Research map / Large language models / Language, safety and interpretability
GlitchMiner: Mining Glitch Tokens in Large Language Models via Gradient-based Discrete Optimization
Zihui Wu, Haichang Gao, Ping Wang, Shudong Zhang, Zhaoxiang Liu, Shiguo Lian
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GlitchMiner instead of Magikarp at budget 2000 | GlitchMiner · 10-LLM average | Detected@2000: 885.6 -> 980.1 | Improved |
| use GlitchMiner instead of Magikarp at budget 1000 | GlitchMiner · Llama-3.1-8B-Instruct | Detected@1000: 664 -> 568 | Hurt |
| use GlitchMiner instead of Magikarp | GlitchMiner · Llama-2-7B-chat-hf | Detected@2000: 186 -> 532 | Improved |
| add local search (restrict to embedding neighborhood) (ablation) | GlitchMiner · 10 LLMs | Detected@1000 | Improved |
| increase neighborhood size K (ablation) | GlitchMiner · 10 LLMs | Detected@1000 | Hurt |
| increase batch size B (ablation) | GlitchMiner · 10 LLMs | Detected@1000 | No clear effect |
| random initialization instead of min l2-norm token (ablation) | GlitchMiner · 10 LLMs | Detected@1000 | No clear effect |
Closest papers by meaning
- Language Models Can Predict Their Own Behavior (NeurIPS 2025 · similarity 0.86)
- Large language models suffer from a curse of ambiguity (NeurIPS 2026 · similarity 0.85)
- Systematic Outliers in Large Language Models (ICLR 2025 · similarity 0.85)
- Interpreting the Repeated Token Phenomenon in Large Language Models (ICML 2025 · similarity 0.85)
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.85)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
