mlsanityData to 14 Oct 2026

Lost at the Beginning of Reasoning

arXiv 2506.22058 ↗First posted 27 Jun 2025cs.CLcited by 15

On the map: LLM reasoning → Test-time compute and efficient thinking

Abstract (arXiv)

Recent advancements in large language models (LLMs) have significantly advanced complex reasoning capabilities, particularly through extended chain-of-thought (CoT) reasoning that incorporates mechanisms such as backtracking, self-reflection, and self-correction. Despite these developments, the self-correction abilities of LLMs during long CoT reasoning remain underexplored. And recent findings on overthinking suggest that such models often engage in unnecessarily redundant reasoning. In this work, we empirically show that the first reasoning step exerts a disproportionately large influence on the final prediction. I.e., errors introduced at this stage can substantially degrade subsequent reasoning quality. This phenomenon is consistently observed across various state-of-the-art open- and closed-source reasoning models. Leveraging this insight, we propose an efficient sampling strategy that leverages a reward model to identify and retain high-quality first reasoning steps while discarding suboptimal ones, achieving up to a 70% reduction in inference cost without sacrificing any accuracy. Our work highlights the central role of the first reasoning step in generating a high-quality reasoning trajectory, and thus enabling significantly efficient sampling.

Key facts. Cited by 15 papers in the map (data through 14 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 54 of its 71 references that are in the map. First posted on arXiv 27 Jun 2025.

Builds on

54 of 71 references.

Closest ancestors · 14

Same direction, newest first

  1. 2025Dynamic Early Exit in Reasoning Models22 Apr
  2. 2025ThinkPrune: Pruning Long Chain-of-Thought of LLMs via Reinforcement Learning2 Apr
  3. 2025s1: Simple test-time scaling31 Jan
  4. 2025Reward-Guided Speculative Decoding for Efficient LLM Reasoning31 Jan
  5. 2025Thoughts Are All Over the Place: On the Underthinking of o1-Like LLMs30 Jan
  6. 2025O1-Pruner: Length-Harmonizing Fine-Tuning for O1-Like Reasoning Pruning22 Jan
  7. 2024Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs30 Dec
  8. 2024LLM Self-Correction with DeCRIM: Decompose, Critique, and Refine for Enhanced Following of Instructions with Multiple Constraints9 Oct
  9. 2024Training Language Models to Self-Correct via Reinforcement Learning19 Sep
  10. 2024Reasoning Aware Self-Consistency: Leveraging Reasoning Paths for Efficient LLM Sampling30 Aug
  11. 2024Escape Sky-high Cost: Early-stopping Self-Consistency for Multi-step Reasoning19 Jan
  12. 2023Large Language Models Cannot Self-Correct Reasoning Yet3 Oct
2 more
  1. 2023Automatically Correcting Large Language Models: Surveying the landscape of diverse self-correction strategies6 Aug
  2. 2023Let's Sample Step by Step: Adaptive-Consistency for Efficient Reasoning and Coding with LLMs19 May

Other references · 40

  1. 2020Language Models are Few-Shot Learnerscited by 32,419
  2. 2023GPT-4 Technical Reportcited by 18,396
  3. 2019Sentence-BERT: Sentence Embeddings using Siamese BERT-Networkscited by 9,170
  4. 2025Qwen3 Technical Reportcited by 8,820
  5. 2020Scaling Laws for Neural Language Modelscited by 6,586
  6. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,330
  7. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,944
  8. 2022Self-Consistency Improves Chain of Thought Reasoning in Language Modelscited by 5,708
  9. 2024GPT-4o System Cardcited by 4,588
  10. 2023Lost in the Middle: How Language Models Use Long Contextscited by 3,455
  11. 2024Gemini 1.5: Unlocking multimodal understanding across millions of tokens of contextcited by 3,224
  12. 2022Training Compute-Optimal Large Language Modelscited by 3,019
28 more
  1. 2023GPQA: A Graduate-Level Google-Proof Q&A Benchmarkcited by 2,879
  2. 2023Efficient Streaming Language Models with Attention Sinkscited by 2,186
  3. 2024LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Codecited by 2,004
  4. 2020MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformerscited by 1,463
  5. 2024Qwen2.5-Math Technical Report: Toward Mathematical Expert Model via Self-Improvementcited by 1,090
  6. 2023Large Language Models are not Fair Evaluatorscited by 1,085
  7. 2023Teaching Large Language Models to Self-Debugcited by 988
  8. 2017Deep Learning Scaling is Predictable, Empiricallycited by 582
  9. 2023Large Language Models Are Not Robust Multiple Choice Selectorscited by 486
  10. 2025The Lessons of Developing Process Reward Models in Mathematical Reasoningcited by 415
  11. 2025MathArena: Evaluating LLMs on Uncontaminated Math Competitionscited by 297
  12. 2021SPLADE: Sparse Lexical and Expansion Model for First Stage Rankingcited by 286
  13. 2023Large Language Models Sensitivity to The Order of Options in Multiple-Choice Questionscited by 263
  14. 2024OpenCodeInterpreter: Integrating Code Generation with Execution and Refinementcited by 247
  15. 2025Light-R1: Curriculum SFT, DPO and RL for Long COT from Scratch and Beyondcited by 175
  16. 2024Recursive Introspection: Teaching Language Model Agents How to Self-Improvecited by 165
  17. 2024NATURAL PLAN: Benchmarking LLMs on Natural Language Planningcited by 128
  18. 2025Why do LLMs attend to the first token?cited by 127
  19. 2022From Distillation to Hard Negative Sampling: Making Sparse Neural IR Models More Effectivecited by 98
  20. 2024Found in the Middle: How Language Models Use Long Contexts Better via Plug-and-Play Positional Encodingcited by 72
  21. 2023Positional Description Matters for Transformers Arithmeticcited by 56
  22. 2024Over-Reasoning and Redundant Calculation of Large Language Modelscited by 27
  23. 2022The NLP Task Effectiveness of Long-Range Transformerscited by 22
  24. 2024The SIFo Benchmark: Investigating the Sequential Instruction Following Ability of Large Language Modelscited by 18
  25. 2023Instruction Position Matters in Sequence Generation with Large Language Modelscited by 10
  26. 20243-in-1: 2D Rotary Adaptation for Efficient Finetuning, Efficient Batching and Composabilitycited by 10
  27. 2025Fractured Chain-of-Thought Reasoningcited by 7
  28. 2024Verbosity neq Veracity: Demystify Verbosity Compensation Behavior of Large Language Modelscited by 5

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 5; 2026: 10.

15 papers.

Sort:

By area:

  1. 2025Toward Effective Tool-Integrated Reasoning via Self-Evolved Preference Learningcited by 17
  2. 2025Let it Calm: Exploratory Annealed Decoding for Verifiable Reinforcement Learningcited by 10
  3. 2026Reward Modeling for Reinforcement Learning-Based LLM Reasoning: Design, Challenges, and Evaluationcited by 7
  4. 2025Humans Perceive Wrong Narratives from AI Reasoning Textscited by 6
  5. 2026Unlocking Reasoning Capability on Machine Translation in Large Language Modelscited by 6
  6. 2026ReasoningFlow: Discourse Structures for Understanding LLM Reasoning Tracescited by 5
  7. 2025Explore Briefly, Then Decide: Mitigating LLM Overthinking via Cumulative Entropy Regulationcited by 4
  8. 2026ET-Agent: Incentivizing Effective Tool-Integrated Reasoning Agent via Behavior Calibrationcited by 4
  9. 2025Round-trip Reinforcement Learning: Self-Consistent Training for Better Chemical LLMscited by 2
  10. 2026TeleResilienceBench: Quantifying Resilience for LLM Reasoning in Telecommunicationscited by 2
  11. 2026From Correctness to Utility: Gain-Based Prefix Evaluation for LLM Reasoningcited by 2
  12. 2026The Tell-Tale Trace: Detecting Reasoning Failures in LLMs Using Chain-of-Thought Dynamicscited by 1
3 more
  1. 2026Cut Your Losses! Learning to Prune Paths Early for Efficient Parallel Reasoningcited by 0
  2. 2026Know When To Fold 'Em: Token-Efficient LLM Synthetic Data Generation via Multi-Stage In-Flight Rejectioncited by 0
  3. 2026Consilience for Verifier-Free Test-Time Scalingcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2024Learning to Check: Unleashing Potentials for Self-Correction in Large Language Models
  2. 2025Early Stopping Chain-of-thoughts in Large Language Models
  3. 2025Demystifying Long Chain-of-Thought Reasoning in LLMs
  4. 2026Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation
  5. 2026Correct, Concise and Complete: Multi-stage Training For Adaptive Reasoning
  6. 2025Self-Training Elicits Concise Reasoning in Large Language Models
  7. 2024On the Hardness of Faithful Chain-of-Thought Reasoning in Large Language Models
  8. 2025Large Reasoning Models are not thinking straight: on the unreliability of thinking trajectories