GlitchToken: Detection and Remediation of Glitch Tokens in Large Language Models

GlitchToken is an open-source research project accepted at USENIX Security 2026, aiming to detect and remediate glitch token issues in large language models (LLMs). The project provides the systematic detection tool GlitchQuiz and lightweight remediation solution GlitchEdit to help developers improve model reliability and security.

Source Information:

• Original author/maintainer: kstanghere
• Project URL: GitHub Link
• Conference: USENIX Security 2026
• Paper Title: One Bad Token Spoils the Barrel: Assessment, Detection, and Remediation of Glitch Tokens in Large Language Models
• Release Time: June 2026

Glitch tokens are special tokens in the LLM vocabulary with 'discontinuity' in their embedding representations—their embedding space is discontinuous from adjacent tokens, leading to unpredictable outputs during model processing.

Causes: Stem from edge cases in tokenizer vocabulary construction (e.g., rare character combinations, encoding anomalies, preprocessing artifacts) and insufficient semantic learning during training.

Impacts: May lead to nonsensical text, hallucinations, or even malicious exploitation to bypass security mechanisms, affecting model reliability and safety.

GlitchQuiz is the glitch token detection tool provided by the project, using multi-dimensional geometric feature analysis:

• Core Idea: Embeddings of normal tokens should be continuously distributed with semantically similar tokens, while glitch tokens appear isolated.
• Detection Methods: Calculate embedding distance to neighbors, local geometric structure change rate, discontinuous points in the embedding manifold, and quantify the 'glitch degree'.
• Advantages: Universal (applicable to various Transformer models), provides visualization functions to help intuitively understand the distribution of the embedding space.

GlitchEdit is a lightweight remediation solution that can alleviate the problem without retraining the model:

• Core Idea: Adjust the embedding layer and output layer representations of glitch tokens to enable smooth transition with their semantic neighbors.
• Specific Steps: Identify semantic neighbors of glitch tokens → Fine-tune embedding vectors via interpolation/projection techniques while keeping the performance of other tokens unchanged.
• Advantages: Low computational cost (only modifies the local embedding matrix), can be quickly deployed in resource-constrained environments.

Practical application scenarios of glitch token technology include:

1. Production Environment Operation: Run GlitchQuiz regularly to detect abnormal tokens in time and prevent service quality issues.
2. Model Fine-tuning Phase: Evaluate whether new data generates new glitch tokens and solve problems early (especially suitable for enterprise-level domain adaptation).
3. Security-sensitive Scenarios: Proactively detect and remediate glitch tokens to improve model robustness and reduce the risk of adversarial attacks.

Features and usage of the GlitchToken project:

• Open Source & License: Apache 2.0 license, free to use in commercial/non-commercial scenarios.
• Module Design: Independent detection (GlitchQuiz) and remediation (GlitchEdit) modules, which can be used alone or in combination.
• Ease of Use: Clear code structure, detailed README (installation steps, usage, configuration), and provides experiment scripts and evaluation benchmarks for easy reproduction.

The GlitchToken project provides systematic tools and methods for LLM glitch token issues, helping developers maintain model quality through GlitchQuiz detection and GlitchEdit remediation.

As LLMs are applied in more critical fields, glitch token governance will become an important part of model operation. The open-source project lays a foundation for exploration in this field, and we look forward to the community further improving the solutions.