The Erosion of LLM Signatures: Can We Still Distinguish Human and LLM-Generated Scientific Ideas After Iterative Paraphrasing?

TL;DR

This work tackles the problem of distinguishing human- vs LLM-generated scientific ideas after iterative paraphrasing. It builds a large, multi-stage paraphrase dataset by extracting research problems, generating ideas with six LLMs, and applying five paraphrase stages, then evaluates multiple classifiers and embeddings. The key finding is a substantial erosion of detectable LLM signatures across paraphrase stages, with an average decline of 25.4% in detection performance, though incorporating the research problem as context can yield notable gains up to 2.97%. The study highlights the limitations of surface-level linguistic cues for attribution and suggests directions for stronger signal extraction through RP-idea integration and deeper reasoning traces.

Abstract

With the increasing reliance on LLMs as research agents, distinguishing between LLM and human-generated ideas has become crucial for understanding the cognitive nuances of LLMs' research capabilities. While detecting LLM-generated text has been extensively studied, distinguishing human vs LLM-generated scientific idea remains an unexplored area. In this work, we systematically evaluate the ability of state-of-the-art (SOTA) machine learning models to differentiate between human and LLM-generated ideas, particularly after successive paraphrasing stages. Our findings highlight the challenges SOTA models face in source attribution, with detection performance declining by an average of 25.4\% after five consecutive paraphrasing stages. Additionally, we demonstrate that incorporating the research problem as contextual information improves detection performance by up to 2.97%. Notably, our analysis reveals that detection algorithms struggle significantly when ideas are paraphrased into a simplified, non-expert style, contributing the most to the erosion of distinguishable LLM signatures.

Figures (7)

• Figure 1: Idea Generation and Paraphrasing Workflow: The process begins with extracting the Research Problem from papers and then generate corresponding scientific ideas using six different LLMs. Both human and LLM-generated ideas are first summarized and subsequently paraphrased across five stages using four distinct paraphrasing techniques (To reduce visual clutter and redundancy, we abstracted Stages 3 and 4, as they represent similar paraphrasing strategies).
• Figure 2: Integrated Gradients Visualization: Green highlights words that contribute to classifying the text as human-written, while red highlights words that push the classification toward LLM-generated content. The overall text is LLM-idea-summarized
• Figure 3: F1-score evaluated across different training and testing stages. The purple dashed line represents the highest performance achieved at each stage, pointing to the overall declining performance trend across the models (clockwise from top left) BERT, BigBird , Stella + FFNN (RP + Idea) (c), and Stella + FFNN (Idea Only) (d).
• Figure 4: Visualization of discriminative features between human and LLM-generated ideas. (a) Four different text embedding representations illustrate the decreasing discriminability as we progress from Stage 1 to Stage 5. (b) Word Mover's Distance also shows a declining trend, indicating reduced differentiation between human and LLM-generated ideas over stages.
• Figure 5: Visualization of the training (solid line) and validation (dashed line) loss curves for the MiniLM + FFNN model across the first 25 epochs, providing insights into learning dynamics and model convergence