research
Dynamic Modes Ranked the Tokens Behind a Decision
DMDIntel decomposes hidden-state trajectories before attributing a classifier's output to input tokens.
Summary
DMDIntel decomposes hidden-state trajectories before attributing a classifier's output to input tokens.
Across three datasets and three model families, the authors report stronger ranked attributions than PCA, integrated gradients and SHAP. The comparisons evaluate the stated classification tasks; they do not make every language-model decision transparent.
Why it matters
DMDIntel decomposes hidden-state trajectories before attributing a classifier's output to input tokens.
Limits and context
- The comparisons evaluate the stated classification tasks; they do not make every language-model decision transparent.
Key claims
DMDIntel decomposes hidden-state trajectories before attributing a classifier's output to input tokens.
Qualification: The comparisons evaluate the stated classification tasks; they do not make every language-model decision transparent.
Evidence: source-2026-08-14-018
Sources
- arXiv preprint 2608.13048arXiv · primary research
Corrections
No corrections have been recorded for this story.