mlsanityData to 14 Oct 2026

Does Thinking More always Help? Mirage of Test-Time Scaling in Reasoning Models

arXiv 2506.04210 ↗First posted 4 Jun 2025cs.AI cs.CLcited by 53

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

Abstract (arXiv)

Recent trends in test-time scaling for reasoning models (e.g., OpenAI o1, DeepSeek R1) have led to a popular belief that extending thinking traces using prompts like "Wait" or "Let me rethink" can improve performance. This raises a natural question: Does thinking more at test-time truly lead to better reasoning? To answer this question, we perform a detailed empirical study across models and benchmarks, which reveals a consistent pattern of initial performance improvements from additional thinking followed by a decline, due to "overthinking". To understand this non-monotonic trend, we consider a simple probabilistic model, which reveals that additional thinking increases output variance-creating an illusion of improved reasoning while ultimately undermining precision. Thus, observed gains from "more thinking" are not true indicators of improved reasoning, but artifacts stemming from the connection between model uncertainty and evaluation metric. This suggests that test-time scaling through extended thinking is not an effective way to utilize the inference thinking budget. Recognizing these limitations, we introduce an alternative test-time scaling approach, parallel thinking, inspired by Best-of-N sampling. Our method generates multiple independent reasoning paths within the same inference budget and selects the most consistent response via majority vote, achieving up to 20% higher accuracy compared to extended thinking. This provides a simple yet effective mechanism for test-time scaling of reasoning models.

Key facts. Cited by 53 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 52 of its 67 references that are in the map. First posted on arXiv 4 Jun 2025.

Builds on

52 of 67 references.

Closest ancestors · 19

Same direction, newest first

  1. 2025AdaCtrl: Towards Adaptive and Controllable Reasoning via Difficulty-Aware Budgeting24 May
  2. 2025When to Continue Thinking: Adaptive Thinking Mode Switching for Efficient Reasoning21 May
  3. 2025Think Only When You Need with Large Hybrid-Reasoning Models20 May
  4. 2025ThinkSwitcher: When to Think Hard, When to Think Fast20 May
  5. 2025AdaptThink: Reasoning Models Can Learn When to Think19 May
  6. 2025Thinkless: LLM Learns When to Think19 May
  7. 2025ThinkPrune: Pruning Long Chain-of-Thought of LLMs via Reinforcement Learning2 Apr
  8. 2025L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning6 Mar
  9. 2025Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning25 Feb
  10. 2025Scaling Test-Time Compute Without Verification or RL is Suboptimal17 Feb
  11. 2025s1: Simple test-time scaling31 Jan
  12. 2024O1 Replication Journey -- Part 2: Surpassing O1-preview through Simple Distillation, Big Progress or Bitter Lesson?25 Nov
7 more
  1. 2024Marco-o1: Towards Open Reasoning Models for Open-Ended Solutions21 Nov
  2. 2024O1 Replication Journey: A Strategic Progress Report -- Part 18 Oct
  3. 2024Interpretable Contrastive Monte Carlo Tree Search Reasoning2 Oct
  4. 2024Training Language Models to Self-Correct via Reinforcement Learning19 Sep
  5. 2024Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters6 Aug
  6. 2024Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for Problem-Solving with Language Models1 Aug
  7. 2024Theoretical guarantees on the best-of-n alignment policy3 Jan

Other references · 33

  1. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  2. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 9,379
  3. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,330
  4. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,371
  5. 2024Qwen2.5 Technical Reportcited by 4,942
  6. 2025DAPO: An Open-Source LLM Reinforcement Learning System at Scalecited by 2,810
  7. 2021WebGPT: Browser-assisted question-answering with human feedbackcited by 1,745
  8. 2025Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learningcited by 1,444
  9. 2025Kimi k1.5: Scaling Reinforcement Learning with LLMscited by 1,210
  10. 2021Measuring Coding Challenge Competence With APPScited by 902
  11. 2024DeepSeek-V3 Technical Reportcited by 811
  12. 2025Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Modelscited by 665
21 more
  1. 2025VLM-R1: A Stable and Generalizable R1-style Large Vision-Language Modelcited by 526
  2. 2025Visual-RFT: Visual Reinforcement Fine-Tuningcited by 513
  3. 2023Scaling Relationship on Learning Mathematical Reasoning with Large Language Modelscited by 427
  4. 2023Llemma: An Open Language Model For Mathematicscited by 388
  5. 2023Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Modelscited by 349
  6. 2024Quiet-STaR: Language Models Can Teach Themselves to Think Before Speakingcited by 301
  7. 2025Towards Large Reasoning Models: A Survey of Reinforced Reasoning with Large Language Modelscited by 293
  8. 2025Training Language Models to Reason Efficientlycited by 247
  9. 2024Chain-of-Thought Reasoning Without Promptingcited by 210
  10. 2024Distilling System 2 into System 1cited by 159
  11. 2024GSM-Plus: A Comprehensive Benchmark for Evaluating the Robustness of LLMs as Mathematical Problem Solverscited by 158
  12. 2024InternLM-Math: Open Math Large Language Models Toward Verifiable Reasoningcited by 119
  13. 2024Self-Improvement in Language Models: The Sharpening Mechanismcited by 104
  14. 2024OpenR: An Open Source Framework for Advanced Reasoning with Large Language Modelscited by 94
  15. 2025DeepSeek-R1 Thoughtology: Let's think about LLM Reasoningcited by 92
  16. 2024LLaMA-Berry: Pairwise Optimization for O1-like Olympiad-Level Mathematical Reasoningcited by 91
  17. 2023System 2 Attention (is something you might need too)cited by 89
  18. 2024o1-Coder: an o1 Replication for Codingcited by 78
  19. 2024Variational Best-of-N Alignmentcited by 51
  20. 2024Regularized Best-of-N Sampling with Minimum Bayes Risk Objective for Language Model Alignmentcited by 25
  21. 2023B-Coder: Value-Based Deep Reinforcement Learning for Program Synthesiscited by 17

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 18; 2026: 35.

53 papers.

Sort:

By area:

  1. 2025Inverse Scaling in Test-Time Computecited by 50
  2. 2025ParaThinker: Native Parallel Thinking as a New Paradigm to Scale LLM Test-time Computecited by 38
  3. 2025What Characterizes Effective Reasoning? Revisiting Length, Review, and Structure of CoTcited by 25
  4. 2026LLM Reasoning as Trajectories: Step-Specific Representation Geometry and Correctness Signalscited by 24
  5. 2025MoReBench: Evaluating Procedural and Pluralistic Moral Reasoning in Language Models, More than Outcomescited by 23
  6. 2025Rethinking Thinking Tokens: LLMs as Improvement Operatorscited by 21
  7. 2025Towards Concise and Adaptive Thinking in Large Reasoning Models: A Surveycited by 18
  8. 2025The Evolution of Thought: Tracking LLM Overthinking via Reasoning Dynamics Analysiscited by 18
  9. 2026Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoningcited by 18
  10. 2025DRPO: Efficient Reasoning via Decoupled Reward Policy Optimizationcited by 16
  11. 2026When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scalingcited by 16
  12. 2025The Art of Scaling Test-Time Compute for Large Language Modelscited by 12
38 more
  1. 2026Group Distributionally Robust Optimization-Driven Reinforcement Learning for LLM Reasoningcited by 12
  2. 2026The Hot Mess of AI: How Does Misalignment Scale With Model Intelligence and Task Complexity?cited by 10
  3. 2025R-Horizon: How Far Can Your Large Reasoning Model Really Go in Breadth and Depth?cited by 8
  4. 2025RADAR: Reasoning-Ability and Difficulty-Aware Routing for Reasoning LLMscited by 6
  5. 2026Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failurescited by 6
  6. 2026TACT: Mitigating Overthinking and Overacting in Coding Agents via Activation Steeringcited by 6
  7. 2025PEAR: Phase Entropy Aware Reward for Efficient Reasoningcited by 5
  8. 2025Overthinking Reduction with Decoupled Rewards and Curriculum Data Schedulingcited by 4
  9. 2025The Zero-Step Thinking: An Empirical Study of Mode Selection as Harder Early Exit in Reasoning Modelscited by 4
  10. 2025Refine and Align: Confidence Calibration through Multi-Agent Interaction in VQAcited by 4
  11. 2026Bridging Online and Offline RL: Contextual Bandit Learning for Multi-Turn Code Generationcited by 4
  12. 2026Understanding Performance Gap Between Parallel and Sequential Sampling in Large Reasoning Modelscited by 4
  13. 2026Walking Through Uncertainty: An Empirical Study of Uncertainty Estimation for Audio-Aware Large Language Modelscited by 4
  14. 2025CoreThink: A Symbolic Reasoning Layer to reason over Long Horizon Tasks with LLMscited by 3
  15. 2025ThinkPilot: Steering Reasoning Models via Automated Think-prefixes Optimizationcited by 3
  16. 2026Multi-Agent Reasoning Improves Compute Efficiency: Pareto-Optimal Test-Time Scalingcited by 3
  17. 2026SLMJury: Can Small Language Models Judge as Well as Large Ones?cited by 3
  18. 2026Recursive Self-Improvement via On-Policy Distillation for Reasoningcited by 3
  19. 2026interwhen: A Generalizable Framework for Steering Reasoning Models with Test-time Verificationcited by 2
  20. 2026Stop Unnecessary Reflection: Training LRMs for Efficient Reasoning with Adaptive Reflection and Length Coordinated Penaltycited by 2
  21. 2026StAR: Segment Anything Reasonercited by 2
  22. 2025VISTA: Verification In Sequential Turn-based Assessmentcited by 1
  23. 2026Boule or Baguette? A Study on Task Topology, Length Generalization, and the Benefit of Reasoning Tracescited by 1
  24. 2026VisRef: Visual Refocusing while Thinking Improves Test-Time Scaling in Multi-Modal Large Reasoning Modelscited by 1
  25. 2026One-Token Verification for Reasoning Correctness Estimationcited by 1
  26. 2026Constructing Synthetic Instruction Datasets for Improving Reasoning in Domain-Specific LLMs: A Case Study in the Japanese Financial Domaincited by 1
  27. 2026Test-Time Scaling for Small VLMs on Multilingual Visual MCQcited by 1
  28. 2026It's the Problem, Not the Path: Budget and Difficulty Confounds in LLM Reasoning Trajectoriescited by 1
  29. 2026Miner:Mining Intrinsic Mastery for Data-Efficient RL in Large Reasoning Modelscited by 0
  30. 2026Mitigating Cognitive Inertia in Large Reasoning Models via Latent Spike Steeringcited by 0
  31. 2026Thinking Is Not Telling: Information Disclosure in User-Service LLM Agentscited by 0
  32. 2026Entropy Centroids as Intrinsic Rewards for Test-Time Scalingcited by 0
  33. 2026Exploring Extrinsic and Intrinsic Properties for Effective Reasoning with Code Interpretercited by 0
  34. 2026Does Verbose Chain-of-Thought Really Help? In-Distribution Evidence that Content, Not Length, Matterscited by 0
  35. 2026Solver-Hard Is Not Model-Hard: A Hardness-Controlled Diagnostic for LLM Constraint Reasoningcited by 0
  36. 2026AdaThinking-E: One-Token Entropy Regulation for Adaptive Thinkingcited by 0
  37. 2026Revisiting Complete Reasoning Traces for Post-Trainingcited by 0
  38. 2026Sample Count Is Not Enough: Candidate-Generation Strategy Shapes the Energy and Performance of LLM Test-Time Scalingcited by 0

50 most cited of 53 All 53 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2025Test-Time Scaling in Reasoning Models Is Not Effective for Knowledge-Intensive Tasks Yet
  2. 2025Extending Test-Time Scaling: A 3D Perspective with Context, Batch, and Turn
  3. 2025Rethinking External Slow-Thinking: From Snowball Errors to Probability of Correct Reasoning
  4. 2025Think Twice: Enhancing LLM Reasoning by Scaling Multi-round Test-time Thinking
  5. 2026ThinkRetrieve: Retrieval-Augmented Reasoning Traces for Test-Time Scaling
  6. 2025Scaling over Scaling: Exploring Test-Time Scaling Plateau in Large Reasoning Models
  7. 2025When Can Large Reasoning Models Save Thinking? Mechanistic Analysis of Behavioral Divergence in Reasoning
  8. 2025Rethinking Inference-Time Scaling: Efficiency Limits and Linguistic Signals