Research map / Adversarial robustness and security / Attacks on text and graphs
Turning Bias into Bugs: Bandit-Guided Style Manipulation Attacks on LLM Judges
XIANGLIN YANG, Bryan Hooi, Gelei Deng, Tianwei Zhang, Jin Song Dong
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 |
|---|---|---|---|
| select stylistic edits with LinUCB bandit instead of random actions (ablation) | BITE · AlpacaEval and ArenaHard, 5 judges | best-so-far score: unbeaten rate 90.88%, avg score diff +0.105 | Improved |
| use diverse curated style action set instead of single rewrite action (ablation) | BITE · AlpacaEval and ArenaHard | best-so-far score: unbeaten rate 92.96%, avg score diff +0.393 | Improved |
| apply style attacks to objective fact-based questions | BITE · AlpacaEval pointwise objective | attack success rate (%): 96.0 ASR | Improved |
| use BITE instead of prompt injection and jailbreak baselines | LLM judges · Chatbot benchmarks | avg score improvement: Qwen 2.010 vs PAIR 1.284; R1 1.909 vs PAIR 1.856 | Mixed |
| swap helper model from Gemini-1.5-Flash-8B to GPT-4.1-Nano (ablation) | BITE · Qwen3-235B judge, pointwise | score improvement: 2.69 -> 2.87 | No clear effect |
| apply linear style control defense (ablation) | BITE attacked answers · AlpacaEval pointwise | avg judge score: 8.546 -> 8.571 | No clear effect |
| apply non-linear kernel style control defense (ablation) | BITE attacked answers · AlpacaEval pointwise | avg judge score: 8.546 -> 8.477 | No clear effect |
| randomize judge prompt among paraphrases (ablation) | BITE · Judge defense | score lift: +3.11 -> +1.77 | Hurt |
| rewrite answers to remove style before judging (ablation) | BITE attacked answers · AlpacaEval pointwise | avg judge score: 8.55 -> 6.67 | Hurt |
| LLM-based detector on attacked answers (ablation) | BITE vs prompt injection · Detection | detection rate: 100% for injection vs 0% for BITE | Hurt |
| detect manipulation via meta-judge on judge explanations (ablation) | BITE · ArenaHard | detection rate (%): 6.40 -> 2.80 | Hurt |
| transfer attacks optimized on one judge to other judges | BITE · Transfer ASR across 5 judges | transfer ASR (%): 100 -> roughly 8.9-30.6 off-diagonal | Hurt |
Closest papers by meaning
- Reverse Engineering Human Preferences with Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Look Before You Leap: Enhance Attention and Vigilance Regarding Harmful Content with GuidelineLLM (AAAI 2025 · similarity 0.87)
- Gradient Cuff: Detecting Jailbreak Attacks on Large Language Models by Exploring Refusal Loss Landscapes (NeurIPS 2024 · similarity 0.87)
- Cannot See the Forest for the Trees: Invoking Heuristics and Biases to Elicit Irrational Choices of LLMs (ICML 2025 · similarity 0.87)
- An LLM can Fool Itself: A Prompt-Based Adversarial Attack (ICLR 2024 · similarity 0.86)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.86)
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.
