ReflCtrl: Controlling LLM Reflection via Representation Engineering

TL;DR

ReflCtrl introduces a representation engineering approach to control self-reflection in large language models, reducing inference costs by selectively managing reflection steps without sacrificing reasoning performance.

Contribution

The paper presents ReflCtrl, a novel framework that identifies and steers reflection behavior in LLMs through latent space manipulation, enabling cost-effective reasoning.

Findings

Reflection steps are often redundant in strong models

Up to 33.6% of reasoning tokens can be saved

Reflection behavior correlates with internal uncertainty signals

Abstract

Large language models (LLMs) with Chain-of-Thought (CoT) reasoning have achieved strong performance across diverse tasks, including mathematics, coding, and general reasoning. A distinctive ability of these reasoning models is self-reflection: the ability to review and revise previous reasoning steps. While self-reflection enhances reasoning performance, it also increases inference cost. In this work, we study self-reflection through the lens of representation engineering. We segment the model's reasoning into steps, identify the steps corresponding to reflection, and extract a reflection direction in the latent space that governs this behavior. Using this direction, we propose a stepwise steering method that can control reflection frequency. We call our framework ReflCtrl. Our experiments show that (1) in many cases reflections are redundant, especially in stronger models (in our experiments, we can save up to 33.6 percent of reasoning tokens while preserving performance), and (2) the model's reflection behavior is highly correlated with an internal uncertainty signal, implying self-reflection may be controlled by the model's uncertainty.