mlsanityData to 7 Oct 2026

Base Models Know How to Reason, Thinking Models Learn When

arXiv 2510.07364 ↗First posted 8 Oct 2025cs.AI cs.LGcited by 25

On the map: LLM reasoning → Reasoning benchmarks and evaluation

Abstract (arXiv)

What do thinking language models learn during training that their base models lack? We first present an unsupervised method that discovers a model's reasoning behaviors by training small Sparse Autoencoders on sentence-level activations of reasoning traces, yielding interpretable reasoning taxonomies. Building on this, we introduce constructive model diffing, which aims to reconstruct the base-to-fine-tuned difference from interpretable components: reasoning mechanisms (category vectors that can induce a reasoning behavior in the base model) and reasoning heuristics (a classifier determining when a mechanism should fire). Across nine base/thinking pairs (four RL-trained, four SFT-distilled, one mixed), two independent findings agree: category vectors in the base model converge to far lower loss for taxonomies derived from purely RL-trained models, and hybrid models recover roughly 76% of the RL base-to-thinking gap but only 11% of the SFT gap. This indicates RL primarily teaches heuristics for orchestrating pre-existing base mechanisms, whereas SFT-distillation installs new ones, offering a new lens on what training paradigms teach, with implications for efficient reasoning-model development.

Key facts. Cited by 25 papers in the map (data through 7 Oct 2026). Placed in Reasoning benchmarks and evaluation, one of 17 directions in LLM reasoning. Builds on 38 of its 58 references that are in the map. First posted on arXiv 8 Oct 2025.

Builds on

38 of 58 references.

Closest ancestors · 5

Same direction, newest first

  1. 2025100 Days After DeepSeek-R1: A Survey on Replication Studies and More Directions for Reasoning Language Models1 May
  2. 2025DeepSeek-R1 Thoughtology: Let's think about LLM Reasoning2 Apr
  3. 2025I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders24 Mar
  4. 2023Towards a Mechanistic Interpretation of Multi-Step Reasoning Capabilities of Language Models23 Oct
  5. 2021Show Your Work: Scratchpads for Intermediate Computation with Language Models30 Nov

Other references · 33

  1. 2021Training Verifiers to Solve Math Word Problemscited by 9,743
  2. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,767
  3. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,203
  4. 2024MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmarkcited by 1,758
  5. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,359
  6. 2025s1: Simple test-time scalingcited by 1,311
  7. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,239
  8. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,009
  9. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 960
  10. 2025Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?cited by 948
  11. 2023Steering Language Models With Activation Engineeringcited by 824
  12. 2024Scaling and evaluating sparse autoencoderscited by 530
21 more
  1. 2025Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Modelcited by 450
  2. 2025Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRscited by 373
  3. 2025Reinforcement Learning for Reasoning in Large Language Models with One Training Examplecited by 288
  4. 2023SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Modelscited by 246
  5. 2013k-Sparse Autoencoderscited by 209
  6. 2023TheoremQA: A Theorem-driven Question Answering datasetcited by 189
  7. 2025Thought Anchors: Which LLM Reasoning Steps Matter?cited by 126
  8. 2025Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretrainingcited by 123
  9. 2025Sparse Autoencoders Trained on the Same Data Learn Different Featurescited by 86
  10. 2025NaturalReasoning: Reasoning in the Wild with 2.8M Challenging Questionscited by 81
  11. 2025Understanding Reasoning in Thinking Language Models via Steering Vectorscited by 69
  12. 2025Inverse Scaling in Test-Time Computecited by 49
  13. 2025Overcoming Sparsity Artifacts in Crosscoders to Interpret Chat-Tuningcited by 31
  14. 2025Steering LLM Thinking with Budget Guidancecited by 30
  15. 2025Towards Understanding Distilled Reasoning Models: A Representational Approachcited by 26
  16. 2025Reasoning-Finetuning Repurposes Latent Representations in Base Modelscited by 23
  17. 2024Nudging: Inference-time Alignment of LLMs via Guided Decodingcited by 22
  18. 2025Climbing the Ladder of Reasoning: What LLMs Can-and Still Can't-Solve after SFT?cited by 21
  19. 2025Steering When Necessary: Flexible Steering Large Language Models with Backtrackingcited by 7
  20. 2025Internal states before wait modulate reasoning patternscited by 6
  21. 2025Controlling Large Language Model with Latent Actionscited by 5

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 7; 2026: 18.

25 papers.

Sort:

By area:

  1. 2026Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thoughtcited by 35
  2. 2026How do LLMs Compute Verbal Confidencecited by 23
  3. 2025Verifying Chain-of-Thought Reasoning via Its Computational Graphcited by 20
  4. 2025Measuring Chain-of-Thought Monitorability Through Faithfulness and Verbositycited by 15
  5. 2026Reasoning Models Generate Societies of Thoughtcited by 11
  6. 2026Triggering Chain-of-Thought via Latent Feature Interventions in Large Language Modelscited by 9
  7. 2025State over Tokens: Characterizing the Role of Reasoning Tokenscited by 5
  8. 2026Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failurescited by 5
  9. 2026RISER: Orchestrating Latent Reasoning Skills for Adaptive Activation Steeringcited by 4
  10. 2025Modeling Hierarchical Thinking in Large Reasoning Modelscited by 3
  11. 2026How LLMs Detect and Correct Their Own Errors: The Role of Internal Confidence Signalscited by 3
  12. 2026Reasoning Can Be Restored by Correcting a Few Decision Tokenscited by 2
13 more
  1. 2026Neural Procedural Memory: Empowering LLM Agents with Implicit Activation Steeringcited by 2
  2. 2025Position: On the Methodological Pitfalls of Evaluating Base LLMs for Reasoningcited by 0
  3. 2025Step-Tagging: Toward controlling the generation of Language Reasoning Models through step monitoringcited by 0
  4. 2025The Deepfake Detective: Interpreting Neural Forensics Through Sparse Features and Manifoldscited by 0
  5. 2026Transcoder Adapters for Reasoning-Model Diffingcited by 0
  6. 2026The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignmentcited by 0
  7. 2026TRACES: Tagging Reasoning Steps for Adaptive Cost-Efficient Early-Stoppingcited by 0
  8. 2026Latent Reward Steering: An Adaptive Inference-Time Framework that Implicitly Promotes Cognitive Behaviors in Reasoning LLMscited by 0
  9. 2026Reasoning Fine-Tuning Induces Persistent Latent Policy Statescited by 0
  10. 2026On the Generalization of Steering Vectors for Chain-of-Thought Faithfulnesscited by 0
  11. 2026Math Reasoning in LLMs is Organized by Approach, Not Topiccited by 0
  12. 2026Fine Until Fine-Tuned: Repeated Solutions Make Reasoning Fragilecited by 0
  13. 2026A Polyphonic Conception of AI Understandingcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Thinking About Thinking: Evaluating Reasoning in Post-Trained Language Models
  2. 2026Reasoning as State Transition: A Representational Analysis of Reasoning Evolution in Large Language Models
  3. 2023Phenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinement
  4. 2024Reasoning Circuits in Language Models: A Mechanistic Interpretation of Syllogistic Inference
  5. 2026Convergence Without Understanding: When Language Models Agree on Representations but Disagree on Reasoning
  6. 2022ALERT: Adapting Language Models to Reasoning Tasks
  7. 2025The Challenge of Teaching Reasoning to LLMs Without RL or Distillation
  8. 2026Woodpecker Distillation: Weak Models Diagnose Reasoning Bugs in Strong Models