Source Attribution of Synthetic Data Hits 98.7% Accuracy but Falls to 29% After Style Rewriting
Joss Armstrong · hf · 2026-10-07
- This study tests two questions: how reliably the provenance of model-generated text can be recovered, and whether provenance helps pick better training data.
- On financial-risk text, generator attribution is 98.7% accurate on original passages, drops to 53.1% after paraphrasing and 29.0% after style rewriting; generated-vs-human detection stays near perfect against the tested human set.
- In three rounds of generation and retraining, a source-based selection rule and a reference-model score rule pick different examples, yet no stable difference in model degradation is detected.
- Takeaway: source identity and training-utility are separate problems; neither the provenance score nor the tested proxy suffices to predict recursive-training outcomes.
More from Research
- RaiSim is back: v2.8.0 brings ultra-fast accurate CPU simulation plus photorealistic rendering — ChongZzZhang · 2026-10-07
- DAEDALUS bootstraps agent memory from self-generated tasks, +15.9 points success rate — illuin · 2026-10-07
- Storing model-native numerical memory outside a frozen LLM: replicated from Qwen to Mistral, 127/128 Top-1 — Nearby_Indication474 · 2026-10-07
- Mathematicians release skill files to de-slop LLM-written math papers — IAmTimNguyen · 2026-10-07
- Audit finds 206 buggy graders in Zapier's AutomationBench; fixed scores flip 27.9% of runs — rohanpaul_ai · 2026-10-07
- Cohere Labs research suggests AI has yet to impact nearly half of workforce occupations — Cohere_Labs · 2026-10-07