Papers with Safety alignment
Interpretable Safety Alignment via SAE-Constructed Low-Rank Subspace Adaptation (2026.acl-long)
Copied to clipboard
| Challenge: | Prior work has shown that safety behaviors are governed by low-rank structures . Low-Rank Adaptation (LoRA) consistently underperforms full fine-tuning and reinforcement learning on safety benchmarks . |
| Approach: | They propose a safety alignment system that disentangles safety-relevant directions into monosemantic features and constructs an interpretable safety subspace from SAE directions. |
| Outcome: | Empirically, the proposed model achieves 99.6% safety rates across multiple model families and scales . low-rank Adaptation consistently underperforms full fine-tuning and reinforcement learning on safety benchmarks compared with previous methods . |
Safety-Utility Conflicts Are Not Global: Surgical Alignment via Head-Level Diagnosis (2026.acl-long)
Copied to clipboard
| Challenge: | Existing mitigation strategies rely on global gradient geometry to resolve alignment conflicts . however, they overlook Modular Heterogeneity within Transformers, resulting in suboptimal trade-offs . Conflict-Aware Sparse Tuning (CAST) combines head-level diagnosis with sparse fine-tuning . |
| Approach: | They propose a framework that integrates head-level diagnosis with sparse fine-tuning to address this limitation. |
| Outcome: | The proposed framework integrates head-level diagnosis with sparse fine-tuning to reduce alignment conflicts in LLMs. |
Adversary-Aware DPO: Enhancing Safety Alignment in Vision Language Models via Adversarial Training (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Adversary-aware DPO (ADPO) is a training framework that explicitly considers adversary. |
| Approach: | a new framework integrates adversarial training into a pre-trained large language model to enhance safety alignment . adversary-aware DPO provides a framework that explicitly considers adversary . |
| Outcome: | a new training framework outperforms baselines in safety alignment and general utility of large language models. |
Preventing Safety Drift in Large Language Models via Coupled Weight and Activation Constraints (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing defenses constrain either weights or activations in isolation, without considering their coupled effects on safety. |
| Approach: | They propose a weight-activation constraint that enforces a precomputed safety subspace on weight updates and applies regularization to safety-critical features identified by sparse autoencoders. |
| Outcome: | The proposed model outperforms baselines even under high harmful data ratios. |
Probing the Safety Robustness of LLMs in Latent Space (2026.acl-long)
Copied to clipboard
Tianle Gu, Kexin Huang, Zongqi Wang, Yixu Wang, Jie Li, Xin Wang, Yang Yao, Yujiu Yang, Yan Teng, Yingchun Wang
| Challenge: | Despite substantial progress in safety alignment techniques, aligned large language models can still produce unsafe responses under minor internal perturbations. |
| Approach: | They introduce Activation Steering Attack (ASA) and leverage the Negative Log-Likelihood (NLL) as a diagnostic signal to probe the local sensitivity of safety behaviors in latent space. |
| Outcome: | The proposed method is model-agnostic and supervision-free, enabling a general and reproducible diagnostic metric for analyzing safety robustness. |
Health-ORSC-Bench: A Benchmark for Measuring Over-Refusal and Safety Completion in Health Context (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing safety alignment benchmarks fail to evaluate Safe Completion: the model’s ability to maximise helpfulness on dual-use or borderline queries without crossing into actionable harm. |
| Approach: | They propose a large-scale benchmark to measure Over-Refusal and Safe Completion quality in healthcare. |
| Outcome: | The framework evaluates 30 state-of-the-art LLMs including GPT-5 and Claude-4. |