Detecting LLM-Generated Tokens in Human--LLM Coauthored Text

2026-07-23 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

A new method addresses the growing need for fine-grained detection of LLM-generated content within human-LLM coauthored text. Unlike existing document-level classification methods, this approach operates at the token level, the natural unit for modern language models. The core idea involves smoothing adjacent token detection scores to reduce variability, employing an adaptive Lepski-type rule to select bandwidth based on the local authorship structure. This method is simple to implement, requires no token-level labeled data for training, and theoretically achieves favorable mean square error performance. Empirically, it demonstrates strong performance against various baselines on both synthetic and realistic datasets, with a publicly accessible website implementing the techniques.

Key takeaway

For machine learning engineers developing tools for content moderation or authorship attribution, this token-level detection method offers a significant advancement. You should consider integrating this adaptive smoothing technique to precisely localize LLM-generated segments in mixed-authorship documents, moving beyond coarse document-level classifications. This approach can enhance the accuracy and utility of your detection systems without requiring extensive token-level labeled training data.

Key insights

A token-level detection method smooths adjacent scores using an adaptive Lepski-type rule to identify LLM-generated content in coauthored text.

Principles

• Fine-grained detection requires token-level analysis.
• Smoothing adjacent token scores reduces variability.
• Adaptive bandwidth selection improves localization.

Method

The method smooths adjacent token detection scores, applying an adaptive Lepski-type rule to select bandwidth based on local authorship structure. It requires no token-level labeled training data.

In practice

• Implement token-level smoothing for LLM detection.
• Utilize adaptive bandwidth selection for precision.
• Deploy the public website for practical use.

Topics

• LLM-Generated Text Detection
• Token-level Analysis
• Adaptive Smoothing
• Human-AI Collaboration
• Authorship Attribution
• Text Forensics

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Detecting LLM-Generated Tokens in Human--LLM Coauthored Text — A novel token-level method localizes LLM-generated content in human-AI coauthored text.
• Authorship Attribution in Multilingual Machine-Generated Texts — Multilingual authorship attribution for LLM-generated text is complex, with current monolingual methods showing limited cross-lingual transferability across diverse language families.
• Sem-Detect: Semantic Level Detection of AI Generated Peer-Reviews — AI-generated peer reviews converge semantically, while human reviews, even LLM-refined, retain unique ideas.
• Operation-Guided Progressive Human-to-AI Text Transformation Benchmark for Multi-Granularity AI-Text Detection — AI-text detectability is non-monotonic, influenced by edit operations and revision history, not just AI content proportion.
• Do we really need to detect LLM-generated text? — Masked word prediction by LLMs can distinguish AI-generated text from human text without explicit training.
• HyperparameterOmens at SemEval-2026 Task 13: Various approaches to detecting machine- generated code — Domain shift significantly degrades lexical model performance in machine-generated code detection tasks.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.