Study Audits Decode-Generate-Steer Gap in Frozen LLMs for Geometric Constraints
A recent study published on arXiv has uncovered significant shortcomings in large language models (LLMs) regarding their handling of geometric reasoning and CAD constraints. The research investigates the performance of six static decoder-only LLMs, focusing on aspects such as linear decodability and how behavior is managed. Results indicate that while pretraining improves the models' understanding of local geometric relations—even when positional information is modified—sketch-level degrees-of-freedom (DOF) show minimal enhancements from random input sources. This underscores a notable disconnect between LLMs' encoded knowledge and their practical applications. The paper, titled "Encoded but Not Actionable," is cataloged as 2608.17843.
Key facts
- The preprint is available on arXiv under identifier 2608.17843 and is cross-listed.
- The study investigates the gap between what LLMs encode and how they behave using geometric reasoning.
- Parametric CAD constraints serve as a controlled testbed for the research.
- Six frozen decoder-only LLMs were probed in the experiments.
- Four properties were examined: linear decodability, forced-choice generation, activation-level influence, and behavioral steerability.
- Pretraining substantially improves decoding of local geometric relations, even after controlling for positional cues.
- Sketch-level DOF status is highly decodable from random initialization, with only modest improvement from pretraining.
- The paper is titled 'Encoded but Not Actionable: Auditing the Decode-Generate-Steer Gap in Frozen LLMs for Geometric Constraints'.
Entities
—