Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Internal states before wait modulate reasoning patterns

Dmitrii Troitskii, Koyena Pal, Chris Wendler, Callum Stuart McDougall, Neel Nanda

TL;DR

The paper investigates how internal latent states immediately before wait tokens influence reasoning patterns in a distilled reasoning model. Using crosscoders to extract sparse latent directions and a latent-attribution patching framework, it identifies features that both promote and suppress wait-token predictions and demonstrates their causal impact via steering. The results reveal that many reasoning-related features reside outside the most prominent tops, shaping behaviors such as backtracking, restarting, and uncertainty, and that targeted interventions can modulate reasoning trajectories. This work advances mechanistic interpretability of reasoning chains and offers a framework for controlled manipulation of reasoning behavior, with publicly available data and code.

Abstract

Prior work has shown that a significant driver of performance in reasoning models is their ability to reason and self-correct. A distinctive marker in these reasoning traces is the token wait, which often signals reasoning behavior such as backtracking. Despite being such a complex behavior, little is understood of exactly why models do or do not decide to reason in this particular manner, which limits our understanding of what makes a reasoning model so effective. In this work, we address the question whether model's latents preceding wait tokens contain relevant information for modulating the subsequent reasoning process. We train crosscoders at multiple layers of DeepSeek-R1-Distill-Llama-8B and its base version, and introduce a latent attribution technique in the crosscoder setting. We locate a small set of features relevant for promoting/suppressing wait tokens' probabilities. Finally, through a targeted series of experiments analyzing max activating examples and causal interventions, we show that many of our identified features indeed are relevant for the reasoning process and give rise to different types of reasoning patterns such as restarting from the beginning, recalling prior knowledge, expressing uncertainty, and double-checking.

Internal states before wait modulate reasoning patterns

TL;DR

The paper investigates how internal latent states immediately before wait tokens influence reasoning patterns in a distilled reasoning model. Using crosscoders to extract sparse latent directions and a latent-attribution patching framework, it identifies features that both promote and suppress wait-token predictions and demonstrates their causal impact via steering. The results reveal that many reasoning-related features reside outside the most prominent tops, shaping behaviors such as backtracking, restarting, and uncertainty, and that targeted interventions can modulate reasoning trajectories. This work advances mechanistic interpretability of reasoning chains and offers a framework for controlled manipulation of reasoning behavior, with publicly available data and code.

Abstract

Prior work has shown that a significant driver of performance in reasoning models is their ability to reason and self-correct. A distinctive marker in these reasoning traces is the token wait, which often signals reasoning behavior such as backtracking. Despite being such a complex behavior, little is understood of exactly why models do or do not decide to reason in this particular manner, which limits our understanding of what makes a reasoning model so effective. In this work, we address the question whether model's latents preceding wait tokens contain relevant information for modulating the subsequent reasoning process. We train crosscoders at multiple layers of DeepSeek-R1-Distill-Llama-8B and its base version, and introduce a latent attribution technique in the crosscoder setting. We locate a small set of features relevant for promoting/suppressing wait tokens' probabilities. Finally, through a targeted series of experiments analyzing max activating examples and causal interventions, we show that many of our identified features indeed are relevant for the reasoning process and give rise to different types of reasoning patterns such as restarting from the beginning, recalling prior knowledge, expressing uncertainty, and double-checking.

Paper Structure

This paper contains 13 sections, 7 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Sparse Crosscoders
  4. Data: Reasoning Instances with wait
  5. Crosscoder Latent Attribution
  6. Steering
  7. Results
  8. Related Work
  9. Conclusion
  10. Notations
  11. Crosscoder Training
  12. Patchscope Experiments
  13. LLM Judge Prompt Template

Figures (5)

  • Figure 1: We determine reasoning related crosscoder features via latent attribution with respect to wait tokens' logits.
  • Figure 2: We compute how many characters occur before the first wait token in the continued rollout, while steering with all of our features and different intervention strengths. Steering with positive coefficient for the top features slightly increases the distance to the first wait which is due to oversteering, for bottom features as one would expect the distance to the first wait increases significantly. Negative steering has the opposite effect.
  • Figure 3: Change in reasoning behavior observed when features are steered in the positive and/or negative directions.
  • Figure 4: Crosscoder classification of the features into base, shared, reasoning-finetuned categories, with distribution attributed using the crosscoder relative norm difference
  • Figure 5: Top token probabilities from patchscope: top features promote 'Wait' and related tokens, while bottom features show no clear pattern.