From Adinkras to GPT-2: Testing an Adinkra-Inspired Low-Rank Hypothesis

This paper reports a descriptive pilot testing whether a low-rank representation fitted to pairwise causal interactions in GPT-2 Small can predict triple interactions. Adinkras supplied the intellectual motivation; neither supersymmetry nor an Adinkra algebra was tested. The pilot comprised 384 donor conditions over 128 prompts and eight name-pair clusters, evaluated using six selected attention heads and six independently selected control heads. For each condition, a rank-one predictor used 15 pair interactions and one calibration triple to predict the remaining 19 triple interactions. Under single-cell calibration, the predictor did not demonstrate an average predictive advantage over zero. For the primary heads and ordinary readout, cluster-balanced mean normalized held-out error was 0.794926, compared with 0.783085 for zero prediction. The predictor improved on zero in 190 of 384 conditions and four of eight name-pair clusters. These findings do not establish statistical equivalence. Exploratory diagnostics showed modest directional alignment, but a spectral initialization ablation achieved a slightly higher median cosine. Residual recurrence was also observed, without consistent primary-head superiority over controls. Unequal realized intervention doses limit circuit-specific interpretations. Oracle scaling calculations use held-out information, including possible sign reversal, and are reported only as diagnostic lower bounds on loss. The accompanying reproduction package includes the manuscript, archived analysis outputs, all 10,752 baseline rows, numerical tables, figures, source identities, and a checksum-verifying CPU analysis script. It supports reproduction of reported aggregates from archived outputs; it does not contain the complete original per-case derivative archive. This is a bounded negative pilot result for the tested predictive procedure. The 240 reserved conditions remain unopened, and no reserved-set confirmation is claimed.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24298950
Primary Topic
Explainable Artificial Intelligence (XAI)
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

From Adinkras to GPT-2: Testing an Adinkra-Inspired Low-Rank Hypothesis

Joshua A. Duran
1 links
Zenodo (CERN European Organization for Nuclear Research)
Explainable Artificial Intelligence (XAI)
preprint

From Adinkras to GPT-2: Testing an Adinkra-Inspired Low-Rank Hypothesis

Joshua A. Duran
preprint en

Abstract

This paper reports a descriptive pilot testing whether a low-rank representation fitted to pairwise causal interactions in GPT-2 Small can predict triple interactions. Adinkras supplied the intellectual motivation; neither supersymmetry nor an Adinkra algebra was tested. The pilot comprised 384 donor conditions over 128 prompts and eight name-pair clusters, evaluated using six selected attention heads and six independently selected control heads. For each condition, a rank-one predictor used 15 pair interactions and one calibration triple to predict the remaining 19 triple interactions. Under single-cell calibration, the predictor did not demonstrate an average predictive advantage over zero. For the primary heads and ordinary readout, cluster-balanced mean normalized held-out error was 0.794926, compared with 0.783085 for zero prediction. The predictor improved on zero in 190 of 384 conditions and four of eight name-pair clusters. These findings do not establish statistical equivalence. Exploratory diagnostics showed modest directional alignment, but a spectral initialization ablation achieved a slightly higher median cosine. Residual recurrence was also observed, without consistent primary-head superiority over controls. Unequal realized intervention doses limit circuit-specific interpretations. Oracle scaling calculations use held-out information, including possible sign reversal, and are reported only as diagnostic lower bounds on loss. The accompanying reproduction package includes the manuscript, archived analysis outputs, all 10,752 baseline rows, numerical tables, figures, source identities, and a checksum-verifying CPU analysis script. It supports reproduction of reported aggregates from archived outputs; it does not contain the complete original per-case derivative archive. This is a bounded negative pilot result for the tested predictive procedure. The 240 reserved conditions remain unopened, and no reserved-set confirmation is claimed.

Zenodo (CERN European Organization for Nuclear Research)
Explainable Artificial Intelligence (XAI)
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