Probing Representational Emergence in Large Language Models

Author

• Shawn Ismail, Kennesaw State UniversityFollow

Semester of Graduation

Spring 2026

Degree Type

Thesis

Degree Name

MS in Computer Science

Department

Department of Computer Science

Committee Chair/First Advisor

Ramazan Aygun

Abstract

This thesis investigates whether abrupt behavioral gains in large language models under scaling are accompanied by systematic changes in internal representations. It combines a behavioral screen of 65 tasks per family with targeted layerwise probing across eight decoder-only, open-weight model families. Behavioral emergence is defined for each family-task trajectory using an empirical jump detector, with segmented regression retained only as a diagnostic. The representational follow-up analyzes 27 selected MMLU subtasks shared across all families, spanning 37 checkpoints and 216 family-task units.

For each follow-up checkpoint, frozen linear probes are trained on every layer's hidden states to measure how much task-relevant information is linearly decodable and where it becomes accessible across depth. Representational change is summarized using normalized decodability mass above chance, sustained decoding onset, peak probe depth, depth center of mass, and alignment between behavioral jump windows and adjacent representational changes.

Within the selected matched cohort, the results show a family-sensitive graded relationship between behavioral emergence and representational reorganization. The normalized decodability mass tracks the behavioral accuracy most consistently in strong and weak trajectories and less consistently in smooth ones. Strong emergence shows a coordinated depth signature in which decodable information appears earlier in normalized depth while the peak and center-of-mass signal shift deeper. Strong trajectories are most clearly distinguished by coordinated alignment around the primary behavioral jump. Smooth trajectories can contain a localized or shallow signal, but they do not develop the broad late-layer profile seen in strong cases.