Step-Wise Refusal Dynamics in Autoregressive and Diffusion Language Models
cs.LG, cs.AI
Submitted: 2026-02-01
Updated: 2026-09-06
Comments: Accepted to the EMNLP 2026 Main Conference
Code: https://github.com/yueliu1999/FlipAttack
Project page: https://elironrahimi.github.io/sri-signal
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Terminology
Sources
- GPT-4 Technical Report
- Detecting Language Model Attacks with Perplexity
- Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks
- LLaDA2.0: Scaling Up Diffusion Language Models to 100B
- SDAR: A Synergistic Diffusion-AutoRegression Paradigm for Scalable Sequence Generation
- How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings
- FlipAttack: Jailbreak LLMs via Flipping
- HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal
- Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations
- Large Language Diffusion Models
- LLM Self Defense: By Self Examination, LLMs Know They Are Being Tricked
- A2D: Any-Order, Any-Step Safety Alignment for Diffusion Language Models
- Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks
- Interpretation Meets Safety: A Survey on Interpretation Methods and Tools for Improving LLM Safety
- Large Language Model Safety: A Holistic Survey
- Jailbreak Attack Initializations as Extractors of Compliance Directions
- Gemma: Open Models Based on Gemini Research and Technology
- DiffuGuard: How Intrinsic Safety is Lost and Found in Diffusion Large Language Models
- BERT Rediscovers the Classical NLP Pipeline
- The Devil behind the mask: An emergent safety vulnerability of Diffusion LLMs
Related papers
- Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
- AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
- Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
- Convergence issues in Relational Concept Analysis based on AOC-posets
- Beliefs Beyond Posteriors: Local-Consistency Optimisation for Bayesian Neural Networks
- Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks