PaperlayerView exact revision on arXiv

Source-linked AI summary

How Far Ahead Do LLMs Plan? Uncovering the Latent Horizon in Chain-of-Thought Reasoning

Liyan Xu, Mo Yu, Fandong Meng, Jie Zhou

arXiv:2602.02103v2cs.LGcs.CL

TL;DR

The relationship between LLMs’ internal representations and verbalized reasoning remains unclear, especially whether they plan globally or transition incrementally. Using Tele-Lens to probe hidden states across diverse tasks, the paper finds a myopic planning horizon and shows that latent CoT signals support uncertainty estimation and CoT bypass.

  • Problem

    The relationship between LLMs’ internal representations and verbalized reasoning tokens remains opaque, including whether hidden states encode global reasoning plans or local transitions.

  • Method

    Tele-Lens uses hidden-state probing across 12 diverse datasets to predict future tokens, final answers, reasoning lengths, and other teleological information.

  • Results

    LLMs exhibit a myopic planning horizon: hidden states primarily support local transitions, while compositional tasks reveal precise answer signals only near CoT completion.

  • Takeaways & Limitations

    A sparse set of pivot positions can represent CoT uncertainty, while planning signals enable automatic CoT bypass with minimal performance degradation.

  • Takeaways & Limitations

    CoT necessity estimation relies on a fixed answer space and is presented primarily as a proof-of-concept.

Abstract

from arXiv · show

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.

1. Introduction

The introduction frames a tension between latent planning in LLM hidden states and the necessity of explicit CoT steps, then investigates the resulting planning horizon and its implications for CoT uncertainty and necessity. Using Tele-Lens probing, the work argues that LLMs are generally myopic planners while showing that CoT dynamics support uncertainty estimation and automatic bypass.

  • Motivation: Prior work presents a tension: hidden states can encode future reasoning early, whereas explicit CoT steps remain indispensable for deriving longer computations.The introduction contrasts evidence for internal planning with theoretical arguments that Transformers require intermediate CoT steps because of limited expressivity.
  • Research Questions: The paper asks whether hidden states encode a global reasoning roadmap or mainly support local incremental transitions, and how this horizon affects CoT.These questions motivate examining the relationship between internal representations and verbalized reasoning tokens.
  • Method: Tele-Lens probes hidden states within CoT steps through a trained low-rank adapter to predict teleological information across multiple dimensions, including subsequent tokens and final answers.The probing experiments evaluate internal planning strength with respect to future reasoning trajectories.
  • Findings: Probing results indicate a myopic planning horizon: hidden states primarily support immediate local transitions rather than precise long-range trajectories.Early states may predict final answers on simpler tasks but behave nearly randomly on compositional tasks, becoming reliable only in the final one or two reasoning steps.
  • Applications: The paper leverages CoT dynamics for uncertainty calibration and automatic CoT bypass, with Qwen3-32B reaching up to 16.2% bypass for only a negligible 0.03 overall accuracy drop.The uncertainty proposal follows a Wooden Barrel principle, while bypass directly outputs answers when CoT is unnecessary.

2. CoT Planning Horizon

Tele-Lens probes hidden states for future-token, final-answer, and reasoning-length information, revealing that LLMs usually plan myopically rather than maintaining precise global foresight. Early signals can capture coarse answer patterns or task-specific shortcuts, but precise answers and reasoning paths generally emerge only locally during CoT.

  • Probing method: Tele-Lens transforms hidden states for full-vocabulary prediction, using a nonlinear bottleneck low-rank adapter to reduce overfitting and computational overhead.The method follows the Logit Lens paradigm by connecting transformed intermediate hidden states to the final language-model head.
  • Final-answer planning: Precise final-answer planning is myopic: on compositional tasks such as Parity, Cycle, and Subsum, the correct answer emerges only near reasoning completion.For Parity, probing exceeds 90% only after the final digit count, while preceding positions remain near random guessing at 50%.
  • Final-answer planning: Early hidden states can show coarse answer signals on semantic and knowledge tasks, but these signals reflect pattern-matching gist rather than precise latent reasoning plans.Early final-answer planning performs worse than both standard CoT reasoning and direct answering across almost all tasks, with the gap widening when CoT is applied.
  • Subsequent-token planning: Subsequent-token Top-5 accuracy exceeds 50% for the next two steps but generally declines with prediction distance, indicating limited foresight over later reasoning.Longer-range planning is mainly visible on structurally modular tasks such as Parity and Cycle; Off-the-Shelf LLMs show the same trend with lower accuracy.
  • Reasoning-length planning: Initial hidden states generally fail to predict total reasoning length reliably, while high correlations on Parity and Subsum arise from input-length shortcuts rather than robust global planning.Cycle lacks this shortcut because its reasoning length depends on the path between vertices rather than the input sequence length.

3. Leveraging CoT Dynamics

The section exploits CoT dynamics to estimate reasoning uncertainty from sparse pivot positions and to identify when full CoT generation is unnecessary. These strategies improve calibration and reduce computation with minimal accuracy loss, while necessity estimation remains a proof of concept due to its fixed answer space.

  • Latent Signals by Tele-Lens: Top-k pivot selection over Tele-Lens signals improves uncertainty estimation by up to 9% absolute over the best full-path baseline, with k = 5 performing best.Averaging final-answer entropy at only five pivot tokens strongly indicates uncertainty across the whole reasoning path.
  • Latent Signals by General Metrics: For Qwen3-32B, top-k selection over general metrics yields 3+% absolute improvement across all metrics at k = 100, reaching up to 6%.The strategy causes no negative impact for the selected k values, even though average CoT length exceeds 7K tokens.
  • Latent Signals by General Metrics: Tele-Lens pivots concentrate near CoT completion, whereas general LM entropy selects positions throughout the trajectory, motivating integration of multiple latent signals.The differing spatial distributions suggest that richer signals and selection strategies could further improve calibration.
  • CoT Necessity Estimation: Early final-answer signals can recognize whether full CoT generation is necessary, enabling automatic bypass on non-essential cases without performance degradation.The approach assesses initial CoT positions to determine whether an input requires extended reasoning.
  • CoT Necessity Estimation: With threshold 0.1, Qwen3-32B reduces thinking by 16.2% on CSQA and 12.4% on MMLU with only 0.03% overall accuracy degradation.The heuristic preserves CoT for tasks such as Parity while bypassing it for easier tasks such as CSQA.

4. Conclusion

The paper investigates LLMs’ internal planning capacity and finds a myopic planning horizon during chain-of-thought generation. Using Tele-Lens probing experiments, it demonstrates how latent signals can support CoT uncertainty and necessity estimation.

  • 4. Conclusion: Tele-Lens probing experiments investigate LLMs’ internal planning capacity and uncover a myopic planning horizon during CoT generation.The method is used to analyze latent planning signals.
  • 4. Conclusion: Latent signals can be exploited for both CoT uncertainty estimation and CoT necessity estimation.The paper demonstrates these applications through its analysis.

A. Tasks and Datasets

The probing experiments cover 12 diverse tasks spanning different task types, with the section providing examples, data-processing details, and statistics.

  • Tasks and Datasets: 12 diverse tasks are included in the probing experiments.The experiments are designed to provide a comprehensive view of empirical insights.
  • Tasks and Datasets: The tasks span different types.
  • Tasks and Datasets: The section provides concrete examples, data-processing details, and statistics.

A.1. Task Examples · A.2. Dataset Descriptions

The paper illustrates tasks spanning sequence counting, graph reachability, subsequence optimization, and diverse multiple-choice reasoning formats. Its experiments cover implicit compositional tasks and knowledge- and semantic-understanding benchmarks, including GSM8K, MATH, AIME, MuSR, CSQA, MMLU, and QuALITY.

  • A.1. Task Examples: Another example asks for the least significant digit of the maximum nonadjacent subsequence sum, reporting sequence length 29, maximum sum 84, and answer 4.The response must contain only the requested least significant digit.
  • A.1. Task Examples: The examples further span GSM8K, MATH, and AIME multiple-choice problems involving arithmetic, geometry, and three-dimensional inequalities.These tasks require selecting and outputting an answer option by letter.
  • A.2. Dataset Descriptions: A.2 groups three mathematical and two logical-reasoning datasets under implicit compositional tasks: GSM8K, MATH, AIME, and MuSR.GSM8K uses middle-school problems, MATH follows the MATH-500 test split, and AIME contributes 30 AIME’25 problems.
  • A.2. Dataset Descriptions: A.2 also includes four knowledge-intensive semantic benchmarks: CSQA, MMLU, QuALITY, and a fourth benchmark described in the dataset list.They target commonsense reasoning, broad subject knowledge, and narrative question answering; QuALITY uses dense-retrieved snippets with a maximum 2K context.

A.3. Data Preparation For Existing Datasets

The study converts three free-form mathematics datasets into multiple-choice tasks for final-answer probing, while shuffling options in existing multiple-choice datasets to reduce memorization and positional bias.

  • Mathematics-task conversion: Three mathematics tasks—MATH, GSM8K, and AIME—are converted from free-form generation into multiple-choice formats using GPT-4.1.This enables probing of final answers within a fixed answer space.
  • Mathematics-task conversion: GPT-4.1 generates four plausible but incorrect answer options as misleading distractors for each converted problem.The prompt provides the problem and correct answer solution, then requests distractors matching the correct answer’s format.
  • Existing multiple-choice datasets: Existing multiple-choice datasets have their answer options shuffled to mitigate LLM memorization effects and positional bias.The shuffling is applied separately to each question.

A.4. Data Generation and Sampling

The study uses controllable generation for three explicit compositional tasks, with task-specific random inputs and labels. Other tasks are sampled from original datasets, while generated-task labels are evenly distributed.

  • Data Generation: Three explicit compositional tasks use a fully controllable process to generate problems and corresponding probing labels.The procedure is specified separately for Parity, Cycle, and Subsum.
  • Data Generation: Cycle samples an even edge count from 4–100, creates either one full cycle or two equal half-sized cycles, and randomizes vertex names and edge orderings.Vertex names are drawn from 1000 candidates to diversify identities and orderings.
  • Data Generation: Subsum uses integer lists with values 1–9 and lengths 2–50, obtaining labels through dynamic programming.The supplied description specifies the input ranges and labeling method.
  • Data Sampling: Generated-task labels are evenly distributed, while other tasks draw test problems from original test sets and supplement training or development sets when necessary.For AIME’25, all 30 problems are assigned to test, with none added to train or dev.

A.5. Dataset Construction and Statistics · A.6. Hyperparameters · B. In-Domain LLM

The study constructs train/dev/test hidden-state datasets from CoT rollouts for both LLM backbones, using denser retention for testing and sampled tokens for training and development. Tele-Lens labels states by future-token, answer, and trajectory properties, with probing trained under specified model-specific dimensions and shared adapter rank.

  • A.5. Dataset Construction and Statistics: Test hidden states retain all CoT tokens up to a maximum length of 16,384.Train/dev retention samples 5% of CoT tokens for Off-the-Shelf LLM and 10% for In-Domain LLM to control storage cost.
  • A.5. Dataset Construction and Statistics: Train/dev/test datasets contain 2.4M / 81K / 11M hidden states for Off-the-Shelf LLM, respectively, for each Transformers layer.
  • B. In-Domain LLM: Train/dev/test datasets contain 2.5M / 57K / 2.7M hidden states for In-Domain LLM, respectively, for each Transformers layer.
  • A.5. Dataset Construction and Statistics: Each hidden state is labeled with rollout outcomes such as the i-th subsequent token ID, final predicted answer token ID, and total CoT length.The dataset supports Tele-Lens adapters in learning latent features that discriminate among different labels.
  • A.6. Hyperparameters: Hidden size is d = 5120 for Off-the-Shelf LLM using Qwen3-32B and d = 3584 for In-Domain LLM trained upon Qwen2.5-7B-Instruct.
  • A.6. Hyperparameters: Both models use probing rank r = 256, learning rate 1 × 10−3, batch size 8000, linear decay, dev-set early stopping, and approximately 5000 maximum training steps.Weight decay and a warmup period are not enabled.

B.1. Training Details

The in-domain LLM is trained with GRPO on Qwen2.5-7B-Instruct across 12 tasks using format and correctness rewards. Training uses 48K problems and converges within 800 steps under specified rollout, optimization, and clipping settings.

  • Training setup: GRPO trains Qwen2.5-7B-Instruct on 12 tasks, rewarding format validation and answer correctness with score 1 each.The training set contains 48K problems sampled from existing datasets and auto-generated explicit-composition problems.
  • Prompt format: The system prompt instructs the model to think before answering and encloses the thinking process within <think>...</think> tags.The prompt frames the model as a helpful assistant solving reasoning problems and separates the final answer from the thinking process.
  • Optimization settings: Training uses rollout size 16, batch size 320, mini batch size 80, maximum response length 4096, and initial learning rate 1 × 10−6.A cosine schedule with 10 warmup steps and clip ratios of 0.3 upper and 0.2 lower is used; convergence occurs within 800 steps, with Parity slowest.

B.2. Evaluation · B.3. In-Domain CoT Examples · C. Probing Results

The paper evaluates its adopted LLM backbones, including an in-domain reinforcement-learning model, across 12 tasks. The in-domain model produces shorter CoT trajectories, suggesting more stable and decisive reasoning, with qualitative examples provided for Parity and Cycle.

  • B.2. Evaluation: 12 tasks are used to evaluate the adopted LLM backbones, including an In-Domain LLM trained by reinforcement learning.Results are provided in Table 5.
  • B.3. In-Domain CoT Examples: The In-Domain LLM produces much shorter CoT trajectories than the Off-the-Shelf LLM, indicating more stable and decisive reasoning behavior.Parity examples compare Qwen3-32B with the In-Domain LLM, while another example covers Cycle.

C.1. Collection of Full Results … Results

The full-results collection documents probing evaluations across layers, tasks, reasoning trajectories, subsequent-token prediction, and global-step estimation. Additional illustrations show myopic planning beyond compositional reasoning, while the results section reports full evaluations for uncertainty estimation and CoT bypass.

  • C.1. Collection of Full Results: Parity examples contrast lengthy off-the-shelf reasoning with a much shorter, predictable in-domain trajectory.The examples use Qwen3-32B for the off-the-shelf model and an in-domain model trained via GRPO.
  • C.1. Collection of Full Results: The full-results collection covers probing final answers, subsequent tokens, and global reasoning steps across in-domain and off-the-shelf LLMs.Listed evaluations include final-answer probing dynamics, subsequent-token Top-5 accuracy up to the eighth following token, and reasoning-length heatmaps.
  • D. Leveraging CoT Dynamics: The collection also reports task-accuracy comparisons for off-the-shelf and in-domain LLMs under thinking, non-thinking, early-final-answer-planning, and random-guess settings.Figures 19 and 20 compare these settings across the two model categories.
  • C.1. Collection of Full Results: Cycle reasoning length tracks the path or cycle between vertices rather than the number of input edges, limiting reliable initial prediction without that heuristic.The in-domain example states that total reasoning length is not reliably predictable at CoT’s initial stage without using the path/cycle-length heuristic.
  • C.2. More Myopic Planning Illustrations: Beyond explicit compositional reasoning, tasks also show a myopic planning horizon, with sparse high-confidence positions emerging near CoT completion in math and logical reasoning.Figure 16 illustrates final-answer planning dynamics on these tasks.
  • Results: The results section includes full uncertainty-estimation evaluations using the top-k pivot-selection strategy and a separate full evaluation of CoT bypass.These evaluations correspond to Tables 6–8 and the methods described in Sections 3.1 and 3.2.

E. Related Works and Discussions

The discussion situates Tele-Lens among probing and chain-of-thought dynamics research, linking its myopic planning horizon to the importance of explicit reasoning for compositional problems. It presents the work as a unified perspective on latent planning and hidden signals in CoT.

  • Planning horizon and CoT: The myopic horizon implies that models must begin dynamic reasoning as state search and exploration, making explicit CoT planning important.The discussion connects this implication to empirical validation by recent works on explicit planning within CoT.
  • Related probing work: Prior probing studies examine truthful responses, world knowledge, and global planning in Transformers’ hidden states, while Tele-Lens provides another approach to understanding LLM internal states.The cited probing literature includes both trained probers and studies of latent planning before CoT generation.
  • CoT dynamics: Prior CoT-dynamics studies analyze entropy, pivot-token identification, information gain, and negative effects of chain-of-thought reasoning.One cited finding is that only about 20% of tokens have high entropy.
  • Why CoT is necessary: Theoretical work explains why CoT is functionally necessary: Transformers cannot perform soft multi-step reasoning in one step, and intermediate steps support length generalization.The discussion states that these limitations make CoT especially indispensable for compositional problems, consistent with this work’s experiments.
  • Contribution and outlook: The work claims to be the first to focus explicitly on latent planning horizons and their utilization, unifying prior studies from complementary perspectives.It also calls for identifying and exploiting additional hidden latent signals to deepen understanding of CoT synergy.
Loading 2602.02103v2…