Research map / Large language models / Language, safety and interpretability
Beyond the Prompt: Leveraging Pre-Decoding States for Jailbreak Detection in dLLMs
Adam Hazimeh, Amel Abdelraheem, Ke Wang, Mariam Salman, Ljiljana Dolamic, Gérôme Bovet, Pascal Frossard
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes detecting jailbreaks in diffusion LLMs by probing hidden states of masked response positions before any token is decoded.
Problem
Jailbreak prompts are hard to identify from the prompt alone in diffusion language models, whose generation exposes hidden states over masked future response positions before tokens are finalized.
Approach
Uses the pre-decoding hidden states over masked response slots of dLLMs as a detection signal, exploiting a detection surface absent from autoregressive decoding, so that the model's emerging compliance can be read before output is committed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fuse prompt and pre-decoding response classifier scores (ablation) | ReFuse · Jailbreak suite avg, LLaDA-8B | ASR: 11.27 -> 3.31 | Improved |
| add response-view classifier to prompt-view classifier (DIJA) (ablation) | ReFuse · DIJA, LLaDA-8B | ASR: 63.43 -> 8.57 | Improved |
| add prompt-view classifier to response-view classifier (ablation) | ReFuse · PAIR, LLaDA-8B | ASR: 30.00 -> 0.56 | Improved |
| use calibrated score fusion instead of hard OR of classifiers (ablation) | ReFuse · LLaDA-8B | XSTest refusal: 20.00 -> 15.56 | Improved |
| use calibrated score fusion instead of hard OR of classifiers (ablation) | ReFuse · Jailbreak suite avg, LLaDA-8B | ASR: 3.74 -> 3.31 | Improved |
| add the fused detector versus prompt-only (ablation) | ReFuse · benign utility benchmarks, LLaDA-8B | benign refusal: 0.37 -> 0.74 | Hurt |
| use ReFuse instead of Self-Reminder | ReFuse · XSTest, LLaDA-8B | benign refusal: 20.56 -> 15.56 | Improved |
| use ReFuse instead of undefended model | ReFuse · Jailbreak suite avg, Dream-7B | ASR: 44.27 -> 13.23 | Improved |
| use fused detector on ReNeLLM with Dream-7B (ablation) | ReFuse · ReNeLLM, Dream-7B | ASR: 88.89 -> 87.22 | No clear effect |
| transfer probes from the other LLaDA variant (ablation) | ReFuse · Jailbreak suite avg, LLaDA-8B | ASR: 3.31 -> 4.31 | No clear effect |
| use ReFuse instead of DiffuGuard | ReFuse · Jailbreak suite avg, LLaDA-8B | ASR: 46.09 -> 3.31 | Improved |
Closest papers by meaning
- Why Jailbreaks Succeed in Diffusion Language Models: An Energy Landscape Analysis (NeurIPS 2026 · similarity 0.89)
- MaskForge: Structure-Aware Adaptive Attacks for Jailbreaking Diffusion Large Language Models (NeurIPS 2026 · similarity 0.88)
- An Interpretable N-gram Perplexity Threat Model for Large Language Model Jailbreaks (ICML 2025 · similarity 0.88)
- Gradient Cuff: Detecting Jailbreak Attacks on Large Language Models by Exploring Refusal Loss Landscapes (NeurIPS 2024 · similarity 0.87)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.87)
- Best-of-N Jailbreaking (NeurIPS 2025 · similarity 0.87)
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.
