LLM Assertiveness can be Mechanistically Decomposed into Emotional and Logical Components

TL;DR

This paper uncovers that LLM assertiveness is composed of emotional and logical components, providing mechanistic insights into overconfidence and suggesting ways to mitigate it.

Contribution

It introduces a mechanistic interpretability approach to decompose LLM assertiveness into emotional and logical parts, revealing their distinct roles and causal effects.

Findings

Assertiveness decomposes into emotional and logical components.

Different components have distinct causal influences on model predictions.

Identifies layers most sensitive to assertiveness contrasts.

Abstract

Large Language Models (LLMs) often display overconfidence, presenting information with unwarranted certainty in high-stakes contexts. We investigate the internal basis of this behavior via mechanistic interpretability. Using open-sourced Llama 3.2 models fine-tuned on human annotated assertiveness datasets, we extract residual activations across all layers, and compute similarity metrics to localize assertive representations. Our analysis identifies layers most sensitive to assertiveness contrasts and reveals that high-assertive representations decompose into two orthogonal sub-components of emotional and logical clusters-paralleling the dual-route Elaboration Likelihood Model in Psychology. Steering vectors derived from these sub-components show distinct causal effects: emotional vectors broadly influence prediction accuracy, while logical vectors exert more localized effects. These findings provide mechanistic evidence for the multi-component structure of LLM assertiveness and highlight avenues for mitigating overconfident behavior.