From Latent Signals to Reflection Behavior: Tracing Meta-Cognitive Activation Trajectory in R1-Style LLMs

TL;DR

This work investigates how reflection emerges inside R1-style LLMs by tracing activation signals across depth with a logit-lens. It identifies a depth-wise, three-stage causal trajectory: latent-control layers projecting a thinking-budget direction $d^{(l)}_{ ext{pos}}$, semantic-pivot layers where turning-point and summarization cues compete for probability mass, and behavior-overt layers where explicit reflection markers are sampled. The authors validate a causal chain via prompt-level and activation-steering interventions that propagate stage-by-stage, showing that early-level semantics modulate later discourse cues and ultimately influence reflection-token sampling. The findings support a human-like meta-cognitive progression and generalize across models, markers, and domains, providing a foundation for controllable reflection and interpretable generation. The work contributes reusable protocols and metrics for tracking internal signals, with potential applications in prediction and safety-aware generation.

Abstract

R1-style LLMs have attracted growing attention for their capacity for self-reflection, yet the internal mechanisms underlying such behavior remain unclear. To bridge this gap, we anchor on the onset of reflection behavior and trace its layer-wise activation trajectory. Using the logit lens to read out token-level semantics, we uncover a structured progression: (i) Latent-control layers, where an approximate linear direction encodes the semantics of thinking budget; (ii) Semantic-pivot layers, where discourse-level cues, including turning-point and summarization cues, surface and dominate the probability mass; and (iii) Behavior-overt layers, where the likelihood of reflection-behavior tokens begins to rise until they become highly likely to be sampled. Moreover, our targeted interventions uncover a causal chain across these stages: prompt-level semantics modulate the projection of activations along latent-control directions, thereby inducing competition between turning-point and summarization cues in semantic-pivot layers, which in turn regulates the sampling likelihood of reflection-behavior tokens in behavior-overt layers. Collectively, our findings suggest a human-like meta-cognitive process-progressing from latent monitoring, to discourse-level regulation, and to finally overt self-reflection. Our analysis code can be found at https://github.com/DYR1/S3-CoT.

Figures (26)

• Figure 1: Depth-wise causal chain of reflection. Prompt-level cues modulate latent-control activations toward a deep or quick-thinking direction, shift probability mass between turning-point and summarization cues in semantic-pivot layers, and ultimately change the sampling likelihood of reflection markers in behavior-overt layers.
• Figure 2: Analysis on DeepSeek-R1$_{7B}$. We report analyses of activation-difference and forward-activation, including t-SNE visualizations, logit-lens decoding results, and our predefined sets. The layer-wise trajectories highlight that different layers play distinct functional roles.
• Figure 3: Analysis on Qwen3-Think$_{4B}$. We report analyses of activation-difference and forward-activation, including t-SNE visualizations, logit-lens decoding results, and our predefined sets. The layer-wise trajectories highlight that different layers play distinct functional roles.
• Figure 4: Changes of projection strength in latent-control layers under prompt-level interventions.
• Figure 5: Effect on pivot-semantic layers under prompt-level interventions.