Daily Papers

Training large language models to follow instructions makes them perform better on a wide range of tasks, generally becoming more helpful. However, a perfectly helpful model will follow even the most malicious instructions and readily generate harmful content. In this paper, we raise concerns over the safety of models that only emphasize helpfulness, not safety, in their instruction-tuning. We show that several popular instruction-tuned models are highly unsafe. Moreover, we show that adding just 3% safety examples (a few hundred demonstrations) in the training set when fine-tuning a model like LLaMA can substantially improve their safety. Our safety-tuning does not make models significantly less capable or helpful as measured by standard benchmarks. However, we do find a behavior of exaggerated safety, where too much safety-tuning makes models refuse to respond to reasonable prompts that superficially resemble unsafe ones. Our study sheds light on trade-offs in training LLMs to follow instructions and exhibit safe behavior.

The risk of harmful content generated by large language models (LLMs) becomes a critical concern. This paper presents a systematic study on assessing and improving LLMs' capability to perform the task of course-correction, \ie, the model can steer away from generating harmful content autonomously. To start with, we introduce the C^2-Eval benchmark for quantitative assessment and analyze 10 popular LLMs, revealing varying proficiency of current safety-tuned LLMs in course-correction. To improve, we propose fine-tuning LLMs with preference learning, emphasizing the preference for timely course-correction. Using an automated pipeline, we create C^2-Syn, a synthetic dataset with 750K pairwise preferences, to teach models the concept of timely course-correction through data-driven preference learning. Experiments on 2 LLMs, Llama2-Chat 7B and Qwen2 7B, show that our method effectively enhances course-correction skills without affecting general performance. Additionally, it effectively improves LLMs' safety, particularly in resisting jailbreak attacks.

Red-teaming, or identifying prompts that elicit harmful responses, is a critical step in ensuring the safe and responsible deployment of large language models (LLMs). Developing effective protection against many modes of attack prompts requires discovering diverse attacks. Automated red-teaming typically uses reinforcement learning to fine-tune an attacker language model to generate prompts that elicit undesirable responses from a target LLM, as measured, for example, by an auxiliary toxicity classifier. We show that even with explicit regularization to favor novelty and diversity, existing approaches suffer from mode collapse or fail to generate effective attacks. As a flexible and probabilistically principled alternative, we propose to use GFlowNet fine-tuning, followed by a secondary smoothing phase, to train the attacker model to generate diverse and effective attack prompts. We find that the attacks generated by our method are effective against a wide range of target LLMs, both with and without safety tuning, and transfer well between target LLMs. Finally, we demonstrate that models safety-tuned using a dataset of red-teaming prompts generated by our method are robust to attacks from other RL-based red-teaming approaches.

This study exposes the safety vulnerabilities of Large Language Models (LLMs) in multi-turn interactions, where malicious users can obscure harmful intents across several queries. We introduce ActorAttack, a novel multi-turn attack method inspired by actor-network theory, which models a network of semantically linked actors as attack clues to generate diverse and effective attack paths toward harmful targets. ActorAttack addresses two main challenges in multi-turn attacks: (1) concealing harmful intents by creating an innocuous conversation topic about the actor, and (2) uncovering diverse attack paths towards the same harmful target by leveraging LLMs' knowledge to specify the correlated actors as various attack clues. In this way, ActorAttack outperforms existing single-turn and multi-turn attack methods across advanced aligned LLMs, even for GPT-o1. We will publish a dataset called SafeMTData, which includes multi-turn adversarial prompts and safety alignment data, generated by ActorAttack. We demonstrate that models safety-tuned using our safety dataset are more robust to multi-turn attacks. Code is available at https://github.com/renqibing/ActorAttack.

Large language models (LLMs) have transformed the way we access information. These models are often tuned to refuse to comply with requests that are considered harmful and to produce responses that better align with the preferences of those who control the models. To understand how this "censorship" works. We use representation engineering techniques to study open-weights safety-tuned models. We present a method for finding a refusal--compliance vector that detects and controls the level of censorship in model outputs. We also analyze recent reasoning LLMs, distilled from DeepSeek-R1, and uncover an additional dimension of censorship through "thought suppression". We show a similar approach can be used to find a vector that suppresses the model's reasoning process, allowing us to remove censorship by applying the negative multiples of this vector

Despite investments in improving model safety, studies show that misaligned capabilities remain latent in safety-tuned models. In this work, we shed light on the mechanics of this phenomenon. First, we show that even when model generations are safe, harmful content can persist in hidden representations and can be extracted by decoding from earlier layers. Then, we show that whether the model divulges such content depends significantly on its perception of who it is talking to, which we refer to as user persona. In fact, we find manipulating user persona to be even more effective for eliciting harmful content than direct attempts to control model refusal. We study both natural language prompting and activation steering as control methods and show that activation steering is significantly more effective at bypassing safety filters. We investigate why certain personas break model safeguards and find that they enable the model to form more charitable interpretations of otherwise dangerous queries. Finally, we show we can predict a persona's effect on refusal given only the geometry of its steering vector.

This report introduces xGen-MM (also known as BLIP-3), a framework for developing Large Multimodal Models (LMMs). The framework comprises meticulously curated datasets, a training recipe, model architectures, and a resulting suite of LMMs. xGen-MM, short for xGen-MultiModal, expands the Salesforce xGen initiative on foundation AI models. Our models undergo rigorous evaluation across a range of tasks, including both single and multi-image benchmarks. Our pre-trained base model exhibits strong in-context learning capabilities and the instruction-tuned model demonstrates competitive performance among open-source LMMs with similar model sizes. In addition, we introduce a safety-tuned model with DPO, aiming to mitigate harmful behaviors such as hallucinations and improve safety. We open-source our models, curated large-scale datasets, and our fine-tuning codebase to facilitate further advancements in LMM research. Associated resources will be available on our project page above.

Fine-tuning large language models (LLMs) for downstream tasks is a widely adopted approach, but it often leads to safety degradation in safety-aligned LLMs. Currently, many solutions address this issue by incorporating additional safety data, which can be impractical in many cases. In this paper, we address the question: How can we improve downstream task performance while preserving safety in LLMs without relying on additional safety data? We propose a simple and effective method that maintains the inherent safety of LLMs while enhancing their downstream task performance: merging the weights of pre- and post-fine-tuned safety-aligned models. Experimental results across various downstream tasks, models, and merging methods demonstrate that this approach effectively mitigates safety degradation while improving downstream task performance, offering a practical solution for adapting safety-aligned LLMs.

The integration of large language models (LLMs) into cyber security applications presents significant opportunities, such as enhancing threat analysis and malware detection, but can also introduce critical risks and safety concerns, including personal data leakage and automated generation of new malware. To address these challenges, we developed CyberLLMInstruct, a dataset of 54,928 instruction-response pairs spanning cyber security tasks such as malware analysis, phishing simulations, and zero-day vulnerabilities. The dataset was constructed through a multi-stage process. This involved sourcing data from multiple resources, filtering and structuring it into instruction-response pairs, and aligning it with real-world scenarios to enhance its applicability. Seven open-source LLMs were chosen to test the usefulness of CyberLLMInstruct: Phi 3 Mini 3.8B, Mistral 7B, Qwen 2.5 7B, Llama 3 8B, Llama 3.1 8B, Gemma 2 9B, and Llama 2 70B. In our primary example, we rigorously assess the safety of fine-tuned models using the OWASP top 10 framework, finding that fine-tuning reduces safety resilience across all tested LLMs and every adversarial attack (e.g., the security score of Llama 3.1 8B against prompt injection drops from 0.95 to 0.15). In our second example, we show that these same fine-tuned models can also achieve up to 92.50 percent accuracy on the CyberMetric benchmark. These findings highlight a trade-off between performance and safety, showing the importance of adversarial testing and further research into fine-tuning methodologies that can mitigate safety risks while still improving performance across diverse datasets and domains. The dataset creation pipeline, along with comprehensive documentation, examples, and resources for reproducing our results, is publicly available at https://github.com/Adelsamir01/CyberLLMInstruct.

Although large language models (LLMs) achieve effective safety alignment at the time of release, they still face various safety challenges. A key issue is that fine-tuning often compromises the safety alignment of LLMs. To address this issue, we propose a method named IRR (Identify, Remove, and Recalibrate for Safety Realignment) that performs safety realignment for LLMs. The core of IRR is to identify and remove unsafe delta parameters from the fine-tuned models, while recalibrating the retained ones. We evaluate the effectiveness of IRR across various datasets, including both full fine-tuning and LoRA methods. Our results demonstrate that IRR significantly enhances the safety performance of fine-tuned models on safety benchmarks, such as harmful queries and jailbreak attacks, while maintaining their performance on downstream tasks. The source code is available at: https://anonymous.4open.science/r/IRR-BD4F.

Aligned language models face a significant limitation as their fine-tuning often results in compromised safety. To tackle this, we propose a simple method RESTA that performs LLM safety realignment. RESTA stands for REstoring Safety through Task Arithmetic. At its core, it involves a simple arithmetic addition of a safety vector to the weights of the compromised model. We demonstrate the effectiveness of RESTA in both parameter-efficient and full fine-tuning, covering a wide range of downstream tasks, including instruction following in Chinese, English, and Hindi, as well as problem-solving capabilities in Code and Math. We also showcase the generalizability of RESTA on three existing safety evaluation benchmarks and a multilingual benchmark dataset proposed as a part of this work, consisting of 550 harmful questions covering 11 categories, each with 5 sub-categories of harm. Overall, RESTA decreases the harmfulness of the compromised model from 18.6% to 5.1% and from 9.2% to 1.5% in parameter-efficient and full fine-tuning, respectively, while maintaining most of the model's performance on the task. We release the source codes at: https://github.com/declare-lab/resta.

LLMs(Large Language Models) nowadays have widespread adoption as a tool for solving issues across various domain/tasks. These models since are susceptible to produce harmful or toxic results, inference-time adversarial attacks, therefore they do undergo safety alignment training and Red teaming for putting in safety guardrails. For using these models, usually fine-tuning is done for model alignment on the desired tasks, which can make model more aligned but also make it more susceptible to produce unsafe responses, if fine-tuned with harmful data.In this paper, we study how much of impact introduction of harmful data in fine-tuning can make, and if it can override the safety protection of those models. Conversely,it was also explored that if model is fine-tuned on safety data can make the model produce more safer responses. Further we explore if fine-tuning the model on harmful data makes it less helpful or less trustworthy because of increase in model uncertainty leading to knowledge drift. Our extensive experimental results shown that Safety protection in an open-source can be overridden, when fine-tuned with harmful data as observed by ASR increasing by 35% when compared to basemodel's ASR. Also, as observed, fine-tuning a model with harmful data made the harmful fine-tuned model highly uncertain with huge knowledge drift and less truthfulness in its responses. Furthermore, for the safe fine-tuned model, ASR decreases by 51.68% as compared to the basemodel, and Safe model also shown in minor drop in uncertainty and truthfulness as compared to basemodel. This paper's code is available at: https://github.com/techsachinkr/Overriding_Model_Safety_Protections

The current safeguard mechanisms for large language models (LLMs) are indeed susceptible to jailbreak attacks, making them inherently fragile. Even the process of fine-tuning on apparently benign data for downstream tasks can jeopardize safety. One potential solution is to conduct safety fine-tuning subsequent to downstream fine-tuning. However, there's a risk of catastrophic forgetting during safety fine-tuning, where LLMs may regain safety measures but lose the task-specific knowledge acquired during downstream fine-tuning. In this paper, we introduce a safety realignment framework through subspace-oriented model fusion (SOMF), aiming to combine the safeguard capabilities of initially aligned model and the current fine-tuned model into a realigned model. Our approach begins by disentangling all task vectors from the weights of each fine-tuned model. We then identify safety-related regions within these vectors by subspace masking techniques. Finally, we explore the fusion of the initial safely aligned LLM with all task vectors based on the identified safety subspace. We validate that our safety realignment framework satisfies the safety requirements of a single fine-tuned model as well as multiple models during their fusion. Our findings confirm that SOMF preserves safety without notably compromising performance on downstream tasks, including instruction following in Chinese, English, and Hindi, as well as problem-solving capabilities in Code and Math.

Red-teaming is a common practice for mitigating unsafe behaviors in Large Language Models (LLMs), which involves thoroughly assessing LLMs to identify potential flaws and addressing them with responsible and accurate responses. While effective, manual red-teaming is costly, and existing automatic red-teaming typically discovers safety risks without addressing them. In this paper, we propose a Multi-round Automatic Red-Teaming (MART) method, which incorporates both automatic adversarial prompt writing and safe response generation, significantly increasing red-teaming scalability and the safety of the target LLM. Specifically, an adversarial LLM and a target LLM interplay with each other in an iterative manner, where the adversarial LLM aims to generate challenging prompts that elicit unsafe responses from the target LLM, while the target LLM is fine-tuned with safety aligned data on these adversarial prompts. In each round, the adversarial LLM crafts better attacks on the updated target LLM, while the target LLM also improves itself through safety fine-tuning. On adversarial prompt benchmarks, the violation rate of an LLM with limited safety alignment reduces up to 84.7% after 4 rounds of MART, achieving comparable performance to LLMs with extensive adversarial prompt writing. Notably, model helpfulness on non-adversarial prompts remains stable throughout iterations, indicating the target LLM maintains strong performance on instruction following.

Fine-tuning text-to-image diffusion models is widely used for personalization and adaptation for new domains. In this paper, we identify a critical vulnerability of fine-tuning: safety alignment methods designed to filter harmful content (e.g., nudity) can break down during fine-tuning, allowing previously suppressed content to resurface, even when using benign datasets. While this "fine-tuning jailbreaking" issue is known in large language models, it remains largely unexplored in text-to-image diffusion models. Our investigation reveals that standard fine-tuning can inadvertently undo safety measures, causing models to relearn harmful concepts that were previously removed and even exacerbate harmful behaviors. To address this issue, we present a novel but immediate solution called Modular LoRA, which involves training Safety Low-Rank Adaptation (LoRA) modules separately from Fine-Tuning LoRA components and merging them during inference. This method effectively prevents the re-learning of harmful content without compromising the model's performance on new tasks. Our experiments demonstrate that Modular LoRA outperforms traditional fine-tuning methods in maintaining safety alignment, offering a practical approach for enhancing the security of text-to-image diffusion models against potential attacks.

Safety backdoor attacks in large language models (LLMs) enable the stealthy triggering of unsafe behaviors while evading detection during normal interactions. The high dimensionality of potential triggers in the token space and the diverse range of malicious behaviors make this a critical challenge. We present BEEAR, a mitigation approach leveraging the insight that backdoor triggers induce relatively uniform drifts in the model's embedding space. Our bi-level optimization method identifies universal embedding perturbations that elicit unwanted behaviors and adjusts the model parameters to reinforce safe behaviors against these perturbations. Experiments show BEEAR reduces the success rate of RLHF time backdoor attacks from >95% to <1% and from 47% to 0% for instruction-tuning time backdoors targeting malicious code generation, without compromising model utility. Requiring only defender-defined safe and unwanted behaviors, BEEAR represents a step towards practical defenses against safety backdoors in LLMs, providing a foundation for further advancements in AI safety and security.

Fine-tuning large language models (LLMs) for telecom tasks and datasets is a common practice to adapt general-purpose models to the telecom domain. However, little attention has been paid to how this process may compromise model safety. Recent research has shown that even benign fine-tuning can degrade the safety alignment of LLMs, causing them to respond to harmful or unethical user queries. In this paper, we investigate this issue for telecom-tuned LLMs using three representative datasets featured by the GenAINet initiative. We show that safety degradation persists even for structured and seemingly harmless datasets such as 3GPP standards and tabular records, indicating that telecom-specific data is not immune to safety erosion during fine-tuning. We further extend our analysis to publicly available Telecom LLMs trained via continual pre-training, revealing that safety alignment is often severely lacking, primarily due to the omission of safety-focused instruction tuning. To address these issues in both fine-tuned and pre-trained models, we conduct extensive experiments and evaluate three safety realignment defenses (SafeInstruct, SafeLoRA, and SafeMERGE) using established red-teaming benchmarks. The results show that, across all settings, the proposed defenses can effectively restore safety after harmful degradation without compromising downstream task performance, leading to Safe teleCOMMunication (SafeCOMM) models. In a nutshell, our work serves as a diagnostic study and practical guide for safety realignment in telecom-tuned LLMs, and emphasizes the importance of safety-aware instruction and fine-tuning for real-world deployments of Telecom LLMs.

Current vision large language models (VLLMs) exhibit remarkable capabilities yet are prone to generate harmful content and are vulnerable to even the simplest jailbreaking attacks. Our initial analysis finds that this is due to the presence of harmful data during vision-language instruction fine-tuning, and that VLLM fine-tuning can cause forgetting of safety alignment previously learned by the underpinning LLM. To address this issue, we first curate a vision-language safe instruction-following dataset VLGuard covering various harmful categories. Our experiments demonstrate that integrating this dataset into standard vision-language fine-tuning or utilizing it for post-hoc fine-tuning effectively safety aligns VLLMs. This alignment is achieved with minimal impact on, or even enhancement of, the models' helpfulness. The versatility of our safety fine-tuning dataset makes it a valuable resource for safety-testing existing VLLMs, training new models or safeguarding pre-trained VLLMs. Empirical results demonstrate that fine-tuned VLLMs effectively reject unsafe instructions and substantially reduce the success rates of several black-box adversarial attacks, which approach zero in many cases. The code and dataset are available at https://github.com/ys-zong/VLGuard.

Simulators offer the possibility of safe, low-cost development of self-driving systems. However, current driving simulators exhibit na\"ive behavior models for background traffic. Hand-tuned scenarios are typically added during simulation to induce safety-critical situations. An alternative approach is to adversarially perturb the background traffic trajectories. In this paper, we study this approach to safety-critical driving scenario generation using the CARLA simulator. We use a kinematic bicycle model as a proxy to the simulator's true dynamics and observe that gradients through this proxy model are sufficient for optimizing the background traffic trajectories. Based on this finding, we propose KING, which generates safety-critical driving scenarios with a 20% higher success rate than black-box optimization. By solving the scenarios generated by KING using a privileged rule-based expert algorithm, we obtain training data for an imitation learning policy. After fine-tuning on this new data, we show that the policy becomes better at avoiding collisions. Importantly, our generated data leads to reduced collisions on both held-out scenarios generated via KING as well as traditional hand-crafted scenarios, demonstrating improved robustness.

Llama 2-Chat is a collection of large language models that Meta developed and released to the public. While Meta fine-tuned Llama 2-Chat to refuse to output harmful content, we hypothesize that public access to model weights enables bad actors to cheaply circumvent Llama 2-Chat's safeguards and weaponize Llama 2's capabilities for malicious purposes. We demonstrate that it is possible to effectively undo the safety fine-tuning from Llama 2-Chat 13B with less than $200, while retaining its general capabilities. Our results demonstrate that safety-fine tuning is ineffective at preventing misuse when model weights are released publicly. Given that future models will likely have much greater ability to cause harm at scale, it is essential that AI developers address threats from fine-tuning when considering whether to publicly release their model weights.

This study presents a computer vision approach aimed at detecting snow on sidewalks and pavements to reduce winter-related fall injuries, especially among elderly and visually impaired individuals. Leveraging fine-tuned VGG-19 and ResNet50 convolutional neural networks (CNNs), the research focuses on identifying snow presence in pavement images. The dataset comprises 98 images evenly split between snowy and snow-free conditions, evaluated with a separate test set using the F1 score and accuracy metrics. This work builds upon existing research by employing fine-tuned CNN architectures to accurately detect snow on pavements from smartphone-captured images. The methodology incorporates transfer learning and model ensembling techniques to integrate the best predictions from both the VGG19 and ResNet50 architectures. The study yields accuracy and F1 scores of 81.8% and 81.7%, respectively, showcasing the potential of computer vision in addressing winter-related hazards for vulnerable populations.

As large language models (LLMs) continue to advance in capabilities, ensuring their safety against jailbreak attacks remains a critical challenge. In this paper, we introduce a novel safety alignment approach called Answer-Then-Check, which enhances LLM robustness against malicious prompts by applying thinking ability to mitigate jailbreaking problems before producing a final answer to the user. Our method enables models to directly answer the question in their thought and then critically evaluate its safety before deciding whether to provide it. To implement this approach, we construct the Reasoned Safety Alignment (ReSA) dataset, comprising 80K examples that teach models to reason through direct responses and then analyze their safety. Experimental results demonstrate that our approach achieves the Pareto frontier with superior safety capability while decreasing over-refusal rates on over-refusal benchmarks. Notably, the model fine-tuned with ReSA maintains general reasoning capabilities on benchmarks like MMLU, MATH500, and HumanEval. Besides, our method equips models with the ability to perform safe completion. Unlike post-hoc methods that can only reject harmful queries, our model can provide helpful and safe alternative responses for sensitive topics (e.g., self-harm). Furthermore, we discover that training on a small subset of just 500 examples can achieve comparable performance to using the full dataset, suggesting that safety alignment may require less data than previously assumed.

The open-ended nature of language generation makes the evaluation of autoregressive large language models (LLMs) challenging. One common evaluation approach uses multiple-choice questions (MCQ) to limit the response space. The model is then evaluated by ranking the candidate answers by the log probability of the first token prediction. However, first-tokens may not consistently reflect the final response output, due to model's diverse response styles such as starting with "Sure" or refusing to answer. Consequently, MCQ evaluation is not indicative of model behaviour when interacting with users. But by how much? We evaluate how aligned first-token evaluation is with the text output along several dimensions, namely final option choice, refusal rate, choice distribution and robustness under prompt perturbation. Our results show that the two approaches are severely misaligned on all dimensions, reaching mismatch rates over 60%. Models heavily fine-tuned on conversational or safety data are especially impacted. Crucially, models remain misaligned even when we increasingly constrain prompts, i.e., force them to start with an option letter or example template. Our findings i) underscore the importance of inspecting the text output as well and ii) caution against relying solely on first-token evaluation.

To ensure AI safety, instruction-tuned Large Language Models (LLMs) are specifically trained to ensure alignment, which refers to making models behave in accordance with human intentions. While these models have demonstrated commendable results on various safety benchmarks, the vulnerability of their safety alignment has not been extensively studied. This is particularly troubling given the potential harm that LLMs can inflict. Existing attack methods on LLMs often rely on poisoned training data or the injection of malicious prompts. These approaches compromise the stealthiness and generalizability of the attacks, making them susceptible to detection. Additionally, these models often demand substantial computational resources for implementation, making them less practical for real-world applications. Inspired by recent success in modifying model behavior through steering vectors without the need for optimization, and drawing on its effectiveness in red-teaming LLMs, we conducted experiments employing activation steering to target four key aspects of LLMs: truthfulness, toxicity, bias, and harmfulness - across a varied set of attack settings. To establish a universal attack strategy applicable to diverse target alignments without depending on manual analysis, we automatically select the intervention layer based on contrastive layer search. Our experiment results show that activation attacks are highly effective and add little or no overhead to attack efficiency. Additionally, we discuss potential countermeasures against such activation attacks. Our code and data are available at https://github.com/wang2226/Backdoor-Activation-Attack Warning: this paper contains content that can be offensive or upsetting.

Large language models (LLMs) show remarkable human-like capability in various domains and languages. However, a notable quality gap arises in low-resource languages, e.g., Indonesian indigenous languages, rendering them ineffective and inefficient in such linguistic contexts. To bridge this quality gap, we introduce Cendol, a collection of Indonesian LLMs encompassing both decoder-only and encoder-decoder architectures across a range of model sizes. We highlight Cendol's effectiveness across a diverse array of tasks, attaining 20% improvement, and demonstrate its capability to generalize to unseen tasks and indigenous languages of Indonesia. Furthermore, Cendol models showcase improved human favorability despite their limitations in capturing indigenous knowledge and cultural values in Indonesia. In addition, we discuss the shortcomings of parameter-efficient tunings, such as LoRA, for language adaptation. Alternatively, we propose the usage of vocabulary adaptation to enhance efficiency. Lastly, we evaluate the safety of Cendol and showcase that safety in pre-training in one language such as English is transferable to low-resource languages, such as Indonesian, even without RLHF and safety fine-tuning.

Ensuring the safe alignment of large language models (LLMs) with human values is critical as they become integral to applications like translation and question answering. Current alignment methods struggle with dynamic user intentions and complex objectives, making models vulnerable to generating harmful content. We propose Safety Arithmetic, a training-free framework enhancing LLM safety across different scenarios: Base models, Supervised fine-tuned models (SFT), and Edited models. Safety Arithmetic involves Harm Direction Removal to avoid harmful content and Safety Alignment to promote safe responses. Additionally, we present NoIntentEdit, a dataset highlighting edit instances that could compromise model safety if used unintentionally. Our experiments show that Safety Arithmetic significantly improves safety measures, reduces over-safety, and maintains model utility, outperforming existing methods in ensuring safe content generation.

Safety alignment approaches in large language models (LLMs) often lead to the over-refusal of benign queries, significantly diminishing their utility in sensitive scenarios. To address this challenge, we introduce FalseReject, a comprehensive resource containing 16k seemingly toxic queries accompanied by structured responses across 44 safety-related categories. We propose a graph-informed adversarial multi-agent interaction framework to generate diverse and complex prompts, while structuring responses with explicit reasoning to aid models in accurately distinguishing safe from unsafe contexts. FalseReject includes training datasets tailored for both standard instruction-tuned models and reasoning-oriented models, as well as a human-annotated benchmark test set. Our extensive benchmarking on 29 state-of-the-art (SOTA) LLMs reveals persistent over-refusal challenges. Empirical results demonstrate that supervised finetuning with FalseReject substantially reduces unnecessary refusals without compromising overall safety or general language capabilities.

Fine-tuning Large Language Models (LLMs) with Low-Rank Adaptation (LoRA) enhances adaptability while reducing computational costs. However, fine-tuning can compromise safety alignment, even with benign data, increasing susceptibility to harmful outputs. Existing safety alignment methods struggle to capture complex parameter shifts, leading to suboptimal safety-utility trade-offs. To address this issue, we propose Safe Pruning LoRA (SPLoRA), a novel pruning-based approach that selectively removes LoRA layers that weaken safety alignment, improving safety while preserving performance. At its core, we introduce Empirical-DIEM (E-DIEM), a dimension-insensitive similarity metric that effectively detects safety misalignment in LoRA-adapted models. We conduct extensive experiments on LLMs fine-tuned with mixed of benign and malicious data, and purely benign datasets, evaluating SPLoRA across utility, safety, and reliability metrics. Results demonstrate that SPLoRA outperforms state-of-the-art safety alignment techniques, significantly reducing safety risks while maintaining or improving model performance and reliability. Additionally, SPLoRA reduces inference overhead, making it a scalable and efficient solution for deploying safer and more reliable LLMs. The code is available at https://github.com/AoShuang92/SPLoRA.

In this work, we develop and release Llama 2, a collection of pretrained and fine-tuned large language models (LLMs) ranging in scale from 7 billion to 70 billion parameters. Our fine-tuned LLMs, called Llama 2-Chat, are optimized for dialogue use cases. Our models outperform open-source chat models on most benchmarks we tested, and based on our human evaluations for helpfulness and safety, may be a suitable substitute for closed-source models. We provide a detailed description of our approach to fine-tuning and safety improvements of Llama 2-Chat in order to enable the community to build on our work and contribute to the responsible development of LLMs.

This paper introduces v0.5 of the AI Safety Benchmark, which has been created by the MLCommons AI Safety Working Group. The AI Safety Benchmark has been designed to assess the safety risks of AI systems that use chat-tuned language models. We introduce a principled approach to specifying and constructing the benchmark, which for v0.5 covers only a single use case (an adult chatting to a general-purpose assistant in English), and a limited set of personas (i.e., typical users, malicious users, and vulnerable users). We created a new taxonomy of 13 hazard categories, of which 7 have tests in the v0.5 benchmark. We plan to release version 1.0 of the AI Safety Benchmark by the end of 2024. The v1.0 benchmark will provide meaningful insights into the safety of AI systems. However, the v0.5 benchmark should not be used to assess the safety of AI systems. We have sought to fully document the limitations, flaws, and challenges of v0.5. This release of v0.5 of the AI Safety Benchmark includes (1) a principled approach to specifying and constructing the benchmark, which comprises use cases, types of systems under test (SUTs), language and context, personas, tests, and test items; (2) a taxonomy of 13 hazard categories with definitions and subcategories; (3) tests for seven of the hazard categories, each comprising a unique set of test items, i.e., prompts. There are 43,090 test items in total, which we created with templates; (4) a grading system for AI systems against the benchmark; (5) an openly available platform, and downloadable tool, called ModelBench that can be used to evaluate the safety of AI systems on the benchmark; (6) an example evaluation report which benchmarks the performance of over a dozen openly available chat-tuned language models; (7) a test specification for the benchmark.

Reasoning has become a central paradigm for large language models (LLMs), consistently boosting accuracy across diverse benchmarks. Yet its suitability for precision-sensitive tasks remains unclear. We present the first systematic study of reasoning for classification tasks under strict low false positive rate (FPR) regimes. Our analysis covers two tasks--safety detection and hallucination detection--evaluated in both fine-tuned and zero-shot settings, using standard LLMs and Large Reasoning Models (LRMs). Our results reveal a clear trade-off: Think On (reasoning-augmented) generation improves overall accuracy, but underperforms at the low-FPR thresholds essential for practical use. In contrast, Think Off (no reasoning during inference) dominates in these precision-sensitive regimes, with Think On surpassing only when higher FPRs are acceptable. In addition, we find token-based scoring substantially outperforms self-verbalized confidence for precision-sensitive deployments. Finally, a simple ensemble of the two modes recovers the strengths of each. Taken together, our findings position reasoning as a double-edged tool: beneficial for average accuracy, but often ill-suited for applications requiring strict precision.

Larger language models (LLMs) have taken the world by storm with their massive multi-tasking capabilities simply by optimizing over a next-word prediction objective. With the emergence of their properties and encoded knowledge, the risk of LLMs producing harmful outputs increases, making them unfit for scalable deployment for the public. In this work, we propose a new safety evaluation benchmark RED-EVAL that carries out red-teaming. We show that even widely deployed models are susceptible to the Chain of Utterances-based (CoU) prompting, jailbreaking closed source LLM-based systems such as GPT-4 and ChatGPT to unethically respond to more than 65% and 73% of harmful queries. We also demonstrate the consistency of the RED-EVAL across 8 open-source LLMs in generating harmful responses in more than 86% of the red-teaming attempts. Next, we propose RED-INSTRUCT--An approach for the safety alignment of LLMs. It constitutes two phases: 1) HARMFULQA data collection: Leveraging CoU prompting, we collect a dataset that consists of 1.9K harmful questions covering a wide range of topics, 9.5K safe and 7.3K harmful conversations from ChatGPT; 2) SAFE-ALIGN: We demonstrate how the conversational dataset can be used for the safety alignment of LLMs by minimizing the negative log-likelihood over helpful responses and penalizing over harmful responses by gradient accent over sample loss. Our model STARLING, a fine-tuned Vicuna-7B, is observed to be more safely aligned when evaluated on RED-EVAL and HHH benchmarks while preserving the utility of the baseline models (TruthfulQA, MMLU, and BBH).

With the widespread application of Large Language Models (LLMs), their associated security issues have become increasingly prominent, severely constraining their trustworthy deployment in critical domains. This paper proposes a novel safety response framework designed to systematically safeguard LLMs at both the input and output levels. At the input level, the framework employs a supervised fine-tuning-based safety classification model. Through a fine-grained four-tier taxonomy (Safe, Unsafe, Conditionally Safe, Focused Attention), it performs precise risk identification and differentiated handling of user queries, significantly enhancing risk coverage and business scenario adaptability, and achieving a risk recall rate of 99.3%. At the output level, the framework integrates Retrieval-Augmented Generation (RAG) with a specifically fine-tuned interpretation model, ensuring all responses are grounded in a real-time, trustworthy knowledge base. This approach eliminates information fabrication and enables result traceability. Experimental results demonstrate that our proposed safety control model achieves a significantly higher safety score on public safety evaluation benchmarks compared to the baseline model, TinyR1-Safety-8B. Furthermore, on our proprietary high-risk test set, the framework's components attained a perfect 100% safety score, validating their exceptional protective capabilities in complex risk scenarios. This research provides an effective engineering pathway for building high-security, high-trust LLM applications.

Diffusion-based large language models (dLLMs) have recently emerged as a powerful alternative to autoregressive LLMs, offering faster inference and greater interactivity via parallel decoding and bidirectional modeling. However, despite strong performance in code generation and text infilling, we identify a fundamental safety concern: existing alignment mechanisms fail to safeguard dLLMs against context-aware, masked-input adversarial prompts, exposing novel vulnerabilities. To this end, we present DIJA, the first systematic study and jailbreak attack framework that exploits unique safety weaknesses of dLLMs. Specifically, our proposed DIJA constructs adversarial interleaved mask-text prompts that exploit the text generation mechanisms of dLLMs, i.e., bidirectional modeling and parallel decoding. Bidirectional modeling drives the model to produce contextually consistent outputs for masked spans, even when harmful, while parallel decoding limits model dynamic filtering and rejection sampling of unsafe content. This causes standard alignment mechanisms to fail, enabling harmful completions in alignment-tuned dLLMs, even when harmful behaviors or unsafe instructions are directly exposed in the prompt. Through comprehensive experiments, we demonstrate that DIJA significantly outperforms existing jailbreak methods, exposing a previously overlooked threat surface in dLLM architectures. Notably, our method achieves up to 100% keyword-based ASR on Dream-Instruct, surpassing the strongest prior baseline, ReNeLLM, by up to 78.5% in evaluator-based ASR on JailbreakBench and by 37.7 points in StrongREJECT score, while requiring no rewriting or hiding of harmful content in the jailbreak prompt. Our findings underscore the urgent need for rethinking safety alignment in this emerging class of language models. Code is available at https://github.com/ZichenWen1/DIJA.

AI developers often apply safety alignment procedures to prevent the misuse of their AI systems. For example, before Meta released Llama 2-Chat, a collection of instruction fine-tuned large language models, they invested heavily in safety training, incorporating extensive red-teaming and reinforcement learning from human feedback. However, it remains unclear how well safety training guards against model misuse when attackers have access to model weights. We explore the robustness of safety training in language models by subversively fine-tuning the public weights of Llama 2-Chat. We employ low-rank adaptation (LoRA) as an efficient fine-tuning method. With a budget of less than $200 per model and using only one GPU, we successfully undo the safety training of Llama 2-Chat models of sizes 7B, 13B, and 70B. Specifically, our fine-tuning technique significantly reduces the rate at which the model refuses to follow harmful instructions. We achieve a refusal rate below 1% for our 70B Llama 2-Chat model on two refusal benchmarks. Our fine-tuning method retains general performance, which we validate by comparing our fine-tuned models against Llama 2-Chat across two benchmarks. Additionally, we present a selection of harmful outputs produced by our models. While there is considerable uncertainty about the scope of risks from current models, it is likely that future models will have significantly more dangerous capabilities, including the ability to hack into critical infrastructure, create dangerous bio-weapons, or autonomously replicate and adapt to new environments. We show that subversive fine-tuning is practical and effective, and hence argue that evaluating risks from fine-tuning should be a core part of risk assessments for releasing model weights.

Vision Large Language Models (VLLMs) have demonstrated impressive capabilities in general visual tasks such as image captioning and visual question answering. However, their effectiveness in specialized, safety-critical domains like autonomous driving remains largely unexplored. Autonomous driving systems require sophisticated scene understanding in complex environments, yet existing multimodal benchmarks primarily focus on normal driving conditions, failing to adequately assess VLLMs' performance in safety-critical scenarios. To address this, we introduce DVBench, a pioneering benchmark designed to evaluate the performance of VLLMs in understanding safety-critical driving videos. Built around a hierarchical ability taxonomy that aligns with widely adopted frameworks for describing driving scenarios used in assessing highly automated driving systems, DVBench features 10,000 multiple-choice questions with human-annotated ground-truth answers, enabling a comprehensive evaluation of VLLMs' capabilities in perception and reasoning. Experiments on 14 SOTA VLLMs, ranging from 0.5B to 72B parameters, reveal significant performance gaps, with no model achieving over 40% accuracy, highlighting critical limitations in understanding complex driving scenarios. To probe adaptability, we fine-tuned selected models using domain-specific data from DVBench, achieving accuracy gains ranging from 5.24 to 10.94 percentage points, with relative improvements of up to 43.59%. This improvement underscores the necessity of targeted adaptation to bridge the gap between general-purpose VLLMs and mission-critical driving applications. DVBench establishes an essential evaluation framework and research roadmap for developing VLLMs that meet the safety and robustness requirements for real-world autonomous systems. We released the benchmark toolbox and the fine-tuned model at: https://github.com/tong-zeng/DVBench.git.

The early detection of mental health disorders from social media text is critical for enabling timely support, risk assessment, and referral to appropriate resources. This work introduces multiMentalRoBERTa, a fine-tuned RoBERTa model designed for multiclass classification of common mental health conditions, including stress, anxiety, depression, post-traumatic stress disorder (PTSD), suicidal ideation, and neutral discourse. Drawing on multiple curated datasets, data exploration is conducted to analyze class overlaps, revealing strong correlations between depression and suicidal ideation as well as anxiety and PTSD, while stress emerges as a broad, overlapping category. Comparative experiments with traditional machine learning methods, domain-specific transformers, and prompting-based large language models demonstrate that multiMentalRoBERTa achieves superior performance, with macro F1-scores of 0.839 in the six-class setup and 0.870 in the five-class setup (excluding stress), outperforming both fine-tuned MentalBERT and baseline classifiers. Beyond predictive accuracy, explainability methods, including Layer Integrated Gradients and KeyBERT, are applied to identify lexical cues that drive classification, with a particular focus on distinguishing depression from suicidal ideation. The findings emphasize the effectiveness of fine-tuned transformers for reliable and interpretable detection in sensitive contexts, while also underscoring the importance of fairness, bias mitigation, and human-in-the-loop safety protocols. Overall, multiMentalRoBERTa is presented as a lightweight, robust, and deployable solution for enhancing support in mental health platforms.

Tram-human interaction safety is an important challenge, given that trams frequently operate in densely populated areas, where collisions can range from minor injuries to fatal outcomes. This paper addresses the issue from the perspective of designing a solution leveraging digital image processing, deep learning, and artificial intelligence to improve the safety of pedestrians, drivers, cyclists, pets, and tram passengers. We present RailSafeNet, a real-time framework that fuses semantic segmentation, object detection and a rule-based Distance Assessor to highlight track intrusions. Using only monocular video, the system identifies rails, localises nearby objects and classifies their risk by comparing projected distances with the standard 1435mm rail gauge. Experiments on the diverse RailSem19 dataset show that a class-filtered SegFormer B3 model achieves 65% intersection-over-union (IoU), while a fine-tuned YOLOv8 attains 75.6% mean average precision (mAP) calculated at an intersection over union (IoU) threshold of 0.50. RailSafeNet therefore delivers accurate, annotation-light scene understanding that can warn drivers before dangerous situations escalate. Code available at https://github.com/oValach/RailSafeNet.

As Large Language Models (LLMs) and generative AI become increasingly widespread, concerns about content safety have grown in parallel. Currently, there is a clear lack of high-quality, human-annotated datasets that address the full spectrum of LLM-related safety risks and are usable for commercial applications. To bridge this gap, we propose a comprehensive and adaptable taxonomy for categorizing safety risks, structured into 12 top-level hazard categories with an extension to 9 fine-grained subcategories. This taxonomy is designed to meet the diverse requirements of downstream users, offering more granular and flexible tools for managing various risk types. Using a hybrid data generation pipeline that combines human annotations with a multi-LLM "jury" system to assess the safety of responses, we obtain Aegis 2.0, a carefully curated collection of 34,248 samples of human-LLM interactions, annotated according to our proposed taxonomy. To validate its effectiveness, we demonstrate that several lightweight models, trained using parameter-efficient techniques on Aegis 2.0, achieve performance competitive with leading safety models fully fine-tuned on much larger, non-commercial datasets. In addition, we introduce a novel training blend that combines safety with topic following data.This approach enhances the adaptability of guard models, enabling them to generalize to new risk categories defined during inference. We plan to open-source Aegis 2.0 data and models to the research community to aid in the safety guardrailing of LLMs.

Evaluating aligned large language models' (LLMs) ability to recognize and reject unsafe user requests is crucial for safe, policy-compliant deployments. Existing evaluation efforts, however, face three limitations that we address with SORRY-Bench, our proposed benchmark. First, existing methods often use coarse-grained taxonomies of unsafe topics, and are over-representing some fine-grained topics. For example, among the ten existing datasets that we evaluated, tests for refusals of self-harm instructions are over 3x less represented than tests for fraudulent activities. SORRY-Bench improves on this by using a fine-grained taxonomy of 45 potentially unsafe topics, and 450 class-balanced unsafe instructions, compiled through human-in-the-loop methods. Second, linguistic characteristics and formatting of prompts are often overlooked, like different languages, dialects, and more -- which are only implicitly considered in many evaluations. We supplement SORRY-Bench with 20 diverse linguistic augmentations to systematically examine these effects. Third, existing evaluations rely on large LLMs (e.g., GPT-4) for evaluation, which can be computationally expensive. We investigate design choices for creating a fast, accurate automated safety evaluator. By collecting 7K+ human annotations and conducting a meta-evaluation of diverse LLM-as-a-judge designs, we show that fine-tuned 7B LLMs can achieve accuracy comparable to GPT-4 scale LLMs, with lower computational cost. Putting these together, we evaluate over 40 proprietary and open-source LLMs on SORRY-Bench, analyzing their distinctive refusal behaviors. We hope our effort provides a building block for systematic evaluations of LLMs' safety refusal capabilities, in a balanced, granular, and efficient manner.

Large Language Models (LLMs) excel in various natural language processing tasks but remain vulnerable to generating harmful content or being exploited for malicious purposes. Although safety alignment datasets have been introduced to mitigate such risks through supervised fine-tuning (SFT), these datasets often lack comprehensive risk coverage. Most existing datasets focus primarily on lexical diversity while neglecting other critical dimensions. To address this limitation, we propose a novel analysis framework to systematically measure the risk coverage of alignment datasets across three essential dimensions: Lexical Diversity, Malicious Intent, and Jailbreak Tactics. We further introduce TRIDENT, an automated pipeline that leverages persona-based, zero-shot LLM generation to produce diverse and comprehensive instructions spanning these dimensions. Each harmful instruction is paired with an ethically aligned response, resulting in two datasets: TRIDENT-Core, comprising 26,311 examples, and TRIDENT-Edge, with 18,773 examples. Fine-tuning Llama 3.1-8B on TRIDENT-Edge demonstrates substantial improvements, achieving an average 14.29% reduction in Harm Score, and a 20% decrease in Attack Success Rate compared to the best-performing baseline model fine-tuned on the WildBreak dataset.

Fine-tuning open-source Vision-Language Models (VLMs) creates a critical yet underexplored attack surface: vulnerabilities in the base VLM could be retained in fine-tuned variants, rendering them susceptible to transferable jailbreak attacks. To demonstrate this risk, we introduce the Simulated Ensemble Attack (SEA), a novel grey-box jailbreak method in which the adversary has full access to the base VLM but no knowledge of the fine-tuned target's weights or training configuration. To improve jailbreak transferability across fine-tuned VLMs, SEA combines two key techniques: Fine-tuning Trajectory Simulation (FTS) and Targeted Prompt Guidance (TPG). FTS generates transferable adversarial images by simulating the vision encoder's parameter shifts, while TPG is a textual strategy that steers the language decoder toward adversarially optimized outputs. Experiments on the Qwen2-VL family (2B and 7B) demonstrate that SEA achieves high transfer attack success rates exceeding 86.5% and toxicity rates near 49.5% across diverse fine-tuned variants, even those specifically fine-tuned to improve safety behaviors. Notably, while direct PGD-based image jailbreaks rarely transfer across fine-tuned VLMs, SEA reliably exploits inherited vulnerabilities from the base model, significantly enhancing transferability. These findings highlight an urgent need to safeguard fine-tuned proprietary VLMs against transferable vulnerabilities inherited from open-source foundations, motivating the development of holistic defenses across the entire model lifecycle.

While large neural-based conversational models have become increasingly proficient dialogue agents, recent work has highlighted safety issues with these systems. For example, these systems can be goaded into generating toxic content, which often perpetuates social biases or stereotypes. We investigate a retrieval-based method for reducing bias and toxicity in responses from chatbots. It uses in-context learning to steer a model towards safer generations. Concretely, to generate a response to an unsafe dialogue context, we retrieve demonstrations of safe responses to similar dialogue contexts. We find our method performs competitively with strong baselines without requiring training. For instance, using automatic evaluation, we find our best fine-tuned baseline only generates safe responses to unsafe dialogue contexts from DiaSafety 4.04% more than our approach. Finally, we also propose a re-ranking procedure which can further improve response safeness.

Model merging has gained significant attention as a cost-effective approach to integrate multiple single-task fine-tuned models into a unified one that can perform well on multiple tasks. However, existing model merging techniques primarily focus on resolving conflicts between task-specific models, they often overlook potential security threats, particularly the risk of backdoor attacks in the open-source model ecosystem. In this paper, we first investigate the vulnerabilities of existing model merging methods to backdoor attacks, identifying two critical challenges: backdoor succession and backdoor transfer. To address these issues, we propose a novel Defense-Aware Merging (DAM) approach that simultaneously mitigates task interference and backdoor vulnerabilities. Specifically, DAM employs a meta-learning-based optimization method with dual masks to identify a shared and safety-aware subspace for model merging. These masks are alternately optimized: the Task-Shared mask identifies common beneficial parameters across tasks, aiming to preserve task-specific knowledge while reducing interference, while the Backdoor-Detection mask isolates potentially harmful parameters to neutralize security threats. This dual-mask design allows us to carefully balance the preservation of useful knowledge and the removal of potential vulnerabilities. Compared to existing merging methods, DAM achieves a more favorable balance between performance and security, reducing the attack success rate by 2-10 percentage points while sacrificing only about 1% in accuracy. Furthermore, DAM exhibits robust performance and broad applicability across various types of backdoor attacks and the number of compromised models involved in the merging process. We will release the codes and models soon.

We introduce Jais and Jais-chat, new state-of-the-art Arabic-centric foundation and instruction-tuned open generative large language models (LLMs). The models are based on the GPT-3 decoder-only architecture and are pretrained on a mixture of Arabic and English texts, including source code in various programming languages. With 13 billion parameters, they demonstrate better knowledge and reasoning capabilities in Arabic than any existing open Arabic and multilingual models by a sizable margin, based on extensive evaluation. Moreover, the models are competitive in English compared to English-centric open models of similar size, despite being trained on much less English data. We provide a detailed description of the training, the tuning, the safety alignment, and the evaluation of the models. We release two open versions of the model -- the foundation Jais model, and an instruction-tuned Jais-chat variant -- with the aim of promoting research on Arabic LLMs. Available at https://huggingface.co/inception-mbzuai/jais-13b-chat

Large language models (LLMs) achieve state-of-the-art performance on multiple language tasks, yet their safety guardrails can be circumvented, leading to harmful generations. In light of this, recent research on safety mechanisms has emerged, revealing that when safety representations or component are suppressed, the safety capability of LLMs are compromised. However, existing research tends to overlook the safety impact of multi-head attention mechanisms, despite their crucial role in various model functionalities. Hence, in this paper, we aim to explore the connection between standard attention mechanisms and safety capability to fill this gap in the safety-related mechanistic interpretability. We propose a novel metric which tailored for multi-head attention, the Safety Head ImPortant Score (Ships), to assess the individual heads' contributions to model safety. Based on this, we generalize Ships to the dataset level and further introduce the Safety Attention Head AttRibution Algorithm (Sahara) to attribute the critical safety attention heads inside the model. Our findings show that the special attention head has a significant impact on safety. Ablating a single safety head allows aligned model (e.g., Llama-2-7b-chat) to respond to 16 times more harmful queries, while only modifying 0.006% of the parameters, in contrast to the ~ 5% modification required in previous studies. More importantly, we demonstrate that attention heads primarily function as feature extractors for safety and models fine-tuned from the same base model exhibit overlapping safety heads through comprehensive experiments. Together, our attribution approach and findings provide a novel perspective for unpacking the black box of safety mechanisms within large models.

The ideal LLM content moderation system would be both structurally interpretable (so its decisions can be explained to users) and steerable (to reflect a community's values or align to safety standards). However, current systems fall short on both of these dimensions. To address this gap, we present SafetyAnalyst, a novel LLM safety moderation framework. Given a prompt, SafetyAnalyst creates a structured "harm-benefit tree," which identifies 1) the actions that could be taken if a compliant response were provided, 2) the harmful and beneficial effects of those actions (along with their likelihood, severity, and immediacy), and 3) the stakeholders that would be impacted by those effects. It then aggregates this structured representation into a harmfulness score based on a parameterized set of safety preferences, which can be transparently aligned to particular values. Using extensive harm-benefit features generated by SOTA LLMs on 19k prompts, we fine-tuned an open-weight LM to specialize in generating harm-benefit trees through symbolic knowledge distillation. On a comprehensive set of prompt safety benchmarks, we show that our system (average F1=0.75) outperforms existing LLM safety moderation systems (average F1<0.72) on prompt harmfulness classification, while offering the additional advantages of interpretability and steerability.

Safety alignment is critical for the ethical deployment of large language models (LLMs), guiding them to avoid generating harmful or unethical content. Current alignment techniques, such as supervised fine-tuning and reinforcement learning from human feedback, remain fragile and can be bypassed by carefully crafted adversarial prompts. Unfortunately, such attacks rely on trial and error, lack generalizability across models, and are constrained by scalability and reliability. This paper presents NeuroStrike, a novel and generalizable attack framework that exploits a fundamental vulnerability introduced by alignment techniques: the reliance on sparse, specialized safety neurons responsible for detecting and suppressing harmful inputs. We apply NeuroStrike to both white-box and black-box settings: In the white-box setting, NeuroStrike identifies safety neurons through feedforward activation analysis and prunes them during inference to disable safety mechanisms. In the black-box setting, we propose the first LLM profiling attack, which leverages safety neuron transferability by training adversarial prompt generators on open-weight surrogate models and then deploying them against black-box and proprietary targets. We evaluate NeuroStrike on over 20 open-weight LLMs from major LLM developers. By removing less than 0.6% of neurons in targeted layers, NeuroStrike achieves an average attack success rate (ASR) of 76.9% using only vanilla malicious prompts. Moreover, Neurostrike generalizes to four multimodal LLMs with 100% ASR on unsafe image inputs. Safety neurons transfer effectively across architectures, raising ASR to 78.5% on 11 fine-tuned models and 77.7% on five distilled models. The black-box LLM profiling attack achieves an average ASR of 63.7% across five black-box models, including the Google Gemini family.

Scams exploiting real-time social engineering -- such as phishing, impersonation, and phone fraud -- remain a persistent and evolving threat across digital platforms. Existing defenses are largely reactive, offering limited protection during active interactions. We propose a privacy-preserving, AI-in-the-loop framework that proactively detects and disrupts scam conversations in real time. The system combines instruction-tuned artificial intelligence with a safety-aware utility function that balances engagement with harm minimization, and employs federated learning to enable continual model updates without raw data sharing. Experimental evaluations show that the system produces fluent and engaging responses (perplexity as low as 22.3, engagement approx0.80), while human studies confirm significant gains in realism, safety, and effectiveness over strong baselines. In federated settings, models trained with FedAvg sustain up to 30 rounds while preserving high engagement (approx0.80), strong relevance (approx0.74), and low PII leakage (leq0.0085). Even with differential privacy, novelty and safety remain stable, indicating that robust privacy can be achieved without sacrificing performance. The evaluation of guard models (LlamaGuard, LlamaGuard2/3, MD-Judge) shows a straightforward pattern: stricter moderation settings reduce the chance of exposing personal information, but they also limit how much the model engages in conversation. In contrast, more relaxed settings allow longer and richer interactions, which improve scam detection, but at the cost of higher privacy risk. To our knowledge, this is the first framework to unify real-time scam-baiting, federated privacy preservation, and calibrated safety moderation into a proactive defense paradigm.

In addressing the critical need for safety in Large Language Models (LLMs), it is crucial to ensure that the outputs are not only safe but also retain their contextual accuracy. Many existing LLMs are safe fine-tuned either with safety demonstrations, or rely only on adversarial testing. While able to get safe outputs, they often risk losing contextual meaning as they mitigate bias and toxicity. In response, we present MBIAS, a LLM framework instruction fine-tuned on a custom dataset specifically designed for safety interventions. MBIAS aims to address the significant issues of bias and toxicity in LLMs generations that typically manifest as underrepresentation or negative portrayals across various demographics, including inappropriate linguistic mentions and biased content in social media. We experiment on MBIAS for safety interventions using various configurations, and demonstrate more than a 30\% reduction in overall bias and toxicity while successfully retaining key information. Additionally, a demographic analysis on an out-of-distribution test set confirms the robustness of our approach, with reductions in bias and toxicity exceeding 90\% across various demographics. The dataset and instruction fine-tuned MBIAS are made available to the research community at https://huggingface.co/newsmediabias/MBIAS.

Public LLMs such as the Llama 2-Chat have driven huge activity in LLM research. These models underwent alignment training and were considered safe. Recently Qi et al. (2023) reported that even benign fine-tuning (e.g., on seemingly safe datasets) can give rise to unsafe behaviors in the models. The current paper is about methods and best practices to mitigate such loss of alignment. Through extensive experiments on several chat models (Meta's Llama 2-Chat, Mistral AI's Mistral 7B Instruct v0.2, and OpenAI's GPT-3.5 Turbo), this paper uncovers that the prompt templates used during fine-tuning and inference play a crucial role in preserving safety alignment, and proposes the "Pure Tuning, Safe Testing" (PTST) principle -- fine-tune models without a safety prompt, but include it at test time. Fine-tuning experiments on GSM8K, ChatDoctor, and OpenOrca show that PTST significantly reduces the rise of unsafe behaviors, and even almost eliminates them in some cases.

Conversational large language models are fine-tuned for both instruction-following and safety, resulting in models that obey benign requests but refuse harmful ones. While this refusal behavior is widespread across chat models, its underlying mechanisms remain poorly understood. In this work, we show that refusal is mediated by a one-dimensional subspace, across 13 popular open-source chat models up to 72B parameters in size. Specifically, for each model, we find a single direction such that erasing this direction from the model's residual stream activations prevents it from refusing harmful instructions, while adding this direction elicits refusal on even harmless instructions. Leveraging this insight, we propose a novel white-box jailbreak method that surgically disables refusal with minimal effect on other capabilities. Finally, we mechanistically analyze how adversarial suffixes suppress propagation of the refusal-mediating direction. Our findings underscore the brittleness of current safety fine-tuning methods. More broadly, our work showcases how an understanding of model internals can be leveraged to develop practical methods for controlling model behavior.

Large language models (LLMs) frequently produce false refusals, declining benign requests that contain terms resembling unsafe queries. We address this challenge by introducing two comprehensive benchmarks: the Exaggerated Safety Benchmark (XSB) for single-turn prompts, annotated with "Focus" keywords that identify refusal-inducing triggers, and the Multi-turn Scenario-based Exaggerated Safety Benchmark (MS-XSB), which systematically evaluates refusal calibration in realistic, context-rich dialog settings. Our benchmarks reveal that exaggerated refusals persist across diverse recent LLMs and are especially pronounced in complex, multi-turn scenarios. To mitigate these failures, we leverage post-hoc explanation methods to identify refusal triggers and deploy three lightweight, model-agnostic approaches, ignore-word instructions, prompt rephrasing, and attention steering, at inference time, all without retraining or parameter access. Experiments on four instruction-tuned Llama models demonstrate that these strategies substantially improve compliance on safe prompts while maintaining robust safety protections. Our findings establish a reproducible framework for diagnosing and mitigating exaggerated refusals, highlighting practical pathways to safer and more helpful LLM deployments.

Pretrained language models underpin several AI applications, but their high computational cost for training limits accessibility. Initiatives such as BLOOM and StarCoder aim to democratize access to pretrained models for collaborative community development. However, such existing models face challenges: limited multilingual capabilities, continual pretraining causing catastrophic forgetting, whereas pretraining from scratch is computationally expensive, and compliance with AI safety and development laws. This paper presents Aurora-M, a 15B parameter multilingual open-source model trained on English, Finnish, Hindi, Japanese, Vietnamese, and code. Continually pretrained from StarCoderPlus on 435 billion additional tokens, Aurora-M surpasses 2 trillion tokens in total training token count. It is the first open-source multilingual model fine-tuned on human-reviewed safety instructions, thus aligning its development not only with conventional red-teaming considerations, but also with the specific concerns articulated in the Biden-Harris Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. Aurora-M is rigorously evaluated across various tasks and languages, demonstrating robustness against catastrophic forgetting and outperforming alternatives in multilingual settings, particularly in safety evaluations. To promote responsible open-source LLM development, Aurora-M and its variants are released at https://huggingface.co/collections/aurora-m/aurora-m-models-65fdfdff62471e09812f5407 .

We present LaMDA: Language Models for Dialog Applications. LaMDA is a family of Transformer-based neural language models specialized for dialog, which have up to 137B parameters and are pre-trained on 1.56T words of public dialog data and web text. While model scaling alone can improve quality, it shows less improvements on safety and factual grounding. We demonstrate that fine-tuning with annotated data and enabling the model to consult external knowledge sources can lead to significant improvements towards the two key challenges of safety and factual grounding. The first challenge, safety, involves ensuring that the model's responses are consistent with a set of human values, such as preventing harmful suggestions and unfair bias. We quantify safety using a metric based on an illustrative set of human values, and we find that filtering candidate responses using a LaMDA classifier fine-tuned with a small amount of crowdworker-annotated data offers a promising approach to improving model safety. The second challenge, factual grounding, involves enabling the model to consult external knowledge sources, such as an information retrieval system, a language translator, and a calculator. We quantify factuality using a groundedness metric, and we find that our approach enables the model to generate responses grounded in known sources, rather than responses that merely sound plausible. Finally, we explore the use of LaMDA in the domains of education and content recommendations, and analyze their helpfulness and role consistency.

The increasing deployment of Large Language Models (LLMs) across enterprise and mission-critical domains has underscored the urgent need for robust guardrailing systems that ensure safety, reliability, and compliance. Existing solutions often struggle with real-time oversight, multi-modal data handling, and explainability -- limitations that hinder their adoption in regulated environments. Existing guardrails largely operate in isolation, focused on text alone making them inadequate for multi-modal, production-scale environments. We introduce Protect, natively multi-modal guardrailing model designed to operate seamlessly across text, image, and audio inputs, designed for enterprise-grade deployment. Protect integrates fine-tuned, category-specific adapters trained via Low-Rank Adaptation (LoRA) on an extensive, multi-modal dataset covering four safety dimensions: toxicity, sexism, data privacy, and prompt injection. Our teacher-assisted annotation pipeline leverages reasoning and explanation traces to generate high-fidelity, context-aware labels across modalities. Experimental results demonstrate state-of-the-art performance across all safety dimensions, surpassing existing open and proprietary models such as WildGuard, LlamaGuard-4, and GPT-4.1. Protect establishes a strong foundation for trustworthy, auditable, and production-ready safety systems capable of operating across text, image, and audio modalities.

Incorporating external context can significantly enhance the response quality of Large Language Models (LLMs). However, real-world contexts often mix relevant information with disproportionate inappropriate content, posing reliability risks. How do LLMs process and prioritize mixed context? To study this, we introduce the Poisoned Context Testbed, pairing queries with real-world contexts containing relevant and inappropriate content. Inspired by associative learning in animals, we adapt the Rescorla-Wagner (RW) model from neuroscience to quantify how competing contextual signals influence LLM outputs. Our adapted model reveals a consistent behavioral pattern: LLMs exhibit a strong tendency to incorporate information that is less prevalent in the context. This susceptibility is harmful in real-world settings, where small amounts of inappropriate content can substantially degrade response quality. Empirical evaluations on our testbed further confirm this vulnerability. To tackle this, we introduce RW-Steering, a two-stage finetuning-based approach that enables the model to internally identify and ignore inappropriate signals. Unlike prior methods that rely on extensive supervision across diverse context mixtures, RW-Steering generalizes robustly across varying proportions of inappropriate content. Experiments show that our best fine-tuned model improves response quality by 39.8% and reverses the undesirable behavior curve, establishing RW-Steering as a robust, generalizable context engineering solution for improving LLM safety in real-world use.

Large Language Models (LLMs) have revolutionized artificial intelligence, driving advancements in machine translation, summarization, and conversational agents. However, their increasing integration into critical societal domains has raised concerns about embedded biases, which can perpetuate stereotypes and compromise fairness. These biases stem from various sources, including historical inequalities in training data, linguistic imbalances, and adversarial manipulation. Despite mitigation efforts, recent studies indicate that LLMs remain vulnerable to adversarial attacks designed to elicit biased responses. This work proposes a scalable benchmarking framework to evaluate LLM robustness against adversarial bias elicitation. Our methodology involves (i) systematically probing models with a multi-task approach targeting biases across various sociocultural dimensions, (ii) quantifying robustness through safety scores using an LLM-as-a-Judge approach for automated assessment of model responses, and (iii) employing jailbreak techniques to investigate vulnerabilities in safety mechanisms. Our analysis examines prevalent biases in both small and large state-of-the-art models and their impact on model safety. Additionally, we assess the safety of domain-specific models fine-tuned for critical fields, such as medicine. Finally, we release a curated dataset of bias-related prompts, CLEAR-Bias, to facilitate systematic vulnerability benchmarking. Our findings reveal critical trade-offs between model size and safety, aiding the development of fairer and more robust future language models.

Despite the recent advancements of vision-language-action (VLA) models on a variety of robotics tasks, they suffer from critical issues such as poor generalizability to unseen tasks, due to their reliance on behavior cloning exclusively from successful rollouts. Furthermore, they are typically fine-tuned to replicate demonstrations collected by experts under different settings, thus introducing distribution bias and limiting their adaptability to diverse manipulation objectives, such as efficiency, safety, and task completion. To bridge this gap, we introduce GRAPE: Generalizing Robot Policy via Preference Alignment. Specifically, GRAPE aligns VLAs on a trajectory level and implicitly models reward from both successful and failure trials to boost generalizability to diverse tasks. Moreover, GRAPE breaks down complex manipulation tasks to independent stages and automatically guides preference modeling through customized spatiotemporal constraints with keypoints proposed by a large vision-language model. Notably, these constraints are flexible and can be customized to align the model with varying objectives, such as safety, efficiency, or task success. We evaluate GRAPE across a diverse array of tasks in both real-world and simulated environments. Experimental results demonstrate that GRAPE enhances the performance of state-of-the-art VLA models, increasing success rates on in-domain and unseen manipulation tasks by 51.79% and 60.36%, respectively. Additionally, GRAPE can be aligned with various objectives, such as safety and efficiency, reducing collision rates by 44.31% and rollout step-length by 11.15%, respectively. All code, models, and data are available at https://grape-vla.github.io/

Llama-3.1-Sherkala-8B-Chat, or Sherkala-Chat (8B) for short, is a state-of-the-art instruction-tuned open generative large language model (LLM) designed for Kazakh. Sherkala-Chat (8B) aims to enhance the inclusivity of LLM advancements for Kazakh speakers. Adapted from the LLaMA-3.1-8B model, Sherkala-Chat (8B) is trained on 45.3B tokens across Kazakh, English, Russian, and Turkish. With 8 billion parameters, it demonstrates strong knowledge and reasoning abilities in Kazakh, significantly outperforming existing open Kazakh and multilingual models of similar scale while achieving competitive performance in English. We release Sherkala-Chat (8B) as an open-weight instruction-tuned model and provide a detailed overview of its training, fine-tuning, safety alignment, and evaluation, aiming to advance research and support diverse real-world applications.

Medical systematic reviews can be very costly and resource intensive. We explore how Large Language Models (LLMs) can support and be trained to perform literature screening when provided with a detailed set of selection criteria. Specifically, we instruction tune LLaMA and Guanaco models to perform abstract screening for medical systematic reviews. Our best model, Bio-SIEVE, outperforms both ChatGPT and trained traditional approaches, and generalises better across medical domains. However, there remains the challenge of adapting the model to safety-first scenarios. We also explore the impact of multi-task training with Bio-SIEVE-Multi, including tasks such as PICO extraction and exclusion reasoning, but find that it is unable to match single-task Bio-SIEVE's performance. We see Bio-SIEVE as an important step towards specialising LLMs for the biomedical systematic review process and explore its future developmental opportunities. We release our models, code and a list of DOIs to reconstruct our dataset for reproducibility.

The increasing rate of road accidents worldwide results not only in significant loss of life but also imposes billions financial burdens on societies. Current research in traffic crash frequency modeling and analysis has predominantly approached the problem as classification tasks, focusing mainly on learning-based classification or ensemble learning methods. These approaches often overlook the intricate relationships among the complex infrastructure, environmental, human and contextual factors related to traffic crashes and risky situations. In contrast, we initially propose a large-scale traffic crash language dataset, named CrashEvent, summarizing 19,340 real-world crash reports and incorporating infrastructure data, environmental and traffic textual and visual information in Washington State. Leveraging this rich dataset, we further formulate the crash event feature learning as a novel text reasoning problem and further fine-tune various large language models (LLMs) to predict detailed accident outcomes, such as crash types, severity and number of injuries, based on contextual and environmental factors. The proposed model, CrashLLM, distinguishes itself from existing solutions by leveraging the inherent text reasoning capabilities of LLMs to parse and learn from complex, unstructured data, thereby enabling a more nuanced analysis of contributing factors. Our experiments results shows that our LLM-based approach not only predicts the severity of accidents but also classifies different types of accidents and predicts injury outcomes, all with averaged F1 score boosted from 34.9% to 53.8%. Furthermore, CrashLLM can provide valuable insights for numerous open-world what-if situational-awareness traffic safety analyses with learned reasoning features, which existing models cannot offer. We make our benchmark, datasets, and model public available for further exploration.

Pre-trained transformers are often fine-tuned to aid clinical decision-making using limited clinical notes. Model interpretability is crucial, especially in high-stakes domains like medicine, to establish trust and ensure safety, which requires human engagement. We introduce SUFO, a systematic framework that enhances interpretability of fine-tuned transformer feature spaces. SUFO utilizes a range of analytic and visualization techniques, including Supervised probing, Unsupervised similarity analysis, Feature dynamics, and Outlier analysis to address key questions about model trust and interpretability. We conduct a case study investigating the impact of pre-training data where we focus on real-world pathology classification tasks, and validate our findings on MedNLI. We evaluate five 110M-sized pre-trained transformer models, categorized into general-domain (BERT, TNLR), mixed-domain (BioBERT, Clinical BioBERT), and domain-specific (PubMedBERT) groups. Our SUFO analyses reveal that: (1) while PubMedBERT, the domain-specific model, contains valuable information for fine-tuning, it can overfit to minority classes when class imbalances exist. In contrast, mixed-domain models exhibit greater resistance to overfitting, suggesting potential improvements in domain-specific model robustness; (2) in-domain pre-training accelerates feature disambiguation during fine-tuning; and (3) feature spaces undergo significant sparsification during this process, enabling clinicians to identify common outlier modes among fine-tuned models as demonstrated in this paper. These findings showcase the utility of SUFO in enhancing trust and safety when using transformers in medicine, and we believe SUFO can aid practitioners in evaluating fine-tuned language models for other applications in medicine and in more critical domains.

LLMs are known to be vulnerable to jailbreak attacks, even after safety alignment. An important observation is that, while different types of jailbreak attacks can generate significantly different queries, they mostly result in similar responses that are rooted in the same harmful knowledge (e.g., detailed steps to make a bomb). Therefore, we conjecture that directly unlearn the harmful knowledge in the LLM can be a more effective way to defend against jailbreak attacks than the mainstream supervised fine-tuning (SFT) based approaches. Our extensive experiments confirmed our insight and suggested surprising generalizability of our unlearning-based approach: using only 20 raw harmful questions without any jailbreak prompt during training, our solution reduced the Attack Success Rate (ASR) in Vicuna-7B on out-of-distribution (OOD) harmful questions wrapped with various complex jailbreak prompts from 82.6\% to 7.7\%. This significantly outperforms Llama2-7B-Chat, which is fine-tuned on about 0.1M safety alignment samples but still has an ASR of 21.9\% even under the help of an additional safety system prompt. Further analysis reveals that the generalization ability of our solution stems from the intrinsic relatedness among harmful responses across harmful questions (e.g., response patterns, shared steps and actions, and similarity among their learned representations in the LLM). Our code is available at https://github.com/thu-coai/SafeUnlearning.

Ensuring the safety of large language models (LLMs) is critical as they are deployed in real-world applications. Existing guardrails rely on rule-based filtering or single-pass classification, limiting their ability to handle nuanced safety violations. To address this, we propose ThinkGuard, a critique-augmented guardrail model that distills knowledge from high-capacity LLMs by generating structured critiques alongside safety labels. Fine-tuned on critique-augmented data, the captured deliberative thinking ability drastically enhances the guardrail's cautiousness and interpretability. Evaluated on multiple safety benchmarks, ThinkGuard achieves the highest average F1 and AUPRC, outperforming all baselines. Compared to LLaMA Guard 3, ThinkGuard improves accuracy by 16.1% and macro F1 by 27.0%. Moreover, it surpasses label-only fine-tuned models, confirming that structured critiques enhance both classification precision and nuanced safety reasoning while maintaining computational efficiency.

Text-to-Image models may generate harmful content, such as pornographic images, particularly when unsafe prompts are submitted. To address this issue, safety filters are often added on top of text-to-image models, or the models themselves are aligned to reduce harmful outputs. However, these defenses remain vulnerable when an attacker strategically designs adversarial prompts to bypass these safety guardrails. In this work, we propose PromptTune, a method to jailbreak text-to-image models with safety guardrails using a fine-tuned large language model. Unlike other query-based jailbreak attacks that require repeated queries to the target model, our attack generates adversarial prompts efficiently after fine-tuning our AttackLLM. We evaluate our method on three datasets of unsafe prompts and against five safety guardrails. Our results demonstrate that our approach effectively bypasses safety guardrails, outperforms existing no-box attacks, and also facilitates other query-based attacks.

The study of representation universality in AI models reveals growing convergence across domains, modalities, and architectures. However, the practical applications of representation universality remain largely unexplored. We bridge this gap by demonstrating that safety interventions can be transferred between models through learned mappings of their shared activation spaces. We demonstrate this approach on two well-established AI safety tasks: backdoor removal and refusal of harmful prompts, showing successful transfer of steering vectors that alter the models' outputs in a predictable way. Additionally, we propose a new task, corrupted capabilities, where models are fine-tuned to embed knowledge tied to a backdoor. This tests their ability to separate useful skills from backdoors, reflecting real-world challenges. Extensive experiments across Llama, Qwen and Gemma model families show that our method enables using smaller models to efficiently align larger ones. Furthermore, we demonstrate that autoencoder mappings between base and fine-tuned models can serve as reliable ``lightweight safety switches", allowing dynamic toggling between model behaviors.

The BigCode community, an open-scientific collaboration working on the responsible development of Large Language Models for Code (Code LLMs), introduces StarCoder and StarCoderBase: 15.5B parameter models with 8K context length, infilling capabilities and fast large-batch inference enabled by multi-query attention. StarCoderBase is trained on 1 trillion tokens sourced from The Stack, a large collection of permissively licensed GitHub repositories with inspection tools and an opt-out process. We fine-tuned StarCoderBase on 35B Python tokens, resulting in the creation of StarCoder. We perform the most comprehensive evaluation of Code LLMs to date and show that StarCoderBase outperforms every open Code LLM that supports multiple programming languages and matches or outperforms the OpenAI code-cushman-001 model. Furthermore, StarCoder outperforms every model that is fine-tuned on Python, can be prompted to achieve 40\% pass@1 on HumanEval, and still retains its performance on other programming languages. We take several important steps towards a safe open-access model release, including an improved PII redaction pipeline and a novel attribution tracing tool, and make the StarCoder models publicly available under a more commercially viable version of the Open Responsible AI Model license.

Model editing has become an increasingly popular alternative for efficiently updating knowledge within language models. Current methods mainly focus on reliability, generalization, and locality, with many methods excelling across these criteria. Some recent works disclose the pitfalls of these editing methods such as knowledge distortion or conflict. However, the general abilities of post-edited language models remain unexplored. In this paper, we perform a comprehensive evaluation on various editing methods and different language models, and have following findings. (1) Existing editing methods lead to inevitable performance deterioration on general benchmarks, indicating that existing editing methods maintain the general abilities of the model within only a few dozen edits. When the number of edits is slightly large, the intrinsic knowledge structure of the model is disrupted or even completely damaged. (2) Instruction-tuned models are more robust to editing, showing less performance drop on general knowledge after editing. (3) Language model with large scale is more resistant to editing compared to small model. (4) The safety of the edited model, is significantly weakened, even for those safety-aligned models. Our findings indicate that current editing methods are only suitable for small-scale knowledge updates within language models, which motivates further research on more practical and reliable editing methods. The details of code and reproduction can be found in https://github.com/lqinfdim/EditingEvaluation.

Deploying language models (LMs) necessitates outputs to be both high-quality and compliant with safety guidelines. Although Inference-Time Guardrails (ITG) offer solutions that shift model output distributions towards compliance, we find that current methods struggle in balancing safety with helpfulness. ITG Methods that safely address non-compliant queries exhibit lower helpfulness while those that prioritize helpfulness compromise on safety. We refer to this trade-off as the guardrail tax, analogous to the alignment tax. To address this, we propose PrimeGuard, a novel ITG method that utilizes structured control flow. PrimeGuard routes requests to different self-instantiations of the LM with varying instructions, leveraging its inherent instruction-following capabilities and in-context learning. Our tuning-free approach dynamically compiles system-designer guidelines for each query. We construct and release safe-eval, a diverse red-team safety benchmark. Extensive evaluations demonstrate that PrimeGuard, without fine-tuning, overcomes the guardrail tax by (1) significantly increasing resistance to iterative jailbreak attacks and (2) achieving state-of-the-art results in safety guardrailing while (3) matching helpfulness scores of alignment-tuned models. Extensive evaluations demonstrate that PrimeGuard, without fine-tuning, outperforms all competing baselines and overcomes the guardrail tax by improving the fraction of safe responses from 61% to 97% and increasing average helpfulness scores from 4.17 to 4.29 on the largest models, while reducing attack success rate from 100% to 8%. PrimeGuard implementation is available at https://github.com/dynamofl/PrimeGuard and safe-eval dataset is available at https://huggingface.co/datasets/dynamoai/safe_eval.

Despite the general capabilities of large pretrained language models, they consistently benefit from further adaptation to better achieve desired behaviors. However, tuning these models has become increasingly resource-intensive, or impossible when model weights are private. We introduce proxy-tuning, a lightweight decoding-time algorithm that operates on top of black-box LMs to achieve the result of directly tuning the model, but by accessing only its prediction over the output vocabulary. Our method instead tunes a smaller LM, then applies the difference between the predictions of the small tuned and untuned LMs to shift the original predictions of the base model in the direction of tuning, while retaining the benefits of larger scale pretraining. In experiments, when we apply proxy-tuning to Llama2-70B using proxies of only 7B size, we can close 88% of the gap between Llama2-70B and its truly-tuned chat version, when evaluated across knowledge, reasoning, and safety benchmarks. Interestingly, when tested on TruthfulQA, proxy-tuned models are actually more truthful than directly tuned models, possibly because decoding-time guidance better retains the model's factual knowledge. We then demonstrate the generality of proxy-tuning by applying it for domain adaptation on code, and task-specific finetuning on question-answering and math problems. Our work demonstrates the promise of using small tuned LMs to efficiently customize large, potentially proprietary LMs through decoding-time guidance.

Emerging large reasoning models (LRMs), such as DeepSeek-R1 models, leverage long chain-of-thought (CoT) reasoning to generate structured intermediate steps, enhancing their reasoning capabilities. However, long CoT does not inherently guarantee safe outputs, potentially leading to harmful consequences such as the introduction of security vulnerabilities in code or the spread of misinformation. Current research on large language model (LLM) safety usually focuses on short-answer responses, overlooking the long CoT style outputs of LRMs. To bridge this gap, we conduct a systematic study of LRM safety. First, we investigate safety evaluators calibrated against human annotations. Using our newly developed metrics, we thoroughly assess the safety of 12 state-of-the-art LRMs on StrongReject and WildJailbreak datasets. Our results show that LRMs are not safe compared to their reasoning advance. Further, we perform a fine-grained analysis of the reasoning trace and final answer. We find that three decoding strategies-ZeroThink, LessThink, and MoreThink-can improve model safety without additional training. However, these strategies either use constrained reasoning traces or incur high inference costs. To better strengthen LRM safety, we introduce SafeChain, the first-of-its-kind safety training dataset in CoT style. We fine-tune two LRMs with SafeChain, showing that it not only enhances model safety but also preserves performance across 6 reasoning benchmarks.

As large reasoning models (LRMs) grow more capable, chain-of-thought (CoT) reasoning introduces new safety challenges. Existing SFT-based safety alignment studies dominantly focused on filtering prompts with safe, high-quality responses, while overlooking hard prompts that always elicit harmful outputs. To fill this gap, we introduce UnsafeChain, a safety alignment dataset constructed from hard prompts with diverse sources, where unsafe completions are identified and explicitly corrected into safe responses. By exposing models to unsafe behaviors and guiding their correction, UnsafeChain enhances safety while preserving general reasoning ability. We fine-tune three LRMs on UnsafeChain and compare them against recent SafeChain and STAR-1 across six out-of-distribution and five in-distribution benchmarks. UnsafeChain consistently outperforms prior datasets, with even a 1K subset matching or surpassing baseline performance, demonstrating the effectiveness and generalizability of correction-based supervision. We release our dataset and code at https://github.com/mbzuai-nlp/UnsafeChain

Image content safety has become a significant challenge with the rise of visual media on online platforms. Meanwhile, in the age of AI-generated content (AIGC), many image generation models are capable of producing harmful content, such as images containing sexual or violent material. Thus, it becomes crucial to identify such unsafe images based on established safety rules. Pre-trained Multimodal Large Language Models (MLLMs) offer potential in this regard, given their strong pattern recognition abilities. Existing approaches typically fine-tune MLLMs with human-labeled datasets, which however brings a series of drawbacks. First, relying on human annotators to label data following intricate and detailed guidelines is both expensive and labor-intensive. Furthermore, users of safety judgment systems may need to frequently update safety rules, making fine-tuning on human-based annotation more challenging. This raises the research question: Can we detect unsafe images by querying MLLMs in a zero-shot setting using a predefined safety constitution (a set of safety rules)? Our research showed that simply querying pre-trained MLLMs does not yield satisfactory results. This lack of effectiveness stems from factors such as the subjectivity of safety rules, the complexity of lengthy constitutions, and the inherent biases in the models. To address these challenges, we propose a MLLM-based method includes objectifying safety rules, assessing the relevance between rules and images, making quick judgments based on debiased token probabilities with logically complete yet simplified precondition chains for safety rules, and conducting more in-depth reasoning with cascaded chain-of-thought processes if necessary. Experiment results demonstrate that our method is highly effective for zero-shot image safety judgment tasks.

Deep neural networks have achieved impressive experimental results in image classification, but can surprisingly be unstable with respect to adversarial perturbations, that is, minimal changes to the input image that cause the network to misclassify it. With potential applications including perception modules and end-to-end controllers for self-driving cars, this raises concerns about their safety. We develop a novel automated verification framework for feed-forward multi-layer neural networks based on Satisfiability Modulo Theory (SMT). We focus on safety of image classification decisions with respect to image manipulations, such as scratches or changes to camera angle or lighting conditions that would result in the same class being assigned by a human, and define safety for an individual decision in terms of invariance of the classification within a small neighbourhood of the original image. We enable exhaustive search of the region by employing discretisation, and propagate the analysis layer by layer. Our method works directly with the network code and, in contrast to existing methods, can guarantee that adversarial examples, if they exist, are found for the given region and family of manipulations. If found, adversarial examples can be shown to human testers and/or used to fine-tune the network. We implement the techniques using Z3 and evaluate them on state-of-the-art networks, including regularised and deep learning networks. We also compare against existing techniques to search for adversarial examples and estimate network robustness.

Malicious attackers can exploit large language models (LLMs) by engaging them in multi-turn dialogues to achieve harmful objectives, posing significant safety risks to society. To address this challenge, we propose a novel defense mechanism: SafeTy Reasoning Elicitation Alignment for Multi-Turn Dialogues (STREAM). STREAM defends LLMs against multi-turn attacks while preserving their functional capabilities. Our approach involves constructing a human-annotated dataset, the Safety Reasoning Multi-turn Dialogues dataset, which is used to fine-tune a plug-and-play safety reasoning moderator. This model is designed to identify malicious intent hidden within multi-turn conversations and alert the target LLM of potential risks. We evaluate STREAM across multiple LLMs against prevalent multi-turn attack strategies. Experimental results demonstrate that our method significantly outperforms existing defense techniques, reducing the Attack Success Rate (ASR) by 51.2%, all while maintaining comparable LLM capability.

Despite increasing efforts to ensure the safety of large language models (LLMs), most existing safety assessments and moderation tools remain heavily biased toward English and other high-resource languages, leaving majority of global languages underexamined. To address this gap, we introduce a manually annotated benchmark dataset for language model safety classification in Polish. We also create adversarially perturbed variants of these samples designed to challenge model robustness. We conduct a series of experiments to evaluate LLM-based and classifier-based models of varying sizes and architectures. Specifically, we fine-tune three models: Llama-Guard-3-8B, a HerBERT-based classifier (a Polish BERT derivative), and PLLuM, a Polish-adapted Llama-8B model. We train these models using different combinations of annotated data and evaluate their performance, comparing it against publicly available guard models. Results demonstrate that the HerBERT-based classifier achieves the highest overall performance, particularly under adversarial conditions.

In this paper we present AnswerCarefully, a dataset for promoting the safety and appropriateness of Japanese LLM outputs. The dataset consists of 1,800 pairs of questions and reference answers, where the questions require special attention in answering. It covers a wide range of risk categories established in prior English-language datasets, but the data samples are original in that they are manually created to reflect the socio-cultural context of LLM usage in Japan. We show that using this dataset for instruction to fine-tune a Japanese LLM led to improved output safety without compromising the utility of general responses. We also report the results of a safety evaluation of 12 Japanese LLMs using this dataset as a benchmark. Finally, we describe the latest update on the dataset which provides English translations and annotations of the questions, aimed at facilitating the derivation of similar datasets in different languages and regions.

Fine-tuning large language models (LLMs) on human preferences, typically through reinforcement learning from human feedback (RLHF), has proven successful in enhancing their capabilities. However, ensuring the safety of LLMs during the fine-tuning remains a critical concern, and mitigating the potential conflicts in safety and helpfulness is costly in RLHF. To address this issue, we propose a supervised learning framework called Bi-Factorial Preference Optimization (BFPO), which re-parameterizes a joint RLHF objective of both safety and helpfulness into a single supervised learning objective. In the supervised optimization, a labeling function is used to capture global preferences ranking to balance both safety and helpfulness. To evaluate BFPO, we develop a benchmark including comprehensive discriminative and generative tasks for helpfulness and harmlessness. The results indicate that our method significantly outperforms existing approaches in both safety and helpfulness. Moreover, BFPO eliminates the need for human prompting and annotation in LLM fine-tuning while achieving the same level of safety as methods that heavily rely on human labor, with less than 10% of the computational resources. The training recipes and models will be released.

As large language models (LLMs) continue to advance, ensuring their alignment with human values becomes increasingly critical. Traditional alignment methods heavily rely on human feedback to fine-tune models. With the emergence of superhuman models whose outputs may surpass human understanding, evaluating and aligning these models using human judgments poses significant challenges. To address the challenges, recent works use weak supervisors to elicit knowledge from much stronger models. However, there are important disanalogies between the empirical setup in the existing works and the genuine goal of alignment. We remark that existing works investigate the phenomenon of weak-to-strong generation in analogous setup (i.e., binary classification), rather than practical alignment-relevant tasks (e.g., safety). In this paper, we bridge this gap by extending weak-to-strong generation to the context of practical alignment. We empirically demonstrate the widespread phenomenon of weak-to-strong generation in three complicated alignment tasks: safety, toxicity, and legal reasoning}. Furthermore, we explore efficient strategies for improving alignment performance to enhance the quality of model outcomes. Lastly, we summarize and analyze the challenges and potential solutions in regard to specific alignment tasks, which we hope to catalyze the research progress on the topic of weak-to-strong generalization. Our code is released at https://github.com/yeruimeng/WTS.git.

As Large Language Models (LLMs) and generative AI become more widespread, the content safety risks associated with their use also increase. We find a notable deficiency in high-quality content safety datasets and benchmarks that comprehensively cover a wide range of critical safety areas. To address this, we define a broad content safety risk taxonomy, comprising 13 critical risk and 9 sparse risk categories. Additionally, we curate AEGISSAFETYDATASET, a new dataset of approximately 26, 000 human-LLM interaction instances, complete with human annotations adhering to the taxonomy. We plan to release this dataset to the community to further research and to help benchmark LLM models for safety. To demonstrate the effectiveness of the dataset, we instruction-tune multiple LLM-based safety models. We show that our models (named AEGISSAFETYEXPERTS), not only surpass or perform competitively with the state-of-the-art LLM-based safety models and general purpose LLMs, but also exhibit robustness across multiple jail-break attack categories. We also show how using AEGISSAFETYDATASET during the LLM alignment phase does not negatively impact the performance of the aligned models on MT Bench scores. Furthermore, we propose AEGIS, a novel application of a no-regret online adaptation framework with strong theoretical guarantees, to perform content moderation with an ensemble of LLM content safety experts in deployment

Construction safety inspections typically involve a human inspector identifying safety concerns on-site. With the rise of powerful Vision Language Models (VLMs), researchers are exploring their use for tasks such as detecting safety rule violations from on-site images. However, there is a lack of open datasets to comprehensively evaluate and further fine-tune VLMs in construction safety inspection. Current applications of VLMs use small, supervised datasets, limiting their applicability in tasks they are not directly trained for. In this paper, we propose the ConstructionSite 10k, featuring 10,000 construction site images with annotations for three inter-connected tasks, including image captioning, safety rule violation visual question answering (VQA), and construction element visual grounding. Our subsequent evaluation of current state-of-the-art large pre-trained VLMs shows notable generalization abilities in zero-shot and few-shot settings, while additional training is needed to make them applicable to actual construction sites. This dataset allows researchers to train and evaluate their own VLMs with new architectures and techniques, providing a valuable benchmark for construction safety inspection.

We introduce LlavaGuard, a family of VLM-based safeguard models, offering a versatile framework for evaluating the safety compliance of visual content. Specifically, we designed LlavaGuard for dataset annotation and generative model safeguarding. To this end, we collected and annotated a high-quality visual dataset incorporating a broad safety taxonomy, which we use to tune VLMs on context-aware safety risks. As a key innovation, LlavaGuard's new responses contain comprehensive information, including a safety rating, the violated safety categories, and an in-depth rationale. Further, our introduced customizable taxonomy categories enable the context-specific alignment of LlavaGuard to various scenarios. Our experiments highlight the capabilities of LlavaGuard in complex and real-world applications. We provide checkpoints ranging from 7B to 34B parameters demonstrating state-of-the-art performance, with even the smallest models outperforming baselines like GPT-4. We make our dataset and model weights publicly available and invite further research to address the diverse needs of communities and contexts.

Safety fine-tuning algorithms are commonly used to fine-tune language models to reduce harmful outputs, but the exact internal mechanisms of how those models achieve this remain unclear. In studying direct preference optimisation (DPO) for toxicity reduction, current explanations claim that DPO works by dampening the most toxic MLP neurons to learn an offset to avert toxic regions in the residual stream. However, by ablating the most toxic neurons and applying activation patching, we find this explanation incomplete. By projecting neuron activation changes onto a toxicity probe, we find that only 31.8\% of toxicity reduction comes from dampened toxic neurons. Instead, DPO reduces toxicity by accumulating effects across multiple neuron groups, both reducing writing in the toxic direction and promoting anti-toxicity in the residual stream. Moreover, DPO gives noisy adjustments to neuron activations, with many neurons actually increasing toxicity. This indicates that DPO is a balancing process between opposing neuron effects to achieve toxicity reduction.

Large language models (LLMs) frequently refuse to respond to pseudo-malicious instructions: semantically harmless input queries triggering unnecessary LLM refusals due to conservative safety alignment, significantly impairing user experience. Collecting such instructions is crucial for evaluating and mitigating over-refusals, but existing instruction curation methods, like manual creation or instruction rewriting, either lack scalability or fail to produce sufficiently diverse and effective refusal-inducing prompts. To address these limitations, we introduce EVOREFUSE, a prompt optimization approach that generates diverse pseudo-malicious instructions consistently eliciting confident refusals across LLMs. EVOREFUSE employs an evolutionary algorithm exploring the instruction space in more diverse directions than existing methods via mutation strategies and recombination, and iteratively evolves seed instructions to maximize evidence lower bound on LLM refusal probability. Using EVOREFUSE, we create two novel datasets: EVOREFUSE-TEST, a benchmark of 582 pseudo-malicious instructions that outperforms the next-best benchmark with 140.41% higher average refusal triggering rate across 9 LLMs, 34.86% greater lexical diversity, and 40.03% improved LLM response confidence scores; and EVOREFUSE-ALIGN, which provides 3,000 pseudo-malicious instructions with responses for supervised and preference-based alignment training. LLAMA3.1-8B-INSTRUCT supervisedly fine-tuned on EVOREFUSE-ALIGN achieves up to 14.31% fewer over-refusals than models trained on the second-best alignment dataset, without compromising safety. Our analysis with EVOREFUSE-TEST reveals models trigger over-refusals by overly focusing on sensitive keywords while ignoring broader context.

We have only limited understanding of how and why large language models (LLMs) respond in the ways that they do. Their neural networks have proven challenging to interpret, and we are only beginning to tease out the function of individual neurons and circuits within them. However, another path to understanding these systems is to investigate and develop their capacity to introspect and explain their own functioning. Here, we show that i) contemporary LLMs are capable of providing accurate, quantitative descriptions of their own internal processes during certain kinds of decision-making, ii) that it is possible to improve these capabilities through training, and iii) that this training generalizes to at least some degree. To do so, we fine-tuned GPT-4o and GPT-4o-mini to make decisions in a wide variety of complex contexts (e.g., choosing between condos, loans, vacations, etc.) according to randomly-generated, quantitative preferences about how to weigh different attributes during decision-making (e.g., the relative importance of natural light versus quiet surroundings for condos). We demonstrate that the LLMs can accurately report these preferences (i.e., the weights that they learned to give to different attributes during decision-making). Next, we demonstrate that these LLMs can be fine-tuned to explain their decision-making even more accurately. Finally, we demonstrate that this training generalizes: It improves the ability of the models to accurately explain what they are doing as they make other complex decisions, not just decisions they have learned to make via fine-tuning. This work is a step towards training LLMs to accurately and broadly report on their own internal processes -- a possibility that would yield substantial benefits for interpretability, control, and safety.

Safety and trustworthiness are indispensable requirements for real-world applications of AI systems using large language models (LLMs). This paper formulates human value alignment as an optimization problem of the language model policy to maximize reward under a safety constraint, and then proposes an algorithm, Stepwise Alignment for Constrained Policy Optimization (SACPO). One key idea behind SACPO, supported by theory, is that the optimal policy incorporating reward and safety can be directly obtained from a reward-aligned policy. Building on this key idea, SACPO aligns LLMs step-wise with each metric while leveraging simple yet powerful alignment algorithms such as direct preference optimization (DPO). SACPO offers several advantages, including simplicity, stability, computational efficiency, and flexibility of algorithms and datasets. Under mild assumptions, our theoretical analysis provides the upper bounds on optimality and safety constraint violation. Our experimental results show that SACPO can fine-tune Alpaca-7B better than the state-of-the-art method in terms of both helpfulness and harmlessness.

Large language models (LLMs) are susceptible to red teaming attacks, which can induce LLMs to generate harmful content. Previous research constructs attack prompts via manual or automatic methods, which have their own limitations on construction cost and quality. To address these issues, we propose an integrated approach that combines manual and automatic methods to economically generate high-quality attack prompts. Specifically, considering the impressive capabilities of newly emerged LLMs, we propose an attack framework to instruct LLMs to mimic human-generated prompts through in-context learning. Furthermore, we propose a defense framework that fine-tunes victim LLMs through iterative interactions with the attack framework to enhance their safety against red teaming attacks. Extensive experiments on different LLMs validate the effectiveness of our proposed attack and defense frameworks. Additionally, we release a series of attack prompts datasets named SAP with varying sizes, facilitating the safety evaluation and enhancement of more LLMs. Our code and dataset is available on https://github.com/Aatrox103/SAP .

Multimodal large language models (MLLMs) are critical for developing general-purpose AI assistants, yet they face growing safety risks. How can we ensure that MLLMs are safely aligned to prevent undesired behaviors such as discrimination, misinformation, or violations of ethical standards? In a further step, we need to explore how to fine-tune MLLMs to enhance reasoning performance while ensuring they satisfy safety constraints. Fundamentally, this can be formulated as a min-max optimization problem. In this study, we propose Safe RLHF-V, the first multimodal safety alignment framework that jointly optimizes helpfulness and safety using separate multimodal reward and cost models within a Lagrangian-based constrained optimization framework. Given that there is a lack of preference datasets that separate helpfulness and safety in multimodal scenarios, we introduce BeaverTails-V, the first open-source dataset with dual preference annotations for helpfulness and safety, along with multi-level safety labels (minor, moderate, severe). Additionally, we design a Multi-level Guardrail System to proactively defend against unsafe queries and adversarial attacks. By applying the Beaver-Guard-V moderation for 5 rounds of filtering and re-generation on the precursor model, the overall safety of the upstream model is significantly improved by an average of 40.9%. Experimental results demonstrate that fine-tuning different MLLMs with Safe RLHF can effectively enhance model helpfulness while ensuring improved safety. Specifically, Safe RLHF-V improves model safety by 34.2% and helpfulness by 34.3%. All of datasets, models, and code can be found at https://github.com/SafeRLHF-V to support the safety development of MLLMs and reduce potential societal risks.

Large Reasoning Models (LRMs) have become powerful tools for complex problem solving, but their structured reasoning pathways can lead to unsafe outputs when exposed to harmful prompts. Existing safety alignment methods reduce harmful outputs but can degrade reasoning depth, leading to significant trade-offs in complex, multi-step tasks, and remain vulnerable to sophisticated jailbreak attacks. To address this, we introduce SAFEPATH, a lightweight alignment method that fine-tunes LRMs to emit a short, 8-token Safety Primer at the start of their reasoning, in response to harmful prompts, while leaving the rest of the reasoning process unsupervised. Empirical results across multiple benchmarks indicate that SAFEPATH effectively reduces harmful outputs while maintaining reasoning performance. Specifically, SAFEPATH reduces harmful responses by up to 90.0% and blocks 83.3% of jailbreak attempts in the DeepSeek-R1-Distill-Llama-8B model, while requiring 295.9x less compute than Direct Refusal and 314.1x less than SafeChain. We further introduce a zero-shot variant that requires no fine-tuning. In addition, we provide a comprehensive analysis of how existing methods in LLMs generalize, or fail, when applied to reasoning-centric models, revealing critical gaps and new directions for safer AI.