Unveiling the Latent Directions of Reflection in Large Language Models

Reflection, the ability of large language models (LLMs) to evaluate and revise their own reasoning, has been widely used to improve performance on complex reasoning tasks. Yet, most prior works emphasizes designing reflective prompting strategies or reinforcement learning objectives, leaving the inner mechanisms of reflection underexplored. In this paper, we investigate reflection through the lens of latent directions in model activations. We propose a methodology based on activation steering to characterize how instructions with different reflective intentions: no reflection, intrinsic reflection, and triggered reflection. By constructing steering vectors between these reflection levels, we demonstrate that (1) new reflection-inducing instructions can be systematically identified, (2) reflective behavior can be directly enhanced or suppressed through activation interventions, and (3) suppressing reflection is considerably easier than stimulating it. Experiments on GSM8k-adv and Cruxeval-o-adv with Qwen2.5-3B and Gemma3-4B-IT reveal clear stratification across reflection levels, and steering interventions confirm the controllability of reflection. Our findings highlight both opportunities (e.g., reflection-enhancing defenses) and risks (e.g., adversarial inhibition of reflection in jailbreak attacks). This work opens a path toward mechanistic understanding of reflective reasoning in LLMs.

Key Contributions

• Methodology using activation steering vectors to characterize three reflection levels (no reflection, intrinsic, triggered) in LLM activations
• Empirical finding that suppressing reflection is considerably easier than stimulating it, creating an asymmetric adversarial risk
• Demonstration that reflective behavior can be directly controlled via activation interventions, with clear stratification across reflection levels on GSM8k-adv and Cruxeval-o-adv

🛡️ Threat Analysis

Details

Similar Papers