The Instability of Safety: How Random Seeds and Temperature Expose Inconsistent LLM Refusal Behavior
cs.LG, cs.AI, cs.CL
Submitted: 2025-12-12
Updated: 2026-08-27
Comments: 17 pages, 7 figures, 9 tables. Code and data available at https://github.com/erikl2/safety-refusal-stability . v3: corrects dataset attribution (AdvBench+HarmBench, not BeaverTails), fixes label-derived statistics and CIs, revises the judge-noise analysis to a sensitivity framing; conclusions unchanged
Code: https://github.com/erikl2/safety-refusal-stability
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Terminology
Sources
- Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
- Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned
- The Llama 3 Herd of Models
- Influences of some families of error-correcting codes
- Language Models (Mostly) Know What They Know
- Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback
- HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal
- Red Teaming Language Models with Language Models
- The Effect of Sampling Temperature on Problem Solving in Large Language Models
- Universal and Transferable Adversarial Attacks on Aligned Language Models
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