Research map / Large language models / Reasoning and chain of thought

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

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

placeholder

Problem

Whether LLM hidden states encode a global plan for the reasoning trajectory before explicit chain-of-thought unfolds, and how this planning horizon relates to CoT uncertainty and necessity.

Approach

Tele-Lens trains a bottleneck low-rank adapter on each layer's hidden states (frozen LM head) to probe future subsequent tokens, final answer, and reasoning length across full CoT traces over 12 tasks. It finds a myopic horizon and proposes a "Wooden Barrel" hypothesis that a few low-entropy-selected pivot positions represent the uncertainty of the whole chain.

Key results

• Top-k (k=5) pivot selection on Tele-Lens entropy reaches average AUROC 0.69 vs 0.60 for best baseline (Self-Certainty) with In-Domain LLM.
• With Qwen3-32B, selecting 100 pivots raises average AUROC for perplexity from 0.72 to 0.78 and self-certainty from 0.65 to 0.70.
• CoT bypass with threshold 0.1 on Qwen3-32B bypasses 16.2% of CSQA and 12.4% of MMLU with -0.03 accuracy change.

Surprising findings

• Early hidden states can predict the final answer for knowledge tasks, but such early signals underperform direct answering without CoT.
• Apparent reasoning-length predictability on Parity and Subsum is largely explained by input sequence length shortcuts.

Limitations

• Final-answer probing requires a fixed answer space, so necessity estimation is only a proof-of-concept.
• Off-the-shelf probing relies on 5% sampled train tokens and a single backbone family (Qwen).

Experimental setup

Datasets: GSM8K MATH AIME MuSR Zebra CSQA MMLU QuALITY GPQA Parity Cycle Subsum

Benchmarks: GSM8K MATH AIME MuSR Zebra CSQA MMLU QuALITY GPQA Parity Cycle Subsum

Metrics: accuracy AUROC top-5 accuracy correlation of predicted reasoning length bypass ratio

Baselines: Perplexity Entropy Self-Certainty Qwen3-32B (w/o CoT) Qwen3-8B

Builds on: Logit Lens GRPO DAPO Self-Certainty

Closest papers by meaning

• Eliciting Chain-of-Thought in Base LLMs via Gradient-Based Representation Optimization (AAAI 2026 · similarity 0.92)
• Demystifying Long Chain-of-Thought Reasoning (ICML 2025 · similarity 0.91)
• Deep Hidden Cognition Facilitates Reliable Chain-of-Thought Reasoning (AAAI 2026 · similarity 0.91)
• When More is Less: Understanding Chain-of-Thought Length in LLMs (ICLR 2026 · similarity 0.91)
• Understanding Chain-of-Thought in LLMs through Information Theory (ICML 2025 · similarity 0.90)
• Long Chain-of-Thought Reasoning Across Languages (ICLR 2026 · similarity 0.90)