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

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

Problem Statement
Reasoning-trained Large Language Models (LLMs) frequently allocate more test-time compute to harder problems, resulting in significantly longer chains of thought. However, existing literature has struggled to distinguish whether these models are merely performing more computational steps or fundamentally altering their internal reasoning trajectory. A critical confound in analyzing hidden-state trajectories is generation length: longer paths are mechanically less direct (more "tortuous") regardless of the model's internal logic. Consequently, raw geometric analyses often misattribute the mechanical effects of path length to problem difficulty or model architecture, obscuring whether reasoning training induces distinct internal dynamics.

Methodology
The authors employ a matched experimental design across three domains: competitive programming (Codeforces), mathematics (MATH), and Boolean satisfiability (SAT).

1. Difficulty Calibration: To establish a continuous difficulty scale independent of platform-specific labels, the authors fit a Rasch model (Item Response Theory) using the performance of 32 models across 1,500 items. This yields a latent difficulty variable ($b_i$) for each problem.
2. Trajectory Extraction: Hidden-state trajectories are extracted from the residual stream at five evenly spaced layers during the generation of solution segments. For reasoning models, this segment is the explicit thinking block; for instruction-tuned baselines, it is the text preceding the detected answer boundary.
3. Geometric Metrics: Two primary statistics are computed:
   • Directness ($D$): The ratio of net displacement to total path length ($\Delta / L$), measuring endpoint efficiency.
   • Curvature Variability ($V$): The standard deviation of Menger curvature over consecutive triples, measuring local bending heterogeneity.
4. Length Correction: Recognizing that raw geometry is dominated by generation length ($N$), the authors perform a residualization procedure. They fit a length-only regression ($\bar{D} \sim \log N$) for each model and layer, then correlate the residuals ($D_\perp$) with item difficulty ($b_i$). This isolates the difficulty-geometry coupling independent of path length.
5. Comparative Analysis: The study compares six matched pairs of reasoning-trained models (using R1 distillation, SFT+RL, or o3-mini distillation) against their instruction-tuned baselines.
6. Auxiliary Analyses: The authors employ linear probes to test for difficulty decodability at the prompt stage and use independent LLM judges to annotate generated traces for reasoning behaviors (e.g., strategy shifting, uncertainty monitoring).

Key Results

1. Sign Reversal via Length Correction: In raw trajectories, harder problems appear less direct because they generate longer paths. After length correction, this relationship reverses across all three domains: harder problems elicit more direct corrected trajectories. This indicates that raw geometry is a structural artifact of length, not a faithful representation of reasoning organization.
2. Domain-Dependent Reasoning Separation:
   • Competitive Programming: The clearest distinction emerges here. All six reasoning models exhibit a strong positive correlation between difficulty and corrected directness ($\rho_{D_\perp} \approx +0.41$), whereas matched baselines remain near zero or negative ($\approx -0.06$). Reasoning models also show reduced curvature variability on hard problems.
   • Mathematics: The separation is weaker and more heterogeneous ($\rho_{D_\perp} \approx +0.05$ for reasoning vs. $-0.07$ for baselines).
   • Boolean Satisfiability: Both reasoning and baseline models show positive corrected coupling, suggesting that difficulty-geometry coupling can emerge in instruction-tuned models without specific reasoning training in this domain.
3. Decoupling from Linear Probes: Prompt-stage linear probes do not mirror the reasoning-baseline separation observed in corrected geometry. In the code domain, the geometric gap is large while the difference in linear difficulty decodability is negligible. This suggests the effect is not merely stronger linear access to difficulty information but a change in how the trajectory evolves during generation.
4. Behavioral Correlates: Stronger corrected geometric coupling co-occurs with observable reasoning behaviors, specifically "strategy shifting" and "uncertainty monitoring," particularly in R1-distilled models. These annotations are descriptive correlates measured from the same traces, not causal mechanisms.

Significance and Claims
The paper establishes that generation length is a structural confound in generation-time trajectory analysis and that length correction is a prerequisite for valid interpretation. The authors claim that reasoning training is associated with distinct corrected trajectory geometry, but this effect is domain-dependent, being most pronounced in competitive programming.

The findings suggest that reasoning models do not just "think longer"; they follow different internal paths that become more direct and less locally heterogeneous as problem difficulty increases, specifically in domains requiring complex strategy selection. The paper posits that this geometric signal is complementary to, rather than a restatement of, linear difficulty decodability. Finally, the authors release a large-scale trajectory archive pairing generated chains of thought with sampled hidden-state trajectories to facilitate future research into the internal dynamics of reasoning.

Limitations
The authors acknowledge that the analysis identifies a robust geometric association rather than a causal mechanism. The segmentation of solution segments relies on heuristics for baselines, and the behavioral annotations are descriptive correlates derived from the same data as the geometry, precluding causal claims about how specific behaviors drive the geometric signal. Additionally, fixed-stride sampling may miss finer temporal structures.