mlsanityData to 14 Oct 2026

Reasoning Models Don't Just Think Longer, They Move Differently

arXiv 2605.15454 ↗First posted 14 May 2026cs.CL cs.LG stat.MLcited by 2

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

Abstract (arXiv)

Reasoning-trained language models often spend more tokens on harder problems, but longer chains of thought do not show whether a model is merely computing for more steps or following a different internal trajectory. We study this distinction through hidden-state trajectories during chain-of-thought generation across competitive programming, mathematics, and Boolean satisfiability. Raw trajectory geometry is strongly shaped by generation length: longer generations mechanically alter path statistics, so difficulty-dependent comparisons are misleading without adjustment. After residualizing trajectory statistics on length, difficulty remains systematically coupled to corrected trajectory geometry across all domains studied. The clearest reasoning-specific separation appears in the code domain, where harder problems show more direct corrected trajectories and less heterogeneous local curvature in reasoning-trained models than in matched instruction-tuned baselines. Corrected difficulty-geometry coupling is weaker, but still present, in mathematics and Boolean satisfiability. Prompt-stage linear probes do not mirror the code-domain separation, and behavioral annotations show that stronger corrected coupling co-occurs with strategy shifts and uncertainty monitoring. Together, these findings establish length correction as a prerequisite for generation-time trajectory analysis and show that reasoning training can be associated with distinct corrected trajectory geometry, with the strength of the effect depending on the domain.

Key facts. Cited by 2 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 23 of its 29 references that are in the map. First posted on arXiv 14 May 2026.

Builds on

23 of 29 references.

Closest ancestors · 4

Same direction, newest first

  1. 2026LLM Reasoning as Trajectories: Step-Specific Representation Geometry and Correctness Signals7 Apr
  2. 2025Mitigating Overthinking in Large Reasoning Models via Manifold Steering28 May
  3. 2025Between Underthinking and Overthinking: An Empirical Study of Reasoning Length and correctness in LLMs30 Apr
  4. 2024Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters6 Aug

Other references · 19

  1. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,371
  2. 2016Understanding intermediate layers using linear classifier probescited by 1,353
  3. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  4. 2023Steering Language Models With Activation Engineeringcited by 922
  5. 2019Designing and Interpreting Probes with Control Taskscited by 640
  6. 2020Null It Out: Guarding Protected Attributes by Iterative Nullspace Projectioncited by 455
  7. 2024tinyBenchmarks: evaluating LLMs with fewer examplescited by 310
  8. 2018Estimating the intrinsic dimension of datasets by a minimal neighborhood informationcited by 250
  9. 2025Lost in Benchmarks? Rethinking Large Language Model Benchmarking with Item Response Theorycited by 45
  10. 2025SATBench: Benchmarking LLMs' Logical Reasoning via Automated Puzzle Generation from SAT Formulascited by 33
  11. 2023Large language models implicitly learn to straighten neural sentence trajectories to construct a predictive representation of natural languagecited by 32
  12. 2025The LLM Already Knows: Estimating LLM-Perceived Question Difficulty via Hidden Representationscited by 29
7 more
  1. 2024Easy2Hard-Bench: Standardized Difficulty Labels for Profiling LLM Performance and Generalizationcited by 24
  2. 2025The Geometry of Reasoning: Flowing Logics in Representation Spacecited by 21
  3. 2023Aligned and oblique dynamics in recurrent neural networkscited by 10
  4. 2026Truth as a Trajectory: What Internal Representations Reveal About Large Language Model Reasoningcited by 10
  5. 2025Probing the Difficulty Perception Mechanism of Large Language Modelscited by 8
  6. 2025LLMs Encode How Difficult Problems Arecited by 7
  7. 2025Latency-Response Theory Model: Evaluating Large Language Models via Response Accuracy and Chain-of-Thought Lengthcited by 5

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 2.

2 papers.

  1. 2026It's the Problem, Not the Path: Budget and Difficulty Confounds in LLM Reasoning Trajectoriescited by 1
  2. 2026MetaSteer: Context-Conditioned, nonlinear Steering via Attention-Projection Adaptationcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026How Hard Does It Think? Analyzing Step-Aware Reasoning Energy in LLM Chain-of-Thought Trajectories
  2. 2026Geometric Signatures of Reasoning: A Spectral Perspective on Task Hardness
  3. 2025Performative Thinking? The Brittle Correlation Between CoT Length and Problem Complexity
  4. 2026Beyond Scalars: Evaluating and Understanding LLM Reasoning via Geometric Progress and Stability
  5. 2026Directional Reasoning Trajectory Change (DRTC): Identifying Critical Trace Segments in Reasoning Models
  6. 2026Reasoning as State Transition: A Representational Analysis of Reasoning Evolution in Large Language Models
  7. 2025The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
  8. 2026Answer-Distribution Trajectories: A Stochastic-Dynamics View of LLM Reasoning