One Frozen Edit, Four Ways to Ask: Prompt-Conditioned Functional Effects in a Transformer

We investigate whether the measured functional response to a fixed transformer weight intervention changes when the input question is phrased differently. This research study presents a source-level reanalysis of an archived factorial experiment involving Qwen2.5-1.5B. The dataset contains 1,024 recorded measurements spanning 32 questions, two historical intervention directions, four wording variants, and four instruction templates. The primary outcome is a recorded functional-response metric called attenuation, rather than mathematical answer correctness. For the historical signed edit direction, the mean math-plus-format versus generic instruction attenuation contrast was +0.205542 under original wording, −0.020673 under reworded questions, −0.045819 under formal wording, and −0.002337 under terse wording. The paired original-minus-reworded contrast across the 32 question units averaged +0.226216, with a 95% question-bootstrap interval of [+0.184256, +0.274149], calculated using 20,000 bootstrap resamples. All 32 paired question-level contrasts were positive within the archived evaluation panel. A historical 90-degree rotated control direction exhibited a different response pattern, including a mean contrast of −0.023924 under original wording. These results demonstrate prompt-conditioned variation in a recorded functional-response measure within one historical model and intervention setting. They do not establish improved mathematical reasoning accuracy, general-purpose model repair, or the identification of a unique causal neural circuit. Important limitations include unmatched prompt lengths, changes in token position, historical intervention selection, and the absence of an independent model-side replication. The numerical results were independently recalculated from the archived data, but the underlying weight-editing experiment was not rerun. The publication includes a complete research-study manuscript, four main figures, two supplementary diagnostic figures, archived numerical results, question-level statistical summaries, and reproducibility scripts. The project aims to support transparent evaluation of transformer weight interventions and encourages independent scientific replication. Research status: Working preprint; not peer reviewed. Author: Harmain DastgirAffiliation: Independent ResearcherFunding: Self-funded; no external research grant reported.Competing interests: None declared.Research correspondence: [email protected] The author welcomes scientific collaboration, independent validation, and discussions concerning possible future research funding. The manuscript, original figures, and research data are provided under CC BY 4.0. The accompanying original Python software is separately licensed under the MIT License, as documented in the supplemental files. Harmain Dastgir — LinkedIn profile

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.24280040
Primary Topic
Topic Modeling
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

One Frozen Edit, Four Ways to Ask: Prompt-Conditioned Functional Effects in a Transformer

Harmain Dastgir
5 links
Zenodo (CERN European Organization for Nuclear Research)
Topic Modeling
preprint

One Frozen Edit, Four Ways to Ask: Prompt-Conditioned Functional Effects in a Transformer

Harmain Dastgir
preprint en

Abstract

We investigate whether the measured functional response to a fixed transformer weight intervention changes when the input question is phrased differently. This research study presents a source-level reanalysis of an archived factorial experiment involving Qwen2.5-1.5B. The dataset contains 1,024 recorded measurements spanning 32 questions, two historical intervention directions, four wording variants, and four instruction templates. The primary outcome is a recorded functional-response metric called attenuation, rather than mathematical answer correctness. For the historical signed edit direction, the mean math-plus-format versus generic instruction attenuation contrast was +0.205542 under original wording, −0.020673 under reworded questions, −0.045819 under formal wording, and −0.002337 under terse wording. The paired original-minus-reworded contrast across the 32 question units averaged +0.226216, with a 95% question-bootstrap interval of [+0.184256, +0.274149], calculated using 20,000 bootstrap resamples. All 32 paired question-level contrasts were positive within the archived evaluation panel. A historical 90-degree rotated control direction exhibited a different response pattern, including a mean contrast of −0.023924 under original wording. These results demonstrate prompt-conditioned variation in a recorded functional-response measure within one historical model and intervention setting. They do not establish improved mathematical reasoning accuracy, general-purpose model repair, or the identification of a unique causal neural circuit. Important limitations include unmatched prompt lengths, changes in token position, historical intervention selection, and the absence of an independent model-side replication. The numerical results were independently recalculated from the archived data, but the underlying weight-editing experiment was not rerun. The publication includes a complete research-study manuscript, four main figures, two supplementary diagnostic figures, archived numerical results, question-level statistical summaries, and reproducibility scripts. The project aims to support transparent evaluation of transformer weight interventions and encourages independent scientific replication. Research status: Working preprint; not peer reviewed. Author: Harmain DastgirAffiliation: Independent ResearcherFunding: Self-funded; no external research grant reported.Competing interests: None declared.Research correspondence: [email protected] The author welcomes scientific collaboration, independent validation, and discussions concerning possible future research funding. The manuscript, original figures, and research data are provided under CC BY 4.0. The accompanying original Python software is separately licensed under the MIT License, as documented in the supplemental files. Harmain Dastgir — LinkedIn profile

Zenodo (CERN European Organization for Nuclear Research)
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