mlsanityData to 14 Oct 2026

Position: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!

arXiv 2504.09762 ↗First posted 14 Apr 2025cs.AIcited by 37

On the map: LLM reasoning → Prompting and chain-of-thought reasoning

Abstract (arXiv)

Intermediate token generation (ITG), where a model produces output before the solution, has become a standard method to improve the performance of language models on reasoning tasks. These intermediate tokens have been called sayreasoning traces or even saythinking traces -- implicitly anthropomorphizing the traces, and implying that these traces resemble steps a human might take when solving a challenging problem, and as such can provide an interpretable window into the operation of the model's thinking process to the end user. In this position paper, we present evidence that this anthropomorphization isn't a harmless metaphor, and instead is quite dangerous -- it confuses the nature of these models and how to use them effectively, and leads to questionable research. We call on the community to avoid such anthropomorphization of intermediate tokens.

Key facts. Cited by 37 papers in the map (data through 14 Oct 2026). Placed in Prompting and chain-of-thought reasoning, one of 17 directions in LLM reasoning. Builds on 36 of its 90 references that are in the map. First posted on arXiv 14 Apr 2025.

Builds on

36 of 90 references.

Closest ancestors · 10

Same direction, newest first

  1. 2025Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation14 Mar
  2. 2025Chain-of-Thought Reasoning In The Wild Is Not Always Faithful11 Mar
  3. 2025LLMs Can Easily Learn to Reason from Demonstrations Structure, not content, is what matters!11 Feb
  4. 2024Training Large Language Models to Reason in a Continuous Latent Space9 Dec
  5. 2024On the Hardness of Faithful Chain-of-Thought Reasoning in Large Language Models15 Jun
  6. 2024Towards Better Chain-of-Thought: A Reflection on Effectiveness and Faithfulness29 May
  7. 2024Let's Think Dot by Dot: Hidden Computation in Transformer Language Models24 Apr
  8. 2024Beyond A*: Better Planning with Transformers via Search Dynamics Bootstrapping21 Feb
  9. 2024Making Reasoning Matter: Measuring and Improving Faithfulness of Chain-of-Thought Reasoning21 Feb
  10. 2022Automatic Chain of Thought Prompting in Large Language Models7 Oct

Other references · 26

  1. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,330
  2. 2023Let's Verify Step by Stepcited by 3,789
  3. 2023Universal and Transferable Adversarial Attacks on Aligned Language Modelscited by 3,104
  4. 2025DAPO: An Open-Source LLM Reinforcement Learning System at Scalecited by 2,810
  5. 2023Sparks of Artificial General Intelligence: Early experiments with GPT-4cited by 2,592
  6. 2025s1: Simple test-time scalingcited by 1,341
  7. 2024Tulu 3: Pushing Frontiers in Open Language Model Post-Trainingcited by 1,053
  8. 2021Show Your Work: Scratchpads for Intermediate Computation with Language Modelscited by 886
  9. 2023Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizescited by 737
  10. 2016Adaptive Computation Time for Recurrent Neural Networkscited by 506
  11. 2024Generative Verifiers: Reward Modeling as Next-Token Predictioncited by 480
  12. 2025The Lessons of Developing Process Reward Models in Mathematical Reasoningcited by 415
14 more
  1. 2025Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRscited by 377
  2. 2024Alignment faking in large language modelscited by 360
  3. 2025R1-Zero's "Aha Moment" in Visual Reasoning on a 2B Non-SFT Modelcited by 165
  4. 2023MiniLLM: On-Policy Distillation of Large Language Modelscited by 139
  5. 2025DeepSeek-R1 Thoughtology: Let's think about LLM Reasoningcited by 92
  6. 2024On Designing Effective RL Reward at Training Time for LLM Reasoningcited by 87
  7. 2025Concise Reasoning via Reinforcement Learningcited by 79
  8. 2024Best-of-N Jailbreakingcited by 76
  9. 2024The Ultimate Guide to Fine-Tuning LLMs from Basics to Breakthroughs: An Exhaustive Review of Technologies, Research, Best Practices, Applied Research Challenges and Opportunitiescited by 64
  10. 2025Understanding Aha Moments: from External Observations to Internal Mechanismscited by 50
  11. 2024Mastering Board Games by External and Internal Planning with Language Modelscited by 31
  12. 2025To Backtrack or Not to Backtrack: When Sequential Search Limits Model Reasoningcited by 24
  13. 2025Thinking beyond the anthropomorphic paradigm benefits LLM researchcited by 14
  14. 2024Talking Nonsense: Probing Large Language Models' Understanding of Adversarial Gibberish Inputscited by 11

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 14; 2026: 23.

37 papers.

Sort:

By area:

  1. 2025Trace Length is a Simple Uncertainty Signal in Reasoning Modelscited by 19
  2. 2025Code Like Humans: A Multi-Agent Solution for Medical Codingcited by 11
  3. 2025Lexical Hints of Accuracy in LLM Reasoning Chainscited by 9
  4. 2025Interpretable Traces, Unexpected Outcomes: Investigating the Disconnect in Trace-Based Knowledge Distillationcited by 8
  5. 2025The Need for Verification in AI-Driven Scientific Discoverycited by 7
  6. 2026Probing the Trajectories of Reasoning Traces in Large Language Modelscited by 7
  7. 2025Humans Perceive Wrong Narratives from AI Reasoning Textscited by 6
  8. 2025State over Tokens: Characterizing the Role of Reasoning Tokenscited by 5
  9. 2026Do explanations generalize across large reasoning models?cited by 5
  10. 2026Aletheia: What Makes RLVR For Code Verifiers Tick?cited by 5
  11. 2026Monitoring the Internal Monologue: Probe Trajectories Reveal Reasoning Dynamicscited by 5
  12. 2026Beyond Consensus: Trace-Level Synthesis in Mixture of Agentscited by 4
25 more
  1. 2025Bye-bye, Bluebook? Automating Legal Drudgery With AI-Augmented Rule Followingcited by 3
  2. 2025Temperature and Persona Shape LLM Agent Consensus With Minimal Accuracy Gains in Qualitative Codingcited by 3
  3. 2026The Stepwise Informativeness Assumption: Why are Entropy Dynamics and Reasoning Correlated in LLMs?cited by 3
  4. 2025Retrieval-augmented reasoning with lean language modelscited by 2
  5. 2025On the Notion that Language Models Reasoncited by 2
  6. 2026The Potential of CoT for Reasoning: A Closer Look at Trace Dynamicscited by 2
  7. 2026Faithful or Just Plausible? Evaluating the Faithfulness of Closed-Source LLMs in Medical Reasoningcited by 2
  8. 2026CONSCIENTIA: Can LLM Agents Learn to Strategize? Emergent Deception and Trust in a Multi-Agent NYC Simulationcited by 2
  9. 2025Feedback-Induced Performance Decline in LLM-Based Decision-Makingcited by 1
  10. 2025Illusions of reflection: open-ended task reveals systematic failures in Large Language Models' reflective reasoningcited by 1
  11. 2026Is my model "mind blurting"? Interpreting the dynamics of reasoning tokens with Recurrence Quantification Analysis (RQA)cited by 1
  12. 2026Humans and LLMs Diverge on Probabilistic Inferencescited by 1
  13. 2026Tracing Uncertainty in Language Model "Reasoning"cited by 1
  14. 2026What Am I Missing? Question-Answering as Hidden State Probingcited by 1
  15. 2025Implicature in Interaction: Understanding Implicature Improves Alignment in Human-LLM Interactioncited by 0
  16. 2026Disagreement as Data: Reasoning Trace Analytics in Multi-Agent Systemscited by 0
  17. 2026LLM Reasoning Is Latent, Not the Chain of Thoughtcited by 0
  18. 2026What properties of reasoning supervision are associated with improved downstream model quality?cited by 0
  19. 2026Some hypotheses on how chatbots work in problem-solution-driven conversations: Large Language Models as confirmation of the Innovation Illusioncited by 0
  20. 2026Humans Disengage, Reasoning Models Persist: Separating Difficulty Registration from Deliberation Allocationcited by 0
  21. 2026Solver-Hard Is Not Model-Hard: A Hardness-Controlled Diagnostic for LLM Constraint Reasoningcited by 0
  22. 2026The Disconnect Between Better Descriptive Reasoning Trace Quality and Recommendation Effectivenesscited by 0
  23. 2026agentic-ger: terminology recovery in long-form speech using global contextcited by 0
  24. 2026Correct Answers, Invalid Traces: What Verifiable Grade-School Math Reveals About Chain-of-Thought Tracescited by 0
  25. 2026Anthropomorphism in the age of Large Language Models: An overview of potential risks and mitigationscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens
  2. 2025Performative Thinking? The Brittle Correlation Between CoT Length and Problem Complexity
  3. 2026Revisiting Anthropomorphic Reflection Markers in Large Language Model Reasoning
  4. 2025Humanizing Machines: Rethinking LLM Anthropomorphism Through a Multi-Level Framework of Design
  5. 2024Abstraction-of-Thought Makes Language Models Better Reasoners
  6. 2025Are language models aware of the road not taken? Token-level uncertainty and hidden state dynamics
  7. 2024Rethinking Thinking Tokens: Understanding Why They Underperform in Practice
  8. 2026When Chain-of-Thought Fails, the Solution Hides in the Hidden States