Researchllmsource attributionparaphrasingdetection
Detection Models Struggle Distinguishing LLM-Generated Ideas
7.0
A new study submitted Dec. 4, 2025 evaluates state-of-the-art models' ability to distinguish human-generated from LLM-generated scientific ideas across successive paraphrasing stages. The authors report detection performance declines by an average of 25.4% after five paraphrases, that providing the research problem improves detection up to 2.97%, and that simplifying ideas into a non-expert style most degrades detectable LLM signatures.
Key Points
- 1Evaluate SOTA models' ability to distinguish human versus LLM-generated scientific ideas after paraphrasing
- 2Show detection degrades markedly—average 25.4% drop after five consecutive paraphrasing stages
- 3Recommend including research-problem context; improves detection up to 2.97%, aiding source attribution
Scoring Rationale
Novel experimental evaluation of attribution but limited scope and single-source preprint reduces practical generalizability across domains.
Sources
Public references used for this report.
Practice with real Logistics & Shipping data
90 SQL & Python problems · 15 industry datasets
Used by DS/ML engineers at top companies
High-Value Overnight OrdersEasyDelivered International ShipmentsMediumOn-Time Delivery Rate by CarrierHard
250 free problems · No credit card
See all Logistics & Shipping problems
