Paper 2602.02103
How Far Ahead Do LLMs Plan? Uncovering the Latent Horizon in Chain-of-Thought Reasoning
- Published
- Feb 2026
- Research lab
- Independent
- Citations
- 1
- GitHub
- 7 stars
01 In brief
Summary
The paper investigates the latent planning horizon of Large Language Models (LLMs) during Chain-of-Thought (CoT) reasoning.
The authors introduce Tele-Lens, a probing method using low-rank adapters to predict teleological information (subsequent tokens, final answers, reasoning length) from hidden states across 12 diverse tasks.
Empirical results reveal that LLMs exhibit a myopic planning horizon, primarily supporting incremental local transitions rather than precise global planning.
Early hidden states may capture coarse answer signals for simpler tasks, but these are heuristic and inferior to direct answering.
Based on this, the authors propose the Wooden Barrel principle for uncertainty estimation, showing that focusing on a sparse set of pivot positions (top-k) improves calibration by up to 6% absolute AUROC.
They also demonstrate automatic CoT bypass, achieving up to 16.2% reduction in thinking with negligible accuracy loss (0.03%) using Qwen3-32B.
The findings reconcile prior conflicting views on internal planning and highlight the importance of exploiting CoT dynamics for efficiency and calibration.
02 From the paper
Abstract
Chain-of-thought (CoT) reasoning has become a central mechanism for eliciting multi-step reasoning in Large Language Models (LLMs). Yet recent evidence presents a tension: hidden states appear to already encode future reasoning before CoT fully unfolds, while explicit steps still remain crucial for tasks requiring compositional computation. To deepen the understanding between LLM's internal states and its verbalized reasoning trajectories, we investigate the latent planning strength of LLMs, through our probing method, Tele-Lens, applying to hidden states across diverse task domains. Our empirical results indicate that LLMs exhibit a myopic horizon, primarily conducting incremental transitions without precise global planning. Leveraging this characteristic, we propose a hypothesis on enhancing uncertainty estimation of CoT, which we validate that a sparse set of pivot positions can effectively represent the uncertainty of the entire path. We further underscore the significance of exploiting CoT dynamics, and demonstrate that automatic recognition of CoT bypass can be achieved without performance degradation. Our code, data and models are released at https://github.com/lxucs/tele-lens.