Latent Washout? Geometric convergence and wording-style recoverability under shared continuations
Year: 2026
Type: preprint
Abstract: When two differently worded questions are followed by the same text, their hidden states may become closer. Does that also make the original wording style harder to identify? We study this question using 96 authored everyday/formal question pairs across six domains and three frozen language models: Gemma 4 E4B IT, Phi-4-mini-instruct, and Qwen3.5-4B. Each pair has the same answer instruction and exactly matched token counts under the evaluated model's tokenizer. We compare final hidden states after a short answer anchor and after a longer supplied task continuation, producing 1,152 measurements. Mean paired cosine distance decreases by 64.0%, 40.6%, and 54.4%, respectively. Yet logistic classifiers evaluated on held-out domains identify wording style after the task continuation in 135/192, 137/192, and 138/192 cases, compared with 96/192 for token-count-only controls. Classifier accuracy rises for Gemma and falls for Phi and Qwen; descriptive intervals for these changes all include zero. The results show that geometric convergence can coexist with detectable authored style differences after controlling instruction and token-count cues. Post-hoc checks also find broader geometric contraction and readily predictable wording labels. These results do not establish selective wording loss, semantic abstraction, exact prompt recovery, or unchanged information content. We document the earlier control failures that motivated the corrected experiment. Code and data: https://github.com/theomirzakhanian/latentwashoutartifacts/tree/main
Author: Mirzakhanian, Theo
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7211899750
Cites:
Cited by:
Related to: