Mechanistic Indicators of Steering Effectiveness in Large Language Models

TL;DR

This work investigates activation-based steering in large language models through internal mechanistic signals. It introduces the Normalized Branching Factor ($NBF$) and KL-divergence dynamics as principled indicators of steering effectiveness, framing steering as a residual-intervention within Transformer architectures. By validating with a reliability study using two architecturally distinct LLM judges and comparing stronger baselines (CAA and SAE) with a rotation-based steering method, the authors show that internal signals can predict steering quality and detect failure without external evaluators. The findings support a mechanistic interpretability view of steering and advocate for principled evaluation practices that leverage internal model dynamics to improve reliability and safety in inference-time control.

Abstract

Activation-based steering enables Large Language Models (LLMs) to exhibit targeted behaviors by intervening on intermediate activations without retraining. Despite its widespread use, the mechanistic factors that govern when steering succeeds or fails remain poorly understood, as prior work has relied primarily on black-box outputs or LLM-based judges. In this study, we investigate whether the reliability of steering can be diagnosed using internal model signals. We focus on two information-theoretic measures: the entropy-derived Normalized Branching Factor (NBF), and the Kullback-Leibler (KL) divergence between steered activations and targeted concepts in the vocabulary space. We hypothesize that effective steering corresponds to structured entropy preservation and coherent KL alignment across decoding steps. Building on a reliability study demonstrating high inter-judge agreement between two architecturally distinct LLMs, we use LLM-generated annotations as ground truth and show that these mechanistic signals provide meaningful predictive power for identifying successful steering and estimating failure probability. We further introduce a stronger evaluation baseline for Contrastive Activation Addition (CAA) and Sparse Autoencoder-based steering, the two most widely adopted activation-steering methods.

Figures (9)

• Figure 1: Overall pipeline of the proposed method. (Left) An abstract schematic of the llm, illustrating the extraction of three distinct feature sets (blue, red, and green blocks). (Right) An overview of how the extracted features are utilized within the regression framework.
• Figure 2: unsuccessful steering example. The best performance is achieved $0.06$, corresponding to the Gemma 2–2B model for the London concept using the addition steering function with the sae extraction method. No clear increase in nbf is observed as the steering intensity $\alpha$ increases.
• Figure 3: Successful steering example. The best performance is achieved $0.24$, corresponding to the Gemma 2–2B model for the London concept using the rotational steering function with the sae extraction method. A clear increase in nbf is observed as the steering intensity $\alpha$ increases.
• Figure 4: Maximum probability values extracted from the attention head show a clear drop in confidence (used here as a proxy for semantic consistency and fluency) immediately after the layer where the intervention occurs.
• Figure 5: Unsuccessful steering, as evidenced by the lack of a significant difference in kl divergence between the steered and unsteered representations.