Task-Dependent Intrinsic Dimensionality Across Transformer Layers: An Exploratory Study of Qwen2.5-1.5B

Abstract We ask whether task categories differ in the dimensionality of a pretrained language model’s internal representations and whether their relationship changes with depth. Last-input-token representations are extracted from Qwen2.5-1.5B for 200 prompts spanning fac-tual recall, arithmetic, reasoning, and coding. Representation geometry is summarized using participation ratio, covariance-spectrum entropy effective rank, principal-component vari-ance thresholds, and the TWO-NN local intrinsic-dimension estimator. The primary held-out analysis contains 12 prompts from two template families per category. For raw representa-tions, the reasoning-minus-recall participation-ratio contrast is positive at blocks 1–20 and negative at blocks 21–28. At block 24, participation ratio is 2.869 for factual recall and 1.397 for reasoning; after the final block and normalization, it is 2.235 and 2.126, respectively. The observations establish a descriptive depth-dependent relationship in these prompt sets, rather than a uniform ordering of task categories. Limited held-out family counts prevent reliable primary family-bootstrap intervals, and insufficient length overlap prevents a joint matched-length comparison. We report a bounded exploratory experiment with its analysis code and limitations.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-11703266/v1
Primary Topic
Topic Modeling
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Task-Dependent Intrinsic Dimensionality Across Transformer Layers: An Exploratory Study of Qwen2.5-1.5B

Mitra Varun Gogulapati
Research Square
Topic Modeling
preprint

Task-Dependent Intrinsic Dimensionality Across Transformer Layers: An Exploratory Study of Qwen2.5-1.5B

Mitra Varun Gogulapati
preprint en

Abstract

Abstract We ask whether task categories differ in the dimensionality of a pretrained language model’s internal representations and whether their relationship changes with depth. Last-input-token representations are extracted from Qwen2.5-1.5B for 200 prompts spanning fac-tual recall, arithmetic, reasoning, and coding. Representation geometry is summarized using participation ratio, covariance-spectrum entropy effective rank, principal-component vari-ance thresholds, and the TWO-NN local intrinsic-dimension estimator. The primary held-out analysis contains 12 prompts from two template families per category. For raw representa-tions, the reasoning-minus-recall participation-ratio contrast is positive at blocks 1–20 and negative at blocks 21–28. At block 24, participation ratio is 2.869 for factual recall and 1.397 for reasoning; after the final block and normalization, it is 2.235 and 2.126, respectively. The observations establish a descriptive depth-dependent relationship in these prompt sets, rather than a uniform ordering of task categories. Limited held-out family counts prevent reliable primary family-bootstrap intervals, and insufficient length overlap prevents a joint matched-length comparison. We report a bounded exploratory experiment with its analysis code and limitations.

Research Square
Kakatiya University (IN)
Topic Modeling
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon