mlsanityData to 14 Oct 2026

Inverse Scaling in Test-Time Compute

arXiv 2507.14417 ↗First posted 19 Jul 2025cs.AI cs.CLcited by 50

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

Abstract (arXiv)

We construct evaluation tasks where extending the reasoning length of Large Reasoning Models (LRMs) deteriorates performance, exhibiting an inverse scaling relationship between test-time compute and accuracy. Our evaluation tasks span four categories: simple counting tasks with distractors, regression tasks with spurious features, deduction tasks with constraint tracking, and advanced AI risks. We identify five distinct failure modes when models reason for longer: 1) Claude models become increasingly distracted by irrelevant information; 2) OpenAI o-series models resist distractors but overfit to problem framings; 3) models shift from reasonable priors to spurious correlations; 4) all models show difficulties in maintaining focus on complex deductive tasks; and 5) extended reasoning may amplify concerning behaviors, with Claude Sonnet 4 showing increased expressions of self-preservation. These findings suggest that while test-time compute scaling remains promising for improving model capabilities, it may inadvertently reinforce problematic reasoning patterns. Our results demonstrate the importance of evaluating models across diverse reasoning lengths to identify and address these failure modes in LRMs.

Key facts. Cited by 50 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 38 of its 60 references that are in the map. First posted on arXiv 19 Jul 2025.

Builds on

38 of 60 references.

Closest ancestors · 16

Same direction, newest first

  1. 2025Does Thinking More always Help? Mirage of Test-Time Scaling in Reasoning Models4 Jun
  2. 2025Don't Overthink it. Preferring Shorter Thinking Chains for Improved LLM Reasoning23 May
  3. 2025Between Underthinking and Overthinking: An Empirical Study of Reasoning Length and correctness in LLMs30 Apr
  4. 2025Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models20 Mar
  5. 2025L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning6 Mar
  6. 2025DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models6 Mar
  7. 2025Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning25 Feb
  8. 2025The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks12 Feb
  9. 2025Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach7 Feb
  10. 2025Demystifying Long Chain-of-Thought Reasoning in LLMs5 Feb
  11. 2025Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning5 Feb
  12. 2025s1: Simple test-time scaling31 Jan
4 more
  1. 2025O1-Pruner: Length-Harmonizing Fine-Tuning for O1-Like Reasoning Pruning22 Jan
  2. 2024Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs30 Dec
  3. 2024Token-Budget-Aware LLM Reasoning24 Dec
  4. 2024Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters6 Aug

Other references · 22

  1. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  2. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,330
  3. 2024OpenAI o1 System Cardcited by 1,244
  4. 2025Kimi k1.5: Scaling Reinforcement Learning with LLMscited by 1,210
  5. 2024Training Large Language Models to Reason in a Continuous Latent Spacecited by 797
  6. 2021BBQ: A Hand-Built Bias Benchmark for Question Answeringcited by 701
  7. 2016Solving General Arithmetic Word Problemscited by 360
  8. 2024Deliberative Alignment: Reasoning Enables Safer Language Modelscited by 308
  9. 2025Chain of Draft: Thinking Faster by Writing Lesscited by 297
  10. 2025Logic-RL: Unleashing LLM Reasoning with Rule-Based Reinforcement Learningcited by 262
  11. 2025Training Language Models to Reason Efficientlycited by 247
  12. 2025CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillationcited by 225
10 more
  1. 2023Inverse Scaling: When Bigger Isn't Bettercited by 207
  2. 2025When More is Less: Understanding Chain-of-Thought Length in LLMscited by 183
  3. 2024Evaluation of OpenAI o1: Opportunities and Challenges of AGIcited by 127
  4. 2025BIG-Bench Extra Hardcited by 125
  5. 2025Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuningcited by 123
  6. 2025Trading Inference-Time Compute for Adversarial Robustnesscited by 61
  7. 2024Causal Language Modeling Can Elicit Search and Reasoning Capabilities on Logic Puzzlescited by 38
  8. 2025Evaluating o1-Like LLMs: Unlocking Reasoning for Translation through Comprehensive Analysiscited by 26
  9. 2024Compositional Hardness of Code in Large Language Models -- A Probabilistic Perspectivecited by 3
  10. 2025RELIC: Evaluating Complex Reasoning via the Recognition of Languages In-Contextcited by 3

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 23; 2026: 27.

50 papers.

Sort:

By area:

  1. 2025SimBench: Benchmarking the Ability of Large Language Models to Simulate Human Behaviorscited by 52
  2. 2025OptimalThinkingBench: Evaluating Over and Underthinking in LLMscited by 35
  3. 2026How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Taskscited by 33
  4. 2025Base Models Know How to Reason, Thinking Models Learn Whencited by 27
  5. 2026Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokenscited by 22
  6. 2025Measuring Chain-of-Thought Monitorability Through Faithfulness and Verbositycited by 17
  7. 2025Stop When Enough: Adaptive Early-Stopping for Chain-of-Thought Reasoningcited by 16
  8. 2025The Art of Scaling Test-Time Compute for Large Language Modelscited by 12
  9. 2025Test-Time Scaling in Reasoning Models Is Not Effective for Knowledge-Intensive Tasks Yetcited by 11
  10. 2025PiCSAR: Probabilistic Confidence Selection And Ranking for Reasoning Chainscited by 10
  11. 2026The Hot Mess of AI: How Does Misalignment Scale With Model Intelligence and Task Complexity?cited by 10
  12. 2026From RAG to Agentic RAG for Faithful Islamic Question Answeringcited by 9
38 more
  1. 2026The Illusion of Stochasticity in LLMscited by 9
  2. 2026Lost in the Noise: How Reasoning Models Fail with Contextual Distractorscited by 8
  3. 2025Planner-R1: Reward Shaping Enables Efficient Agentic RL with Smaller LLMscited by 7
  4. 2026How Inference Compute Shapes Frontier LLM Evaluationcited by 7
  5. 2025Generalized Parallel Scaling with Interdependent Generationscited by 6
  6. 2026Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failurescited by 6
  7. 2025Martingale Score: An Unsupervised Metric for Bayesian Rationality in LLM Reasoningcited by 5
  8. 2026InferenceBench: A Benchmark for Open-Ended LLM Inference Optimization by AI Agentscited by 5
  9. 2025Language Models and Logic Programs for Trustworthy Tax Reasoningcited by 4
  10. 2025Scaling Policy Compliance Assessment in Language Models with Policy Reasoning Tracescited by 4
  11. 2026Understanding Performance Gap Between Parallel and Sequential Sampling in Large Reasoning Modelscited by 4
  12. 2025Certainty-Guided Reasoning in Large Language Models: A Dynamic Thinking Budget Approachcited by 3
  13. 2025Word Salad Chopper: Reasoning Models Waste A Ton Of Decoding Budget On Useless Repetitions, Self-Knowinglycited by 3
  14. 2025Evaluating Code Reasoning Abilities of Large Language Models Under Real-World Settingscited by 3
  15. 2026Capture the Flags: Family-Based Evaluation of Agentic LLMs via Semantics-Preserving Transformationscited by 3
  16. 2026Limited Reasoning Space: The cage of long-horizon reasoning in LLMscited by 3
  17. 2025LLM-Based Agents for Competitive Landscape Mapping in Drug Asset Due Diligencecited by 2
  18. 2025Evaluating the Robustness of Chinchilla Compute-Optimal Scalingcited by 2
  19. 2026Critique of Agent Modelcited by 2
  20. 2026Reasoning effort, not tool access, buys first-try reliability in agentic code generation: an observational studycited by 2
  21. 2026Prefix Sliding for efficient test-time scalingcited by 2
  22. 2025Lucy: edgerunning agentic web search on mobile with machine generated task vectorscited by 1
  23. 2025Platonic Representations for Poverty Mapping: Unified Vision-Language Codes or Agent-Induced Novelty?cited by 1
  24. 2025FlagEval Findings Report: A Preliminary Evaluation of Large Reasoning Models on Automatically Verifiable Textual and Visual Questionscited by 1
  25. 2026Test-Time Scaling for Small VLMs on Multilingual Visual MCQcited by 1
  26. 2026It's the Problem, Not the Path: Budget and Difficulty Confounds in LLM Reasoning Trajectoriescited by 1
  27. 2026Fork-and-Flush: Escaping Idea Basins in Autoresearch Agentscited by 1
  28. 2025A Real-Time, Self-Tuning Moderator Framework for Adversarial Prompt Detectioncited by 0
  29. 2025VCORE: Variance-Controlled Optimization-based Reweighting for Chain-of-Thought Supervisioncited by 0
  30. 2026Absurd World: A Simple Yet Powerful Method to Absurdify the Real-world for Probing LLM Reasoning Capabilitiescited by 0
  31. 2026LLM Jaggedness Unlocks Scientific Creativitycited by 0
  32. 2026Stateful Reasoning via Insight Replaycited by 0
  33. 2026Reinforcement learning to improve large language model-based automated code compliance systemscited by 0
  34. 2026Position: Quantum Program Generation Must Prioritize Validity Over Probabilistic Scalingcited by 0
  35. 2026ThinkRetrieve: Retrieval-Augmented Reasoning Traces for Test-Time Scalingcited by 0
  36. 2026Playing log(N)-Questions over Wikipedia Abstracts: How Per-Round Errors Compound Under Information Asymmetrycited by 0
  37. 2026Scaling of Capability and Efficiency at Inference Time in Large Reasoning Modelscited by 0
  38. 2026When Reasoning Goes Astray: Attention Dynamics of Uncontrolled Reasoningcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025ARISE: An Adaptive Resolution-Aware Metric for Test-Time Scaling Evaluation in Large Reasoning Models
  2. 2025Scaling over Scaling: Exploring Test-Time Scaling Plateau in Large Reasoning Models
  3. 2025Rethinking Inference-Time Scaling: Efficiency Limits and Linguistic Signals
  4. 2026When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling
  5. 2026Test-Time Scaling in Reasoning LLMs: Inference Regimes, Evaluation, and Reproducibility
  6. 2025The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
  7. 2025On the Role of Temperature Sampling in Test-Time Scaling
  8. 2026Beyond What Seems Necessary: Hidden Gains from Scaling Training-Time Reasoning Length under Outcome Supervision