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

AI summary

Overview

Research area: Interpretability and reasoning analysis of Large Language Models, specifically the internal planning behavior that underlies Chain-of-Thought (CoT) generation.

Technical level: Intermediate. The paper uses probing methods, low-rank adapters, AUROC-based uncertainty evaluation, and reinforcement learning (GRPO) training, but the framing and findings are accessible to readers familiar with basic LLM reasoning concepts.

Scope: Across 12 datasets and two LLM backbones, the paper measures how far ahead LLM hidden states encode future reasoning content, and then applies those measurements to uncertainty estimation and deciding when CoT can be skipped.

What This Paper Is About

Prior work contains a tension: some studies find that early hidden states already encode future reasoning and the final answer, while other work argues CoT steps are indispensable because Transformers cannot perform compositional computation in a single forward pass. This paper asks how much of the reasoning roadmap is actually present in hidden states before it is verbalized, and whether that latent structure can be used practically. The authors build a probing method called Tele-Lens and test it across many task domains to characterize the "latent horizon" of planning.

Key Contributions

1. Tele-Lens probing method. A transformation-based prober in the spirit of Logit Lens, using a bottleneck low-rank adapter with added nonlinearity (rank r = 256, LM head kept frozen). It probes three "teleological" dimensions of hidden states: subsequent tokens (up to the 8th following token, with the target offset injected via an optional embedding), total reasoning length (via a regression layer), and the final answer (restricted to a fixed answer space).

2. A broad empirical study of planning horizon. Probing is conducted across 12 diverse tasks grouped into explicit compositional tasks (Parity, Cycle, Subsum), implicit compositional tasks (GSM8K, MATH, AIME, MuSR, Zebra), and knowledge/semantic tasks (CSQA, MMLU, QuALITY, GPQA), using both an off-the-shelf Qwen3 series model (Qwen3-32B as primary, Qwen3-8B in some experiments) and an In-Domain LLM trained with GRPO from Qwen2.5-7B-Instruct.

3. The "Wooden Barrel" hypothesis for CoT uncertainty. The claim that a reasoning chain's uncertainty is governed by a small number of pivot positions rather than a global average over all tokens. The authors validate that selecting top-k positions improves uncertainty calibration for both Tele-Lens signals and general metrics.

4. A proof-of-concept for CoT bypass. Using the early final-answer gist revealed by probing, the method recognizes inputs where CoT is unnecessary and outputs the answer directly with negligible performance loss.

Main Findings

• LLMs exhibit a myopic planning horizon. For precise final-answer planning, the answer only emerges immediately before reasoning completes. In the Parity probe (Table 1), final-answer probability at the last four counting positions (-4 to -1) and at position 0 was 0.49 / 0.51 / 0.51 / 0.97 / 0.99 for the In-Domain LLM, and 0.50 / 0.52 / 0.51 / 0.94 / 0.97 for the Off-the-Shelf LLM, with random guessing at 0.5. The example figures use 50% as the random-guessing level for two-choice answers.

• Early answer signals exist but are only a coarse gist. For semantic and knowledge-heavy tasks such as CSQA and MMLU, probing accuracy spikes early in the CoT. However, across almost all tasks this early planning signal yields lower task accuracy than both reasoning with CoT and direct answering without CoT, indicating it is not the product of a precise pre-planned reasoning path.

• Foresight over subsequent tokens is limited. In-Domain LLM Top-5 accuracy for subsequent-token prediction exceeds 50% for the next two steps but declines as position advances, with the sharpest declines on semantic/knowledge tasks (MMLU, GPQA). Longer-range predictability appears mainly on structurally patterned tasks such as Parity or Cycle. The Off-the-Shelf LLM shows a similar trend with much lower accuracy overall.

• Reasoning length is not known in advance, except through shortcuts. Predictions of total reasoning length from initial hidden states show unstable and often low correlations across most tasks. Parity and Subsum look like exceptions, but the paper attributes this to a task-specific confound: reasoning length is typically proportional to input sequence length in those tasks. Cycle does not show this pattern because its reasoning length scales with the path between two vertices rather than input length.

• Pivot-based uncertainty estimation improves calibration. With In-Domain LLM latent signals from Tele-Lens (last layer), top-5 selection reached average AUROC 0.69 versus 0.60 for the best baseline (Self-Certainty), with per-task values of 0.87 (GSM8K), 0.77 (Zebra), 0.73 (MMLU), 0.56 (GPQA). The paper reports this as up to 9% absolute improvement over the best baseline. Top-10 (0.68), Top-20 (0.63), and Top-50 (0.64) were weaker than Top-5.

• General metrics also benefit from top-k pivots. For Qwen3-32B, selecting k = 100 positions produced average AUROC gains across all three metrics: Perplexity 0.72 to 0.78, Entropy 0.72 to 0.75, and Self-Certainty 0.65 to 0.70. The paper states this is a consistent 3+% absolute improvement, reaching up to 6%. Average CoT length across inputs here exceeds 7K tokens.

• CoT bypass works with minimal accuracy cost. With a normalized-entropy threshold of 0.1, Qwen3-32B bypassed CoT on 16.2% of CSQA, 12.4% of MMLU, 1.2% of GPQA, and 0% of Parity, averaging 2.8% bypass with a 0.03 overall accuracy change. At threshold 0.2, bypass averaged 6.2% with a 0.37 accuracy change. The In-Domain LLM at threshold 0.1 averaged 13.3% bypass (40.2% CSQA, 30.4% MMLU, 7% GPQA, 0% Parity) with a 0.47 accuracy change; at threshold 0.2 it averaged 21.6% bypass with a 1.42 accuracy change.

• Pivot positions from different signal sources differ. Tele-Lens-selected pivots tend to concentrate near CoT completion, whereas positions selected by general LM entropy are distributed across the whole trajectory, suggesting the two signals may be complementary.

• Backbone behavior. Direct answering without CoT achieves near-random performance on compositional tasks such as Parity and Cycle. The In-Domain LLM underperforms Qwen3 models on some datasets but achieves the best performance on three compositional tasks with roughly 1K+ characters per CoT versus 10K+ for Qwen3. The best probing layer was not the final layer but an intermediate one: layer 48 of 64 for Off-the-Shelf Qwen3 and layer 21 of 28 for the In-Domain LLM.

Methodology in Plain English

The authors attach a small trainable adapter to each Transformer layer of a language model. This adapter transforms the hidden state at each token of a reasoning trace into a prediction over the model's full vocabulary (with the output head frozen), so the researchers can ask: from this internal state alone, what does the model seem to "already know"?

Three questions are asked at every point along the trace: what are the next tokens going to be, how long will the whole reasoning be, and what is the final answer. A separate adapter is trained per layer per probing dimension, for about 5K steps with early stopping on a dev set, using up to 4000 / 100 / 500 train / dev / test problems per task. The final-answer label space is 20 tokens because all 12 tasks were made to have fixed answer choices, with math and logic problems converted to multiple-choice using GPT-4.1 to generate distractor options.

To make the results more interpretable, the authors use two backbones. The off-the-shelf Qwen3-32B represents a widely used model with native thinking mode. The In-Domain LLM, trained with GRPO from Qwen2.5-7B-Instruct (a model without native thinking mode), is meant to provide an "upper bound" on internal planning since it reasons more decisively and with much shorter traces.

For the uncertainty work, the authors start from standard metrics (perplexity/NLL, average entropy, self-certainty) that average over every token, then replace the whole-sequence average with an average over only the most uncertain positions (the "shortest staves"). For the CoT bypass work, they generate the first five CoT tokens, compute the normalized entropy over the 20 probing classes, and if any initial position falls below a threshold, they stop generation and answer directly in non-thinking mode.

Why This Matters

Impact on research: The paper offers a unified reading of previously conflicting evidence. Hidden states do carry answer-related information for easier tasks, but this is a coarse, pattern-matching gist rather than a genuine multi-step plan, which reconciles early-planning observations with the theoretical argument that CoT is essential for compositional computation.

Real-world applications:

• Adaptive thinking systems, like those that route requests to models of different sizes or allocate different reasoning effort, need a way to sense input complexity from internal signals.
• CoT bypass or early exiting reduces the "overthinking" cost, useful for latency- and compute-sensitive deployment.
• Uncertainty calibration for reasoning outputs supports confidence-aware decisions where wrong answers are costly.
• CoT compression and length steering can trim verbose traces while preserving accuracy.

Industry relevance: Both mechanisms the paper targets, uncertainty estimation and deciding when reasoning is needed, map directly to inference cost and reliability. The reported 16.2% / 12.4% CoT reduction on CSQA / MMLU at a 0.03 accuracy cost is a concrete efficiency claim of the kind that matters for serving large reasoning models. The code, data, and models are released at https://github.com/lxucs/tele-lens.

Future Directions

1. Integrating multiple latent signal sources. Because Tele-Lens pivots concentrate near CoT completion while entropy-based pivots spread across the trace, combining them could identify critical positions more robustly and improve uncertainty calibration.

2. Going beyond a fixed answer space. Both the final-answer probing and the necessity estimation depend on a fixed answer space (20 tokens). The authors explicitly present both analyses as limited by this constraint; the CoT bypass study is framed as a proof-of-concept.

3. Unlocking the coarse gist for training and steering. The paper suggests the latent signals could support model training and CoT compression/steering, but does not demonstrate these applications.

4. Understanding why intermediate layers predict best and why semantic tasks show earlier spikes. The paper observes that the best probing layer is intermediate rather than final, and that early answer gist is strongest on semantic tasks, but leaves the underlying mechanism open.

Target Audience

This paper suits researchers and engineers working on LLM interpretability, reasoning efficiency, and inference-time compute allocation. It is also relevant to practitioners building adaptive-thinking or early-exit systems who need to know how much internal planning they can actually rely on. Readers need some familiarity with Chain-of-Thought, hidden states, and uncertainty metrics such as AUROC, perplexity, and entropy to follow the experimental sections.

Authors’ 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.