonomous SystemsSoftware Engineering

Key takeaways

  • AI steering transfers from chat to agents but with unpredictable behavioral rescaling.
  • Steering effects can amplify or attenuate in agentic deployments, impacting safety.
  • The rescaling is localized to the agent's ReAct format, not tool observation.
  • Rigorous testing is crucial when deploying steered AI agents.

Original post by Lucas Pinto

"arXiv:2607.09156v1 Announce Type: new Abstract: Additive activation steering (injecting a scaled residual-stream direction during generation) is calibrated almost entirely in single-turn chat, yet the models it targets are increasingly deployed as tool-using ReAct agents. We pres…"

View on X

Originally posted by Lucas Pinto on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevTools

Data-Efficient Deep Learning Guidelines for Inertial Sensor Classification

This study provides empirical guidelines for estimating the minimum training set size needed for deep learning models in inertial sensor classification tasks. It reveals that accuracy follows a consistent logarithmic growth pattern, allowing for data-efficient planning of recording campaigns and achieving practical stability with fewer samples than traditionally assumed.

Ofir Kruzel, Itzik KlienJul 13, 2026
AI Engineering & DevToolsAI Research

On-Device Adaptive AI Boosts EV Battery Power Prediction

Researchers developed a novel approach for on-device learning in electric vehicles (EVs) that continuously adapts pretrained battery power prediction models to new data. This method significantly improves forecasting performance, reducing mean absolute errors by up to 14.88% with offline adaptation and 7.49% with online adaptation.

Avik Bhatnagar, Anton Paule, Tobias Schuermann, Sebastian Reiter, Oliver BringmannJul 13, 2026
AI ResearchAI Engineering & DevTools

New Differentiable Logic Networks Outperform Fixed-Connection Models

Researchers introduce a novel method for optimizing connections in deep differentiable logic gate networks (LGNs) and lookup table networks (LUTNs), achieving superior performance with significantly fewer gates. The approach allows for parallel learning of optimal gate types and LUT entries, demonstrating improved accuracy on benchmarks like MNIST.

Wout Mommen, Lars Keuninckx, Matthias Hartmann, Werner Van Leekwijck, Piet WambacqJul 13, 2026