What does this research mean for the field?

Different reasoning types leave distinct geometric signatures in transformer embedding spaces, which can be classified with 84.2% accuracy using a lightweight classifier. Novelty: ClaimNovelty.NOVEL_FINDING. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

The aim is to identify and route different reasoning types based on their geometric patterns in embedding spaces.

March 1, 2026Open Access

ARTI: Detecting and Routing Reasoning Types from Embedding Geometry

Key Points

The aim is to identify and route different reasoning types based on their geometric patterns in embedding spaces.
Developed a lightweight classifier with 20K parameters
Classified 10 reasoning types in embedding spaces
Conducted falsification experiments to validate findings
Used GCMC for validation with R² metrics
Achieved 84.2% accuracy in classifying reasoning types
ARTI-routed scoring reached 81.0% in StrategyQA with a significant improvement
Validated the necessity of a frozen encoder with a falsification chain
GCMC validation yielded R² = 0.678

Abstract

We show that different reasoning types leave distinct geometric signatures in transformer embedding spaces, detectable by a lightweight classifier with only 20K parameters. ARTI classifies 10 reasoning types at 84.2% accuracy (8.4x chance). ARTI-routed scoring achieves StrategyQA 81.0% (+32.6pp). Three-experiment falsification chain proves frozen encoder is structurally necessary. GCMC validation: R² = 0.678.

ARTI: Detecting and Routing Reasoning Types from Embedding Geometry

Key Points

Abstract

Cite This Study