Answers are grounded in this paper and cite their sources.
Fixed-cohort study demonstrates compression detects model signatures in varied prompts, suggesting implications for prompt analysis.
Can a general-purpose compressor detect model-associated signatures across different prompt sets? This fixed-cohort study tests that hypothesis on very short continuations from 127 model endpoints/configurations. Five grouped-prompt folds compare symmetric Zstandard compression distance with an exact-response-length control, tie-aware chance calibration, character n-gram and stylometric baselines, family/backend decompositions, matched same-prompt controls, prefix budgets, and deduplication sensitivities. Raw compression MRR was 0.185 versus 0.083 for the length control in math and 0.489 versus 0.098 in code; hashed character n-gram TF-IDF was stronger at 0.347 and 0.687. The evidence supports fixed-cohort model-associated candidate-ranking structure beyond exact length sequences, not open-set, passage-level, causal, temporal, deployment, or forensic attribution. The deposit includes the paper and a complete non-sensitive repository snapshot with analysis reports, fold/model-level results, overlap audits, provenance hashes, matrix metadata and hashes, matched controls, figures, source code, and tests.
No takes yet. Share an insight, caveat, or question.
Gabriel Kahen (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: