Mehdi Jafari

Mehdi Jafari, Hao Xue, Flora Salim

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.

Quoc Tuan Pham, Mehdi Jafari, Flora Salim

Test-time compute is central to large reasoning models, yet analysing their reasoning behaviour through generated text is increasingly impractical and unreliable. Response length is often used as a brute proxy for reasoning effort, but this metric fails to capture the dynamics and effectiveness of the Chain of Thoughts (CoT) or the generated tokens. We propose Recurrence Quantification Analysis (RQA) as a non-textual alternative for analysing model's reasoning chains at test time. By treating token generation as a dynamical system, we extract hidden embeddings at each generation step and apply RQA to the resulting trajectories. RQA metrics, including Determinism and Laminarity, quantify patterns of repetition and stalling in the model's latent representations. Analysing 3,600 generation traces from DeepSeek-R1-Distill, we show that RQA captures signals not reflected by response length, but also substantially improves prediction of task complexity by 8\%. These results help establish RQA as a principled tool for studying the latent token generation dynamics of test-time scaling in reasoning models.