![[2602.15332] Directional Reasoning Trajectory Change (DRTC): Identifying Critical Trace Segments in Reasoning Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.15332] Directional Reasoning Trajectory Change (DRTC): Identifying Critical Trace Segments in Reasoning Models
Summary
The paper introduces Directional Reasoning Trajectory Change (DRTC), a framework for interpreting long-horizon reasoning in language models by identifying critical decision points and their causal influences.
Why It Matters
Understanding how language models reason is crucial for developing more transparent AI systems. DRTC offers insights into the decision-making processes of these models, potentially improving their interpretability and reliability in applications.
Key Takeaways
- DRTC identifies pivotal decision points in reasoning models using uncertainty signals.
- The framework provides a causally grounded view of how context influences reasoning.
- Empirical results show that learned spans outperform random spans in reasoning tasks.
- DRTC measures intervention effects on model trajectories, enhancing interpretability.
- The study highlights the concentration of directional influence across reasoning models.
Computer Science > Machine Learning arXiv:2602.15332 (cs) [Submitted on 17 Feb 2026] Title:Directional Reasoning Trajectory Change (DRTC): Identifying Critical Trace Segments in Reasoning Models Authors:Waldemar Chang View a PDF of the paper titled Directional Reasoning Trajectory Change (DRTC): Identifying Critical Trace Segments in Reasoning Models, by Waldemar Chang View PDF HTML (experimental) Abstract:Understanding how language models carry out long-horizon reasoning remains an open challenge. Existing interpretability methods often highlight tokens or spans correlated with an answer, but they rarely reveal where the model makes consequential reasoning turns, which earlier context causally triggers those turns, or whether the highlighted text actually steers the reasoning process. We introduce Directional Reasoning Trajectory Change (DRTC), a process-causal framework for interpreting long-form reasoning from a single on-policy rollout. DRTC detects pivot decision points using uncertainty and distribution-shift signals, then applies receiver-side interventions that preserve the realized rollout without resampling the continuation while blocking information flow from selected earlier chunks only at a pivot. It measures whether each intervention redirects the direction of the model's log-probability trajectory relative to the realized rollout direction, producing a signed per-chunk attribution score. We also compute turning-angle curvature changes on raw logits as a comp...