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

2026-07-23 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

A new token-level detection method addresses the growing need for fine-grained localization of LLM-generated content within human-AI coauthored documents. Unlike existing document-level classification approaches, this method identifies specific parts of text produced by large language models. The core innovation involves smoothing adjacent token detection scores to reduce variability, coupled with an adaptive Lepski-type rule that selects bandwidth based on the local authorship structure. This approach is straightforward to implement and does not require token-level labeled data for training. Theoretically, the method demonstrates favorable mean square error performance in signal estimation. Empirically, it shows strong performance against various baselines across both synthetic and realistic datasets. A publicly accessible website has been deployed to implement these detection capabilities.

Key takeaway

For NLP Engineers or content integrity analysts working with human-LLM coauthored text, this token-level detection method offers a precise way to identify AI-generated segments. You can now localize specific LLM contributions, moving beyond document-level classifications. This capability is vital for verifying content authenticity and managing risks associated with AI-generated text. Consider integrating this method, which requires no labeled training data, into your content analysis workflows to enhance detection granularity.

Key insights

A novel token-level method localizes LLM-generated content in human-AI coauthored text.

Principles

• Token-level analysis is crucial for mixed-authorship detection.
• Smoothing adjacent scores reduces variability in detection.
• Adaptive bandwidth selection improves local authorship structure detection.

Method

Smooth adjacent token detection scores, then apply an adaptive Lepski-type rule to select bandwidth based on local authorship structure.

In practice

• Implement token-level detection without labeled training data.
• Utilize the public website for LLM-generated token detection.

Topics

• LLM-Generated Token Detection
• Token-Level Analysis
• Human-AI Coauthorship
• Adaptive Lepski Rule
• Content Authenticity

Best for: Research Scientist, AI Scientist, NLP Engineer, AI Security Engineer

Related on AIssential

• Detecting LLM-Generated Tokens in Human--LLM Coauthored Text — A token-level detection method smooths adjacent scores using an adaptive Lepski-type rule to identify LLM-generated content in coauthored text.
• How to Tell If Something Was Written by AI (And Why It’s Getting Harder) — Advanced LLMs now produce AI-generated content nearly indistinguishable from human writing.
• Improving the Faithfulness of LLM-based Abstractive Summarization with Span-level Unlikelihood Training — Unlikelihood training with span-level hallucination annotations significantly improves LLM abstractive summarization faithfulness.
• Reddit is using LLMs to solve a problem LLMs largely created — Platforms are deploying LLMs to counter AI-generated spam, recognizing that human moderation is still essential for comprehensive content safety.
• Beating the Style Detector: Three Hours of Agentic Research on the AI-Text Arms Race — Agentic research rapidly reproduces and extends NLP studies, demonstrating LLMs' ability to evade AI-text detection.
• How to Detect AI-Generated Text Using Signs of AI Writing — AI text detection can leverage community-curated signs to combat the proliferation of LLM-generated content.

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 Takara TLDR - Daily AI Papers.