Work

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

Year: 2026
Type: preprint
Abstract: Fluid Tensors — Paper 1 is an open, reproducible research preprint examining how question wording and instruction format change the measured response to one frozen transformer weight intervention. The experiments use a single Qwen2.5-1.5B-Instruct model/edit setting and measure final-token response attenuation, not answer accuracy or general-purpose model repair. The original archival study analyzed 1,024 response-geometry measurements from 32 GSM8K question identities, two historical intervention directions, four wording variants, and four instruction templates. The original signed-edit original-versus-reworded paired contrast was +0.226216, with a question-level bootstrap interval of approximately [+0.184256, +0.274149]. NEW GPU CONFIRMATION — NOVEMBER 18, 2026: Using the same pinned Qwen2.5-1.5B-Instruct checkpoint and the same frozen rank-four layer-19 intervention, we executed a 128-cell original contrast replay and evaluated 64 additional GSM8K train question identities disjoint from the frozen historical registry. The 1,152 recorded cells included fresh raw-prompt, explicit token-count-matched and norm-matched orthogonal-edit comparisons. All numerical-quality checks passed; 128/128 replay measurements agreed with the archived values. On 64 new questions, the token-count-matched signed-edit paired interaction mean was +0.193491, with a 95% question-bootstrap interval of [+0.166896, +0.219002]; 62/64 question-level effects were positive. The equal-norm orthogonal-control mean was +0.004060, interval [−0.004614, +0.012629]. Under raw prompts, the signed mean was +0.229246, showing that formatting and matching choices influence magnitude. The outcome is last-token output-logit response attenuation, NOT how often the model answered a mathematics problem correctly. The additional questions came from GSM8K train, so model-pretraining exposure cannot be excluded. Token-count matching inserted explicit newline token IDs and does not eliminate all semantic, positional or formatting confounds. This confirms a narrowly defined within-model response pattern, not independent-model replication, an isolated neural circuit, improved math reasoning accuracy or general-purpose model repair. Version 3 preserves all prior published source materials and adds the dated 17-page v1.2 reading copy and a fully hashed confirmation kit with GPU logs, figures, detailed methods, checker scripts, and limitations. Author: Harmain Dastgir, Independent Researcher. The work remains a working preprint, not peer reviewed. The manuscript and research data remain CC BY 4.0; third-party code/model/dataset licenses remain documented in the archive. See https://github.com/Harmain11/fluid-tensors and https://harmain11.github.io/fluid-tensors/papers/one-frozen-edit.html.
Language: English
Open Access status: green
OpenAlex ID: W7233963769
Cites:
Cited by:
Related to: 10