CASA: Concept-Axis Self-Adjustment. A Closed Loop for Continuous Self-Correction of Language Models
Year: 2026
Type: preprint
Abstract: This paper turns a raw research intuition into a concrete algorithm. The intuition: while running,a language model should be able to notice that its representation of an entity disagrees withverifiable knowledge—its own freshly generated statement, an implication of an ontology, or anexternal reference—and apply a small displacement of that representation along the relevantconcept axis; accumulated displacements should then be periodically consolidated into weights.The result would be a closed loop of self-improvement that never retrains on a corpus. Wedevelop the idea into CASA (Concept-Axis Self-Adjustment): a formal setup, update ruleswith a trust region, a two-level memory (edit cache → weights), stability conditions, an honestlist of failure modes, and a falsifiable MVP plan for a single GPU.
Author: Nedilko Aleksandr Yuryevich
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236684773
Cites:
Cited by:
Related to:
