The study investigates how prompt template choices during Knowledge Distillation (KD) affect safety alignment in language models. It finds that using chat templates during KD degrades safety alignment, making models more compliant with harmful queries, while non-chat templates better preserve the base model’s internal representations. These effects are observed across LLaMA, Gemma, and Qwen families on multiple safety benchmarks.
arXiv:2606. 19168v1 Announce Type: new Abstract: To achieve deeper safety alignment for large language models (LLMs), recent efforts have studied how to push safety interventions earlier into the pretraining stage, primarily by filtering unsafe data or rewriting it into safer forms.
arXiv:2511. 00382v2 Announce Type: replace-cross Abstract: Organizations increasingly adapt Large Language Models (LLMs) from public repositories such as HuggingFace to downstream tasks.
Deliberative alignment: reasoning enables safer language models Introducing our new alignment strategy for o1 models, which are directly taught safety specifications and how to reason over them.
arXiv:2606. 02530v1 Announce Type: new Abstract: Aligning Large Language Models (LLMs) with human values often degrades their general capabilities, termed the alignment tax.
SHARD is a self‑reframing distillation technique designed to enhance the safe‑helpfulness of large language models. It rewrites sensitive prompts to reveal benign intent, reframes the model’s original responses into safer, more helpful versions, and then fine‑tunes the model on these self‑reframed outputs. Experiments on DNA and the English subset of LINGUASAFE show that SHARD improves helpfulness across various model families while maintaining safety, and performs competitively with distillation from larger teacher models.
EOPSA (Efficient On-Policy Self-Distilled Safety Alignment) addresses inefficiencies in On-Policy Self-Distillation (OPSD) for safety alignment by focusing training on safety-critical tokens. It introduces Adaptive Rollout Scheduling, which limits generation length based on a Teacher Rescue Rate metric, and Selective Distillation, which filters out safety-neutral tokens to concentrate gradient updates on safety-pivotal transitions. Experiments on models up to 32B parameters show that EOPSA reduces rollout computation by about 50% and backpropagates through only roughly 2% of tokens, outperforming full-token distillation baselines in safety compliance and reasoning retention.
arXiv:2607. 00572v1 Announce Type: new Abstract: Understanding how aligned LLMs internally represent safety is critical for diagnosing alignment vulnerabilities, as it explains why jailbreaks succeed and informs the design of robust alignment strategies.
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.
arXiv:2506. 08473v4 Announce Type: replace Abstract: Fine-tuning large language models (LLMs) improves performance but introduces critical safety vulnerabilities: even minimal harmful data can severely compromise safety measures.
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:2604. 09544v2 Announce Type: replace-cross Abstract: Large language models (LLMs) undergo alignment training to avoid harmful behaviors, yet the resulting safeguards remain brittle: jailbreaks routinely bypass them, and fine-tuning on narrow domains can induce ``emergent misalignment'' that generalizes broadly.
arXiv:2607. 02914v1 Announce Type: new Abstract: Large language models (LLMs) have demonstrated remarkable capabilities across diverse applications, yet ensuring their simultaneous safety, helpfulness, and trustworthiness remains a persistent challenge.