The paper investigates why safety‑aligned large language models still over‑refuse harmless queries, attributing the issue to overlapping representations in deeper layers and imbalanced gradient coupling that favors refusal. It proposes GOAL, a framework that orthogonalizes over‑refusal gradients against a low‑rank refusal subspace and adds contrastive learning to sharpen harmful‑vs‑over‑refusal representation separation. Experiments show GOAL reduces over‑refusal while maintaining safety and general performance.
arXiv:2608. 11583v1 Announce Type: new Abstract: Safety alignment in large language models is often treated as a distributed property of the entire network, yet its practical brittleness suggests that refusal behavior may be concentrated in a smaller set of parameters.
Warning: This paper studies stereotypes and biases, and contains potentially disturbing examples, used for illustration purposes only. Our findings should not be interpreted as an argument against alignment.
arXiv:2607. 08883v1 Announce Type: new Abstract: Behavioral alignment in large language models often masks fragile internal safety representations.
arXiv:2605.01913v2 Announce Type: replace-cross Abstract: Fine-tuning safety-aligned language models for downstream tasks often leads to substantial degradation of refusal behavior, making models vul...
arXiv:2611.07216v1 Announce Type: cross Abstract: Large vision-language models (LVLMs) are vulnerable to compositional harm: an image and a prompt that are benign in isolation become harmful through...
The paper investigates how large language models balance helpfulness and safety by refusing harmful queries while responding to benign ones. It decomposes safety-tuning responses into a boilerplate refusal statement and a rationale, finding that the statement causes false refusals by relying on superficial cues. Training on rationales alone reduces false refusals without compromising safety performance, suggesting that fine‑grained safety supervision is essential for better alignment.
The paper introduces SimpleLoRIF, a scalable gradient‑embedding technique that pinpoints fine‑tuning examples whose gradients align with unsafe behavior in large language models. By applying this method, the authors create SafeTulu, a version of Tulu 3 with the 20 % of most unsafe‑aligned examples removed, which reduces GCG attack success by up to 22 % across several LLMs while largely preserving or improving capability. The study also highlights previously under‑studied safety‑degrading data sources such as tool‑execution failures and narrative arithmetic.
arXiv:2609.36862v1 Announce Type: cross Abstract: Fine-tuning-as-a-service lets users adapt a safety-aligned language model to their own data, but it also creates a harmful fine-tuning attack surface...
The paper introduces "cunning questions"—non‑safety prompts that contain misleading premises or subtle inconsistencies—to train large language models (LLMs) to scrutinize underlying intent and assumptions. Experiments show that incorporating these questions improves robustness against out‑of‑distribution jailbreak attacks and enhances subsequent safety fine‑tuning, achieving a new state‑of‑the‑art reduction in mean ASR from 17.40% to 15.05% across nine backbone–benchmark combinations. The authors argue that this training fosters vigilance, enabling models to prioritize safety judgments before engaging in harmful planning.
The paper introduces Weighted Refusal Direction Optimization (WRDO), a method that constructs a composite refusal direction by learning aggregation weights over per‑prompt refusal directions. WRDO aims to strengthen refusal interventions in aligned language models by allowing heterogeneous harmful prompts to contribute unequally. Experiments on three open‑weight models, using guard judges and HarmBench, demonstrate that weighted aggregation outperforms single‑direction methods and can also be used in reverse to restore refusal behavior in a damaged checkpoint.
The paper investigates how safety‑aligned large language models (LLMs) can still be tricked by jailbreaks or over‑refuse benign requests. By analyzing residual‑stream updates in decoder‑only Transformers, the authors find that harmfulness is more readily decoded from attention updates—especially in early layers—while the MLP component is primarily responsible for writing refusal responses. Post‑training enhances MLP‑driven refusal writing without improving harmfulness decodability, and causal interventions confirm that downstream MLP updates are crucial for the refusal shift caused by upstream attention changes. Building on these insights, the authors propose Sentinel, a system that combines an attention‑derived perception score with an MLP‑derived decision score to better detect jailbreaks and over‑refusals across multiple models.
The paper investigates how different post‑training methods—supervised fine‑tuning, reasoning‑augmented fine‑tuning, and preference optimization (ORPO)—affect the internal computation of refusal behavior in language models. Experiments on Llama‑3.1‑8B, Gemma‑2‑9B, and Qwen3‑8B show that reasoning‑augmented training consistently creates a distinct refusal computation across models, while the architecture influences the internal structure and steerability of refusal. None of the studied methods simultaneously achieve a distributed refusal mechanism, preserve general capability, and allow easy corrective edits, indicating that current post‑training approaches are not a fully reliable defense for safety-critical applications.