mlsanityData to 14 Oct 2026

From Candidate Tokens to Final Outcomes: Decoding Relative Long-Horizon Candidate Value from Reasoning Model Representations

arXiv 2610.15412 ↗First posted 11 Oct 2026cs.CLcited by 0

On the map: LLM reasoning → Test-time compute and efficient thinking also fits: Reasoning benchmarks and evaluation

Abstract (arXiv)

Large reasoning models generate reasoning trajectories token by token. At a given prefix, multiple next-token candidates may be plausible, yet shape subsequent reasoning and differ in their chances of leading to the correct answer. Whether these differences are reflected in internal representations remains unclear. To study this question, we define relative long-horizon candidate value to quantify within-prefix differences and construct the Reasoning Candidate-Token Value (RCTV) dataset, pairing mathematical reasoning prefixes with plausible next-token candidates and their empirical long-horizon candidate values. Using RCTV, we examine hidden states, multilayer perceptron outputs, and attention outputs at candidate positions with sparse linear probes. Experimental results provide evidence that relative long-horizon candidate value is linearly decodable from these representations at multiple layers. Candidate-value probes can also outperform probability- and embedding-based baselines. Furthermore, a controlled online intervention experiment shows that using the decoded signals to guide token selection during generation can improve the chances of eventual success. Our findings provide evidence that internal representations encode information about relative long-horizon candidate value, offering insights into how local candidate-level representations relate to long-horizon reasoning outcomes.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 28 of its 37 references that are in the map. First posted on arXiv 11 Oct 2026.

Builds on

28 of 37 references.

Closest ancestors · 8

Same direction, newest first

  1. 2026Efficient and Trainable Language Model Test-Time Scaling via Local Branch Routing24 Jun
  2. 2026LLM Reasoning as Trajectories: Step-Specific Representation Geometry and Correctness Signals7 Apr
  3. 2026LLMs Encode Their Failures: Predicting Success from Pre-Generation Activations10 Feb
  4. 2026Towards Long-Horizon Interpretability: Efficient and Faithful Multi-Token Attribution for Reasoning LLMs2 Feb
  5. 2026Martingale Foresight Sampling: A Principled Approach to Inference-Time LLM Decoding21 Jan
  6. 2025Think Only When You Need with Large Hybrid-Reasoning Models20 May
  7. 2025Reasoning Models Know When They're Right: Probing Hidden States for Self-Verification7 Apr
  8. 2024Critical Tokens Matter: Token-Level Contrastive Estimation Enhances LLM's Reasoning Capability29 Nov

Other references · 20

  1. 2021Evaluating Large Language Models Trained on Codecited by 8,320
  2. 2023Finding Neurons in a Haystack: Case Studies with Sparse Probingcited by 334
  3. 2025Big-Math: A Large-Scale, High-Quality Math Dataset for Reinforcement Learning in Language Modelscited by 109
  4. 2026Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thoughtcited by 40
  5. 2024Forking Paths in Neural Text Generationcited by 34
  6. 2025The LLM Already Knows: Estimating LLM-Perceived Question Difficulty via Hidden Representationscited by 29
  7. 2024Non-myopic Generation of Language Models for Reasoning and Planningcited by 26
  8. 2025Knowing Before Saying: LLM Representations Encode Information About Chain-of-Thought Success Before Completioncited by 23
  9. 2024The Hidden Space of Transformer Language Adapterscited by 17
  10. 2025Are language models aware of the road not taken? Token-level uncertainty and hidden state dynamicscited by 11
  11. 2025Token Hidden Reward: Steering Exploration-Exploitation in Group Relative Deep Reinforcement Learningcited by 10
  12. 2026Truth as a Trajectory: What Internal Representations Reveal About Large Language Model Reasoningcited by 10
8 more
  1. 2026Hidden Error Awareness in Chain-of-Thought Reasoning: The Signal Is Diagnostic, Not Causalcited by 7
  2. 2025Chain-of-Thought Tokens are Computer Program Variablescited by 6
  3. 2026Your Language Model is Its Own Critic: Reinforcement Learning with Value Estimation from Actor's Internal Statescited by 5
  4. 2026Sparse Reward Subsystem in Large Language Modelscited by 4
  5. 2026The Cancellation Hypothesis in Critic-Free RL: From Outcome Rewards to Token Creditscited by 3
  6. 2026Beyond the Target: From Imitation to Collaboration in Speculative Decodingcited by 2
  7. 2026Unleashing Implicit Rewards: Prefix-Value Learning for Distribution-Level Optimizationcited by 1
  8. 2026FutureBridge: Token Selection Beyond Local Preference in Collaborative Decodingcited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning
  2. 2025Internal states before wait modulate reasoning patterns
  3. 2026The Horizon Gap: Planning, Memory, Execution, Training, and Evaluation for Long-Horizon LLM Agents
  4. 2025Do Latent Tokens Think? A Causal and Adversarial Analysis of Chain-of-Continuous-Thought
  5. 2026Unveiling the Reasoning Process of Large Language Models
  6. 2024Calibrating Reasoning in Language Models with Internal Consistency
  7. 2026Reasoning as State Transition: A Representational Analysis of Reasoning Evolution in Large Language Models
  8. 2026The Tell-Tale Norm: ℓ2 Magnitude as a Signal for Reasoning Dynamics in Large Language Models