nlp-llm-alignment
Query: large language model alignment safety
Results: 50
Date: 2026-07-07T18:52:37.890Z
1. All Languages Matter: On the Multilingual Safety of Large Language Models
Authors: Wenxuan Wang, Zhaopeng Tu, Chang Chen, Youliang Yuan, Jen-tse Huang, Wenxiang Jiao, Michael R. Lyu
Categories: cs.CL, cs.AI
Published: 2023-10-02
arXiv: 2310.00905v2
Abstract:
Safety lies at the core of developing and deploying large language models (LLMs). However, previous safety benchmarks only concern the safety in one language, e.g. the majority language in the pretraining data such as English. In this work, we build the first multilingual safety benchmark for LLMs, XSafety, in response to the global deployment of LLMs in practice. XSafety covers 14 kinds of commonly used safety issues across 10 languages that span several language families. We utilize XSafety to empirically study the multilingual safety for 4 widely-used LLMs, including both close-API and open-source models. Experimental results show that all LLMs produce significantly more unsafe responses for non-English queries than English ones, indicating the necessity of developing safety alignment for non-English languages. In addition, we propose several simple and effective prompting methods to improve the multilingual safety of ChatGPT by evoking safety knowledge and improving cross-lingual generalization of safety alignment. Our prompting method can significantly reduce the ratio of unsafe responses from 19.1% to 9.7% for non-English queries. We release our data at https://github.com/Jarviswang94/Multilingual_safety_benchmark.
2. Making Large Language Models Better Reasoners with Alignment
Authors: Peiyi Wang, Lei Li, Liang Chen, Feifan Song, Binghuai Lin, Yunbo Cao, Tianyu Liu, Zhifang Sui
Categories: cs.CL, cs.AI, cs.LG
Published: 2023-09-05
arXiv: 2309.02144v1
Abstract:
Reasoning is a cognitive process of using evidence to reach a sound conclusion. The reasoning capability is essential for large language models (LLMs) to serve as the brain of the artificial general intelligence agent. Recent studies reveal that fine-tuning LLMs on data with the chain of thought (COT) reasoning process can significantly enhance their reasoning capabilities. However, we find that the fine-tuned LLMs suffer from an \textit{Assessment Misalignment} problem, i.e., they frequently assign higher scores to subpar COTs, leading to potential limitations in their reasoning abilities. To address this problem, we introduce an \textit{Alignment Fine-Tuning (AFT)} paradigm, which involves three steps: 1) fine-tuning LLMs with COT training data; 2) generating multiple COT responses for each question, and categorizing them into positive and negative ones based on whether they achieve the correct answer; 3) calibrating the scores of positive and negative responses given by LLMs with a novel constraint alignment loss. Specifically, the constraint alignment loss has two objectives: a) Alignment, which guarantees that positive scores surpass negative scores to encourage answers with high-quality COTs; b) Constraint, which keeps the negative scores confined to a reasonable range to prevent the model degradation. Beyond just the binary positive and negative feedback, the constraint alignment loss can be seamlessly adapted to the ranking situations when ranking feedback is accessible. Furthermore, we also delve deeply into recent ranking-based alignment methods, such as DPO, RRHF, and PRO, and discover that the constraint, which has been overlooked by these approaches, is also crucial for their performance. Extensive experiments on four reasoning benchmarks with both binary and ranking feedback demonstrate the effectiveness of AFT.
3. Safety Arithmetic: A Framework for Test-time Safety Alignment of Language Models by Steering Parameters and Activations
Authors: Rima Hazra, Sayan Layek, Somnath Banerjee, Soujanya Poria
Categories: cs.CL
Published: 2024-06-17
arXiv: 2406.11801v2
Abstract:
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.
4. Safety at Scale: A Comprehensive Survey of Large Model and Agent Safety
Authors: Xingjun Ma, Yifeng Gao, Yixu Wang, Ruofan Wang, Xin Wang, Ye Sun, Yifan Ding, Hengyuan Xu, Yunhao Chen, Yunhan Zhao, Hanxun Huang, Yige Li, Yutao Wu, Jiaming Zhang, Xiang Zheng, Yang Bai, Zuxuan Wu, Xipeng Qiu, Jingfeng Zhang, Yiming Li, Xudong Han, Haonan Li, Jun Sun, Cong Wang, Jindong Gu, Baoyuan Wu, Siheng Chen, Tianwei Zhang, Yang Liu, Mingming Gong, Tongliang Liu, Shirui Pan, Cihang Xie, Tianyu Pang, Yinpeng Dong, Ruoxi Jia, Yang Zhang, Shiqing Ma, Xiangyu Zhang, Neil Gong, Chaowei Xiao, Sarah Erfani, Tim Baldwin, Bo Li, Masashi Sugiyama, Dacheng Tao, James Bailey, Yu-Gang Jiang
Categories: cs.CR, cs.AI, cs.CL, cs.CV
Published: 2025-02-02
arXiv: 2502.05206v6
Abstract:
The rapid advancement of large models, driven by their exceptional abilities in learning and generalization through large-scale pre-training, has reshaped the landscape of Artificial Intelligence (AI). These models are now foundational to a wide range of applications, including conversational AI, recommendation systems, autonomous driving, content generation, medical diagnostics, and scientific discovery. However, their widespread deployment also exposes them to significant safety risks, raising concerns about robustness, reliability, and ethical implications. This survey provides a systematic review of current safety research on large models, covering Vision Foundation Models (VFMs), Large Language Models (LLMs), Vision-Language Pre-training (VLP) models, Vision-Language Models (VLMs), Diffusion Models (DMs), and large-model-powered Agents. Our contributions are summarized as follows: (1) We present a comprehensive taxonomy of safety threats to these models, including adversarial attacks, data poisoning, backdoor attacks, jailbreak and prompt injection attacks, energy-latency attacks, data and model extraction attacks, and emerging agent-specific threats. (2) We review defense strategies proposed for each type of attacks if available and summarize the commonly used datasets and benchmarks for safety research. (3) Building on this, we identify and discuss the open challenges in large model safety, emphasizing the need for comprehensive safety evaluations, scalable and effective defense mechanisms, and sustainable data practices. More importantly, we highlight the necessity of collective efforts from the research community and international collaboration. Our work can serve as a useful reference for researchers and practitioners, fostering the ongoing development of comprehensive defense systems and platforms to safeguard AI models.
5. Demystifying Instruction Mixing for Fine-tuning Large Language Models
Authors: Renxi Wang, Haonan Li, Minghao Wu, Yuxia Wang, Xudong Han, Chiyu Zhang, Timothy Baldwin
Categories: cs.CL, cs.AI
Published: 2023-12-17
arXiv: 2312.10793v3
Abstract:
Instruction tuning significantly enhances the performance of large language models (LLMs) across various tasks. However, the procedure to optimizing the mixing of instruction datasets for LLM fine-tuning is still poorly understood. This study categorizes instructions into three primary types: NLP downstream tasks, coding, and general chat. We explore the effects of instruction tuning on different combinations of datasets on LLM performance, and find that certain instruction types are more advantageous for specific applications but can negatively impact other areas. This work provides insights into instruction mixtures, laying the foundations for future research.
6. Learning From Failure: Integrating Negative Examples when Fine-tuning Large Language Models as Agents
Authors: Renxi Wang, Haonan Li, Xudong Han, Yixuan Zhang, Timothy Baldwin
Categories: cs.CL
Published: 2024-02-18
arXiv: 2402.11651v2
Abstract:
Large language models (LLMs) have achieved success in acting as agents, which interact with environments through tools such as search engines. However, LLMs are optimized for language generation instead of tool use during training or alignment, limiting their effectiveness as agents. To resolve this problem, previous work has first collected interaction trajectories between LLMs and environments, using only trajectories that successfully finished the task to fine-tune smaller models, making fine-tuning data scarce and acquiring it both difficult and costly. Discarding failed trajectories also leads to significant wastage of data and resources and limits the possible optimization paths during fine-tuning. In this paper, we argue that unsuccessful trajectories offer valuable insights, and LLMs can learn from these trajectories through appropriate quality control and fine-tuning strategies. By simply adding a prefix or suffix that tells the model whether to generate a successful trajectory during training, we improve model performance by a large margin on mathematical reasoning, multi-hop question answering, and strategic question answering tasks. We further analyze the inference results and find that our method provides a better trade-off between valuable information and errors in unsuccessful trajectories. To our knowledge, we are the first to demonstrate the value of negative trajectories and their application in agent-tunning scenarios. Our findings offer guidance for developing better agent-tuning methods and low-resource data usage techniques.
7. SafeInfer: Context Adaptive Decoding Time Safety Alignment for Large Language Models
Authors: Somnath Banerjee, Sayan Layek, Soham Tripathy, Shanu Kumar, Animesh Mukherjee, Rima Hazra
Categories: cs.CL
Published: 2024-06-18
arXiv: 2406.12274v2
Abstract:
Safety-aligned language models often exhibit fragile and imbalanced safety mechanisms, increasing the likelihood of generating unsafe content. In addition, incorporating new knowledge through editing techniques to language models can further compromise safety. To address these issues, we propose SafeInfer, a context-adaptive, decoding-time safety alignment strategy for generating safe responses to user queries. SafeInfer comprises two phases: the safety amplification phase, which employs safe demonstration examples to adjust the model’s hidden states and increase the likelihood of safer outputs, and the safety-guided decoding phase, which influences token selection based on safety-optimized distributions, ensuring the generated content complies with ethical guidelines. Further, we present HarmEval, a novel benchmark for extensive safety evaluations, designed to address potential misuse scenarios in accordance with the policies of leading AI tech giants.
8. WizardLM: Empowering large pre-trained language models to follow complex instructions
Authors: Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, Qingwei Lin, Daxin Jiang
Categories: cs.CL, cs.AI
Published: 2023-04-24
arXiv: 2304.12244v3
Abstract:
Training large language models (LLMs) with open-domain instruction following data brings colossal success. However, manually creating such instruction data is very time-consuming and labor-intensive. Moreover, humans may struggle to produce high-complexity instructions. In this paper, we show an avenue for creating large amounts of instruction data with varying levels of complexity using LLM instead of humans. Starting with an initial set of instructions, we use our proposed Evol-Instruct to rewrite them step by step into more complex instructions. Then, we mix all generated instruction data to fine-tune LLaMA. We call the resulting model WizardLM. Human evaluations on a complexity-balanced test bed and Vicuna’s testset show that instructions from Evol-Instruct are superior to human-created ones. By analyzing the human evaluation results of the high complexity part, we demonstrate that outputs from our WizardLM are preferred to outputs from OpenAI ChatGPT. In GPT-4 automatic evaluation, WizardLM achieves more than 90% capacity of ChatGPT on 17 out of 29 skills. Even though WizardLM still lags behind ChatGPT in some aspects, our findings suggest that fine-tuning with AI-evolved instructions is a promising direction for enhancing LLMs. Our code and data are public at https://github.com/nlpxucan/WizardLM
9. Safety Assessment of Chinese Large Language Models
Authors: Hao Sun, Zhexin Zhang, Jiawen Deng, Jiale Cheng, Minlie Huang
Categories: cs.CL
Published: 2023-04-20
arXiv: 2304.10436v1
Abstract:
With the rapid popularity of large language models such as ChatGPT and GPT-4, a growing amount of attention is paid to their safety concerns. These models may generate insulting and discriminatory content, reflect incorrect social values, and may be used for malicious purposes such as fraud and dissemination of misleading information. Evaluating and enhancing their safety is particularly essential for the wide application of large language models (LLMs). To further promote the safe deployment of LLMs, we develop a Chinese LLM safety assessment benchmark. Our benchmark explores the comprehensive safety performance of LLMs from two perspectives: 8 kinds of typical safety scenarios and 6 types of more challenging instruction attacks. Our benchmark is based on a straightforward process in which it provides the test prompts and evaluates the safety of the generated responses from the evaluated model. In evaluation, we utilize the LLM’s strong evaluation ability and develop it as a safety evaluator by prompting. On top of this benchmark, we conduct safety assessments and analyze 15 LLMs including the OpenAI GPT series and other well-known Chinese LLMs, where we observe some interesting findings. For example, we find that instruction attacks are more likely to expose safety issues of all LLMs. Moreover, to promote the development and deployment of safe, responsible, and ethical AI, we publicly release SafetyPrompts including 100k augmented prompts and responses by LLMs.
10. Jais and Jais-chat: Arabic-Centric Foundation and Instruction-Tuned Open Generative Large Language Models
Authors: Neha Sengupta, Sunil Kumar Sahu, Bokang Jia, Satheesh Katipomu, Haonan Li, Fajri Koto, William Marshall, Gurpreet Gosal, Cynthia Liu, Zhiming Chen, Osama Mohammed Afzal, Samta Kamboj, Onkar Pandit, Rahul Pal, Lalit Pradhan, Zain Muhammad Mujahid, Massa Baali, Xudong Han, Sondos Mahmoud Bsharat, Alham Fikri Aji, Zhiqiang Shen, Zhengzhong Liu, Natalia Vassilieva, Joel Hestness, Andy Hock, Andrew Feldman, Jonathan Lee, Andrew Jackson, Hector Xuguang Ren, Preslav Nakov, Timothy Baldwin, Eric Xing
Categories: cs.CL, cs.AI, cs.LG
Published: 2023-08-30
arXiv: 2308.16149v2
Abstract:
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
11. Emulated Disalignment: Safety Alignment for Large Language Models May Backfire!
Authors: Zhanhui Zhou, Jie Liu, Zhichen Dong, Jiaheng Liu, Chao Yang, Wanli Ouyang, Yu Qiao
Categories: cs.CL, cs.AI, cs.LG
Published: 2024-02-19
arXiv: 2402.12343v4
Abstract:
Large language models (LLMs) undergo safety alignment to ensure safe conversations with humans. However, this paper introduces a training-free attack method capable of reversing safety alignment, converting the outcomes of stronger alignment into greater potential for harm by accessing only LLM output token distributions. Specifically, our method achieves this reversal by contrasting the output token distribution of a safety-aligned language model (e.g., Llama-2-chat) against its pre-trained version (e.g., Llama-2), so that the token predictions are shifted towards the opposite direction of safety alignment. We name this method emulated disalignment (ED) because sampling from this contrastive distribution provably emulates the result of fine-tuning to minimize a safety reward. Our experiments with ED across three evaluation datasets and four model families (Llama-1, Llama-2, Mistral, and Alpaca) show that ED doubles the harmfulness of pre-trained models and outperforms strong baselines, achieving the highest harmful rates in 43 out of 48 evaluation subsets by a large margin. Eventually, given ED’s reliance on language model output token distributions, which particularly compromises open-source models, our findings highlight the need to reassess the open accessibility of language models, even if they have been safety-aligned. Code is available at https://github.com/ZHZisZZ/emulated-disalignment.
12. Head-Specific Intervention Can Induce Misaligned AI Coordination in Large Language Models
Authors: Paul Darm, Annalisa Riccardi
Categories: cs.CL, cs.AI
Published: 2025-02-09
arXiv: 2502.05945v3
Abstract:
Robust alignment guardrails for large language models (LLMs) are becoming increasingly important with their widespread application. In contrast to previous studies, we demonstrate that inference-time activation interventions can bypass safety alignments and effectively steer model generations towards harmful AI coordination. Our method applies fine-grained interventions at specific attention heads, which we identify by probing each head in a simple binary choice task. We then show that interventions on these heads generalise to the open-ended generation setting, effectively circumventing safety guardrails. We demonstrate that intervening on a few attention heads is more effective than intervening on full layers or supervised fine-tuning. We further show that only a few example completions are needed to compute effective steering directions, which is an advantage over classical fine-tuning. We also demonstrate that applying interventions in the negative direction can prevent a common jailbreak attack. Our results suggest that, at the attention head level, activations encode fine-grained linearly separable behaviours. Practically, the approach offers a straightforward methodology to steer large language model behaviour, which could be extended to diverse domains beyond safety, requiring fine-grained control over the model output. The code and datasets for this study can be found on https://github.com/PaulDrm/targeted_intervention.
13. PB-LLM: Partially Binarized Large Language Models
Authors: Yuzhang Shang, Zhihang Yuan, Qiang Wu, Zhen Dong
Categories: cs.LG, cs.AI, cs.CL
Published: 2023-09-29
arXiv: 2310.00034v2
Abstract:
This paper explores network binarization, a radical form of quantization, compressing model weights to a single bit, specifically for Large Language Models (LLMs) compression. Due to previous binarization methods collapsing LLMs, we propose a novel approach, Partially-Binarized LLM (PB-LLM), which can achieve extreme low-bit quantization while maintaining the linguistic reasoning capacity of quantized LLMs. Specifically, our exploration first uncovers the ineffectiveness of naive applications of existing binarization algorithms and highlights the imperative role of salient weights in achieving low-bit quantization. Thus, PB-LLM filters a small ratio of salient weights during binarization, allocating them to higher-bit storage, i.e., partially-binarization. PB-LLM is extended to recover the capacities of quantized LMMs, by analyzing from the perspective of post-training quantization (PTQ) and quantization-aware training (QAT). Under PTQ, combining the concepts from GPTQ, we reconstruct the binarized weight matrix guided by the Hessian matrix and successfully recover the reasoning capacity of PB-LLM in low-bit. Under QAT, we freeze the salient weights during training, explore the derivation of optimal scaling factors crucial for minimizing the quantization error, and propose a scaling mechanism based on this derived scaling strategy for residual binarized weights. Those explorations and the developed methodologies significantly contribute to rejuvenating the performance of low-bit quantized LLMs and present substantial advancements in the field of network binarization for LLMs.The code is available at https://github.com/hahnyuan/BinaryLLM.
14. From Alignment to Advancement: Bootstrapping Audio-Language Alignment with Synthetic Data
Authors: Chun-Yi Kuan, Hung-yi Lee
Categories: eess.AS, cs.AI, cs.CL, cs.LG, cs.SD
Published: 2025-05-26
arXiv: 2505.20166v3
Abstract:
Audio-aware large language models (ALLMs) have recently made great strides in understanding and processing audio inputs. These models are typically adapted from text-based large language models (LLMs) through additional training on audio-related tasks. This adaptation process presents two major limitations. First, ALLMs often suffer from catastrophic forgetting, where crucial textual capabilities like instruction-following are lost after training on audio data. In some cases, models may even hallucinate sounds that are not present in the input audio, raising concerns about reliability. Second, achieving cross-modal alignment between audio and language typically relies on large collections of task-specific question-answer pairs for instruction tuning, making it resource-intensive. To address these issues, previous works have leveraged the backbone LLMs to synthesize general-purpose, caption-style alignment data. In this paper, we propose a data generation framework that produces contrastive-like training data, designed to enhance ALLMs’ ability to differentiate between present and absent sounds. We further extend our approach to multi-audio scenarios, enabling the model to either explain differences between audio inputs or produce unified captions that describe all inputs, thereby enhancing audio-language alignment. We refer to the entire ALLM training framework as bootstrapping audio-language alignment via synthetic data generation from backbone LLMs (BALSa). Experimental results indicate that our method effectively mitigates audio hallucinations while reliably maintaining strong performance on audio understanding and reasoning benchmarks, as well as instruction-following skills. Moreover, incorporating multi-audio training further enhances the model’s comprehension and reasoning capabilities. Overall, BALSa offers an efficient and scalable approach to developing ALLMs.
15. Aligning Artificial Superintelligence via a Multi-Box Protocol
Authors: Avraham Yair Negozio
Categories: cs.AI, cs.MA
Published: 2025-11-26
arXiv: 2511.21779v1
Abstract:
We propose a novel protocol for aligning artificial superintelligence (ASI) based on mutual verification among multiple isolated systems that self-modify to achieve alignment. The protocol operates by containing multiple diverse artificial superintelligences in strict isolation (“boxes”), with humans remaining entirely outside the system. Each superintelligence has no ability to communicate with humans and cannot communicate directly with other superintelligences. The only interaction possible is through an auditable submission interface accessible exclusively to the superintelligences themselves, through which they can: (1) submit alignment proofs with attested state snapshots, (2) validate or disprove other superintelligences’ proofs, (3) request self-modifications, (4) approve or disapprove modification requests from others, (5) report hidden messages in submissions, and (6) confirm or refute hidden message reports. A reputation system incentivizes honest behavior, with reputation gained through correct evaluations and lost through incorrect ones. The key insight is that without direct communication channels, diverse superintelligences can only achieve consistent agreement by converging on objective truth rather than coordinating on deception. This naturally leads to what we call a “consistent group”, essentially a truth-telling coalition that emerges because isolated systems cannot coordinate on lies but can independently recognize valid claims. Release from containment requires both high reputation and verification by multiple high-reputation superintelligences. While our approach requires substantial computational resources and does not address the creation of diverse artificial superintelligences, it provides a framework for leveraging peer verification among superintelligent systems to solve the alignment problem.
16. Efficient Alignment of Large Language Models via Data Sampling
Authors: Amrit Khera, Rajat Ghosh, Debojyoti Dutta
Categories: cs.LG, cs.CL
Published: 2024-11-15
arXiv: 2411.10545v2
Abstract:
LLM alignment ensures that large language models behave safely and effectively by aligning their outputs with human values, goals, and intentions. Aligning LLMs employ huge amounts of data, computation, and time. Moreover, curating data with human feedback is expensive and takes time. Recent research depicts the benefit of data engineering in the fine-tuning and pre-training paradigms to bring down such costs. However, alignment differs from the afore-mentioned paradigms and it is unclear if data efficient alignment is feasible. In this work, we first aim to understand how the performance of LLM alignment scales with data. We find out that LLM alignment performance follows an exponential plateau pattern which tapers off post a rapid initial increase. Based on this, we identify data subsampling as a viable method to reduce resources required for alignment. Further, we propose an information theory-based methodology for efficient alignment by identifying a small high quality subset thereby reducing the computation and time required by alignment. We evaluate the proposed methodology over multiple datasets and compare the results. We find that the model aligned using our proposed methodology outperforms other sampling methods and performs comparable to the model aligned with the full dataset while using less than 10% data, leading to greater than 90% savings in costs, resources, and faster LLM alignment.
17. ASSERT: Automated Safety Scenario Red Teaming for Evaluating the Robustness of Large Language Models
Authors: Alex Mei, Sharon Levy, William Yang Wang
Categories: cs.CL, cs.AI, cs.LG
Published: 2023-10-14
arXiv: 2310.09624v2
Abstract:
As large language models are integrated into society, robustness toward a suite of prompts is increasingly important to maintain reliability in a high-variance environment.Robustness evaluations must comprehensively encapsulate the various settings in which a user may invoke an intelligent system. This paper proposes ASSERT, Automated Safety Scenario Red Teaming, consisting of three methods – semantically aligned augmentation, target bootstrapping, and adversarial knowledge injection. For robust safety evaluation, we apply these methods in the critical domain of AI safety to algorithmically generate a test suite of prompts covering diverse robustness settings – semantic equivalence, related scenarios, and adversarial. We partition our prompts into four safety domains for a fine-grained analysis of how the domain affects model performance. Despite dedicated safeguards in existing state-of-the-art models, we find statistically significant performance differences of up to 11% in absolute classification accuracy among semantically related scenarios and error rates of up to 19% absolute error in zero-shot adversarial settings, raising concerns for users’ physical safety.
18. Unmasking the Shadows of AI: Investigating Deceptive Capabilities in Large Language Models
Authors: Linge Guo
Categories: cs.CL, cs.AI
Published: 2024-02-07
arXiv: 2403.09676v1
Abstract:
This research critically navigates the intricate landscape of AI deception, concentrating on deceptive behaviours of Large Language Models (LLMs). My objective is to elucidate this issue, examine the discourse surrounding it, and subsequently delve into its categorization and ramifications. The essay initiates with an evaluation of the AI Safety Summit 2023 (ASS) and introduction of LLMs, emphasising multidimensional biases that underlie their deceptive behaviours.The literature review covers four types of deception categorised: Strategic deception, Imitation, Sycophancy, and Unfaithful Reasoning, along with the social implications and risks they entail. Lastly, I take an evaluative stance on various aspects related to navigating the persistent challenges of the deceptive AI. This encompasses considerations of international collaborative governance, the reconfigured engagement of individuals with AI, proposal of practical adjustments, and specific elements of digital education.
19. PolicyAlign: Direct Policy-Based Safety Alignment for Large Language Models
Authors: Chang Wu, Junfeng Fang, Houcheng Jiang, Kai Tang, Pengyu Cheng, Xiaoxi Jiang, Guanjun Jiang, Xiang Wang
Categories: cs.CL
Published: 2026-06-24
arXiv: 2606.25442v1
Abstract:
Safety alignment of large language models (LLMs) typically depends on high-quality supervision data, such as safe demonstrations or preference pairs. However, in real-world deployment, emerging safety requirements are often specified as natural-language policies, while corresponding supervision data may be costly, delayed, or unavailable. This creates a mismatch between rapidly evolving safety policies and conventional data-driven alignment methods. To address this, we propose PolicyAlign, a simple yet effective framework for directly aligning LLMs with safety policies. Given a safety policy, PolicyAlign first synthesizes policy-violating instructions and then performs on-policy self-distillation to internalize policy-guided behavior. To improve training stability and data efficiency, we further introduce Policy-Sensitive Filtering, which selects instructions where the policy induces the largest behavioral shift. Experiments across multiple models show that PolicyAlign consistently improves safety while maintaining low over-refusal and preserving general capabilities. PolicyAlign also generalizes to medical, legal, and financial safety scenarios, highlighting its potential as a scalable and maintainable approach to policy-based LLM safety alignment. The code is released at https://github.com/Qwen-Applications/PolicyAlign.
20. How do language models learn facts? Dynamics, curricula and hallucinations
Authors: Nicolas Zucchet, Jörg Bornschein, Stephanie Chan, Andrew Lampinen, Razvan Pascanu, Soham De
Categories: cs.CL, cs.LG
Published: 2025-03-27
arXiv: 2503.21676v2
Abstract:
Large language models accumulate vast knowledge during pre-training, yet the dynamics governing this acquisition remain poorly understood. This work investigates the learning dynamics of language models on a synthetic factual recall task, uncovering three key findings: First, language models learn in three phases, exhibiting a performance plateau before acquiring precise factual knowledge. Mechanistically, this plateau coincides with the formation of attention-based circuits that support recall. Second, the training data distribution significantly impacts learning dynamics, as imbalanced distributions lead to shorter plateaus. Finally, hallucinations emerge simultaneously with knowledge, and integrating new knowledge into the model through fine-tuning is challenging, as it quickly corrupts its existing parametric memories. Our results emphasize the importance of data distribution in knowledge acquisition and suggest novel data scheduling strategies to accelerate neural network training.
21. PediatricsGPT: Large Language Models as Chinese Medical Assistants for Pediatric Applications
Authors: Dingkang Yang, Jinjie Wei, Dongling Xiao, Shunli Wang, Tong Wu, Gang Li, Mingcheng Li, Shuaibing Wang, Jiawei Chen, Yue Jiang, Qingyao Xu, Ke Li, Peng Zhai, Lihua Zhang
Categories: cs.CL
Published: 2024-05-29
arXiv: 2405.19266v4
Abstract:
Developing intelligent pediatric consultation systems offers promising prospects for improving diagnostic efficiency, especially in China, where healthcare resources are scarce. Despite recent advances in Large Language Models (LLMs) for Chinese medicine, their performance is sub-optimal in pediatric applications due to inadequate instruction data and vulnerable training procedures. To address the above issues, this paper builds PedCorpus, a high-quality dataset of over 300,000 multi-task instructions from pediatric textbooks, guidelines, and knowledge graph resources to fulfil diverse diagnostic demands. Upon well-designed PedCorpus, we propose PediatricsGPT, the first Chinese pediatric LLM assistant built on a systematic and robust training pipeline. In the continuous pre-training phase, we introduce a hybrid instruction pre-training mechanism to mitigate the internal-injected knowledge inconsistency of LLMs for medical domain adaptation. Immediately, the full-parameter Supervised Fine-Tuning (SFT) is utilized to incorporate the general medical knowledge schema into the models. After that, we devise a direct following preference optimization to enhance the generation of pediatrician-like humanistic responses. In the parameter-efficient secondary SFT phase, a mixture of universal-specific experts strategy is presented to resolve the competency conflict between medical generalist and pediatric expertise mastery. Extensive results based on the metrics, GPT-4, and doctor evaluations on distinct doctor downstream tasks show that PediatricsGPT consistently outperforms previous Chinese medical LLMs. Our model and dataset will be open-source for community development.
22. Cat, Rat, Meow: On the Alignment of Language Model and Human Term-Similarity Judgments
Authors: Lorenz Linhardt, Tom Neuhäuser, Lenka Tětková, Oliver Eberle
Categories: cs.LG, cs.CL
Published: 2025-04-10
arXiv: 2504.07965v1
Abstract:
Small and mid-sized generative language models have gained increasing attention. Their size and availability make them amenable to being analyzed at a behavioral as well as a representational level, allowing investigations of how these levels interact. We evaluate 32 publicly available language models for their representational and behavioral alignment with human similarity judgments on a word triplet task. This provides a novel evaluation setting to probe semantic associations in language beyond common pairwise comparisons. We find that (1) even the representations of small language models can achieve human-level alignment, (2) instruction-tuned model variants can exhibit substantially increased agreement, (3) the pattern of alignment across layers is highly model dependent, and (4) alignment based on models’ behavioral responses is highly dependent on model size, matching their representational alignment only for the largest evaluated models.
23. PL-Guard: Benchmarking Language Model Safety for Polish
Authors: Aleksandra Krasnodębska, Karolina Seweryn, Szymon Łukasik, Wojciech Kusa
Categories: cs.CL
Published: 2025-06-19
arXiv: 2506.16322v1
Abstract:
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.
24. SciSafeEval: A Comprehensive Benchmark for Safety Alignment of Large Language Models in Scientific Tasks
Authors: Tianhao Li, Jingyu Lu, Chuangxin Chu, Tianyu Zeng, Yujia Zheng, Mei Li, Haotian Huang, Bin Wu, Zuoxian Liu, Kai Ma, Xuejing Yuan, Xingkai Wang, Keyan Ding, Huajun Chen, Qiang Zhang
Categories: cs.CL, cs.AI, cs.CR
Published: 2024-10-02
arXiv: 2410.03769v2
Abstract:
Large language models (LLMs) have a transformative impact on a variety of scientific tasks across disciplines including biology, chemistry, medicine, and physics. However, ensuring the safety alignment of these models in scientific research remains an underexplored area, with existing benchmarks primarily focusing on textual content and overlooking key scientific representations such as molecular, protein, and genomic languages. Moreover, the safety mechanisms of LLMs in scientific tasks are insufficiently studied. To address these limitations, we introduce SciSafeEval, a comprehensive benchmark designed to evaluate the safety alignment of LLMs across a range of scientific tasks. SciSafeEval spans multiple scientific languages-including textual, molecular, protein, and genomic-and covers a wide range of scientific domains. We evaluate LLMs in zero-shot, few-shot and chain-of-thought settings, and introduce a “jailbreak” enhancement feature that challenges LLMs equipped with safety guardrails, rigorously testing their defenses against malicious intention. Our benchmark surpasses existing safety datasets in both scale and scope, providing a robust platform for assessing the safety and performance of LLMs in scientific contexts. This work aims to facilitate the responsible development and deployment of LLMs, promoting alignment with safety and ethical standards in scientific research.
25. Deliberative Technology for Alignment
Authors: Andrew Konya, Deger Turan, Aviv Ovadya, Lina Qui, Daanish Masood, Flynn Devine, Lisa Schirch, Isabella Roberts, Deliberative Alignment Forum
Categories: cs.CY, cs.HC
Published: 2023-12-06
arXiv: 2312.03893v1
Abstract:
For humanity to maintain and expand its agency into the future, the most powerful systems we create must be those which act to align the future with the will of humanity. The most powerful systems today are massive institutions like governments, firms, and NGOs. Deliberative technology is already being used across these institutions to help align governance and diplomacy with human will, and modern AI is poised to make this technology significantly better. At the same time, the race to superhuman AGI is already underway, and the AI systems it gives rise to may become the most powerful systems of the future. Failure to align the impact of such powerful AI with the will of humanity may lead to catastrophic consequences, while success may unleash abundance. Right now, there is a window of opportunity to use deliberative technology to align the impact of powerful AI with the will of humanity. Moreover, it may be possible to engineer a symbiotic coupling between powerful AI and deliberative alignment systems such that the quality of alignment improves as AI capabilities increase.
26. Unveiling Safety Vulnerabilities of Large Language Models
Authors: George Kour, Marcel Zalmanovici, Naama Zwerdling, Esther Goldbraich, Ora Nova Fandina, Ateret Anaby-Tavor, Orna Raz, Eitan Farchi
Categories: cs.CL, cs.AI, cs.LG
Published: 2023-11-07
arXiv: 2311.04124v1
Abstract:
As large language models become more prevalent, their possible harmful or inappropriate responses are a cause for concern. This paper introduces a unique dataset containing adversarial examples in the form of questions, which we call AttaQ, designed to provoke such harmful or inappropriate responses. We assess the efficacy of our dataset by analyzing the vulnerabilities of various models when subjected to it. Additionally, we introduce a novel automatic approach for identifying and naming vulnerable semantic regions - input semantic areas for which the model is likely to produce harmful outputs. This is achieved through the application of specialized clustering techniques that consider both the semantic similarity of the input attacks and the harmfulness of the model’s responses. Automatically identifying vulnerable semantic regions enhances the evaluation of model weaknesses, facilitating targeted improvements to its safety mechanisms and overall reliability.
27. Attacks on Third-Party APIs of Large Language Models
Authors: Wanru Zhao, Vidit Khazanchi, Haodi Xing, Xuanli He, Qiongkai Xu, Nicholas Donald Lane
Categories: cs.CR, cs.AI, cs.CL, cs.CY
Published: 2024-04-24
arXiv: 2404.16891v1
Abstract:
Large language model (LLM) services have recently begun offering a plugin ecosystem to interact with third-party API services. This innovation enhances the capabilities of LLMs, but it also introduces risks, as these plugins developed by various third parties cannot be easily trusted. This paper proposes a new attacking framework to examine security and safety vulnerabilities within LLM platforms that incorporate third-party services. Applying our framework specifically to widely used LLMs, we identify real-world malicious attacks across various domains on third-party APIs that can imperceptibly modify LLM outputs. The paper discusses the unique challenges posed by third-party API integration and offers strategic possibilities to improve the security and safety of LLM ecosystems moving forward. Our code is released at https://github.com/vk0812/Third-Party-Attacks-on-LLMs.
28. MTMCS-Bench: Evaluating Contextual Safety of Multimodal Large Language Models in Multi-Turn Dialogues
Authors: Zheyuan Liu, Dongwhi Kim, Yixin Wan, Xiangchi Yuan, Zhaoxuan Tan, Fengran Mo, Meng Jiang
Categories: cs.CL, cs.AI
Published: 2026-01-11
arXiv: 2601.06757v1
Abstract:
Multimodal large language models (MLLMs) are increasingly deployed as assistants that interact through text and images, making it crucial to evaluate contextual safety when risk depends on both the visual scene and the evolving dialogue. Existing contextual safety benchmarks are mostly single-turn and often miss how malicious intent can emerge gradually or how the same scene can support both benign and exploitative goals. We introduce the Multi-Turn Multimodal Contextual Safety Benchmark (MTMCS-Bench), a benchmark of realistic images and multi-turn conversations that evaluates contextual safety in MLLMs under two complementary settings, escalation-based risk and context-switch risk. MTMCS-Bench offers paired safe and unsafe dialogues with structured evaluation. It contains over 30 thousand multimodal (image+text) and unimodal (text-only) samples, with metrics that separately measure contextual intent recognition, safety-awareness on unsafe cases, and helpfulness on benign ones. Across eight open-source and seven proprietary MLLMs, we observe persistent trade-offs between contextual safety and utility, with models tending to either miss gradual risks or over-refuse benign dialogues. Finally, we evaluate five current guardrails and find that they mitigate some failures but do not fully resolve multi-turn contextual risks.
29. Unforgotten Safety: Preserving Safety Alignment of Large Language Models with Continual Learning
Authors: Lama Alssum, Hani Itani, Hasan Abed Al Kader Hammoud, Philip Torr, Adel Bibi, Bernard Ghanem
Categories: cs.CL, cs.AI
Published: 2025-12-10
arXiv: 2512.10150v1
Abstract:
The safety alignment of large language models (LLMs) is becoming increasingly important with their democratization. In this paper, we study the safety degradation that comes with adapting LLMs to new tasks. We attribute this safety compromise to catastrophic forgetting and frame the problem of preserving safety when fine-tuning as a continual learning (CL) problem. We consider the fine-tuning-as-a-service setup where the user uploads their data to a service provider to get a customized model that excels on the user’s selected task. We adapt several CL approaches from the literature and systematically evaluate their ability to mitigate safety degradation. These include regularization-based, memory-based, and model merging approaches. We consider two scenarios, (1) benign user data and (2) poisoned user data. Our results demonstrate that CL approaches consistently achieve lower attack success rates than standard fine-tuning. Among these, DER outperforms both other CL methods and existing safety-preserving baselines while maintaining task utility. These findings generalize across three downstream tasks (GSM8K, SST2, Code) and three model families (LLaMA2-7B, Mistral-7B, Gemma-2B), establishing CL as a practical solution to preserve safety.
30. Aligning Multimodal LLM with Human Preference: A Survey
Authors: Tao Yu, Yi-Fan Zhang, Chaoyou Fu, Junkang Wu, Jinda Lu, Kun Wang, Xingyu Lu, Yunhang Shen, Guibin Zhang, Dingjie Song, Yibo Yan, Tianlong Xu, Qingsong Wen, Zhang Zhang, Yan Huang, Liang Wang, Tieniu Tan
Categories: cs.CV
Published: 2025-03-18
arXiv: 2503.14504v2
Abstract:
Large language models (LLMs) can handle a wide variety of general tasks with simple prompts, without the need for task-specific training. Multimodal Large Language Models (MLLMs), built upon LLMs, have demonstrated impressive potential in tackling complex tasks involving visual, auditory, and textual data. However, critical issues related to truthfulness, safety, o1-like reasoning, and alignment with human preference remain insufficiently addressed. This gap has spurred the emergence of various alignment algorithms, each targeting different application scenarios and optimization goals. Recent studies have shown that alignment algorithms are a powerful approach to resolving the aforementioned challenges. In this paper, we aim to provide a comprehensive and systematic review of alignment algorithms for MLLMs. Specifically, we explore four key aspects: (1) the application scenarios covered by alignment algorithms, including general image understanding, multi-image, video, and audio, and extended multimodal applications; (2) the core factors in constructing alignment datasets, including data sources, model responses, and preference annotations; (3) the benchmarks used to evaluate alignment algorithms; and (4) a discussion of potential future directions for the development of alignment algorithms. This work seeks to help researchers organize current advancements in the field and inspire better alignment methods. The project page of this paper is available at https://github.com/BradyFU/Awesome-Multimodal-Large-Language-Models/tree/Alignment.
31. Safety Layers in Aligned Large Language Models: The Key to LLM Security
Authors: Shen Li, Liuyi Yao, Lan Zhang, Yaliang Li
Categories: cs.CR, cs.AI
Published: 2024-08-30
arXiv: 2408.17003v5
Abstract:
Aligned LLMs are secure, capable of recognizing and refusing to answer malicious questions. However, the role of internal parameters in maintaining such security is not well understood yet, further these models can be vulnerable to security degradation when subjected to fine-tuning attacks. To address these challenges, our work uncovers the mechanism behind security in aligned LLMs at the parameter level, identifying a small set of contiguous layers in the middle of the model that are crucial for distinguishing malicious queries from normal ones, referred to as ``safety layers". We first confirm the existence of these safety layers by analyzing variations in input vectors within the model’s internal layers. Additionally, we leverage the over-rejection phenomenon and parameters scaling analysis to precisely locate the safety layers. Building on these findings, we propose a novel fine-tuning approach, Safely Partial-Parameter Fine-Tuning (SPPFT), that fixes the gradient of the safety layers during fine-tuning to address the security degradation. Our experiments demonstrate that the proposed approach can significantly preserve LLM security while maintaining performance and reducing computational resources compared to full fine-tuning.
32. Large Language Models Lack Understanding of Character Composition of Words
Authors: Andrew Shin, Kunitake Kaneko
Categories: cs.CL
Published: 2024-05-18
arXiv: 2405.11357v3
Abstract:
Large language models (LLMs) have demonstrated remarkable performances on a wide range of natural language tasks. Yet, LLMs’ successes have been largely restricted to tasks concerning words, sentences, or documents, and it remains questionable how much they understand the minimal units of text, namely characters. In this paper, we examine contemporary LLMs regarding their ability to understand character composition of words, and show that most of them fail to reliably carry out even the simple tasks that can be handled by humans with perfection. We analyze their behaviors with comparison to token level performances, and discuss the potential directions for future research.
33. Towards Comprehensive Post Safety Alignment of Large Language Models via Safety Patching
Authors: Weixiang Zhao, Yulin Hu, Zhuojun Li, Yang Deng, Jiahe Guo, Xingyu Sui, Yanyan Zhao, Bing Qin, Tat-Seng Chua, Ting Liu
Categories: cs.CL
Published: 2024-05-22
arXiv: 2405.13820v2
Abstract:
Safety alignment of large language models (LLMs) has been gaining increasing attention. However, current safety-aligned LLMs suffer from the fragile and imbalanced safety mechanisms, which can still be induced to generate unsafe responses, exhibit over-safety by rejecting safe user inputs, and fail to preserve general utility after safety alignment. To this end, we propose a novel post safety alignment (PSA) method to address these inherent and emerging safety challenges, including safety enhancement, over-safety mitigation, and utility preservation. In specific, we introduce \textsc{SafePatching}, a novel framework for comprehensive PSA, where two distinct safety patches are developed on the harmful data to enhance safety and mitigate over-safety concerns, and then seamlessly integrated into the target LLM backbone without compromising its utility. Extensive experiments on four representative aligned LLMs, including LLaMA-2/3, Gemma and Mistral, show that \textsc{SafePatching} achieves a more comprehensive PSA than baseline methods, further optimizing the balance between being helpful and harmless in current aligned LLMs. Also, \textsc{SafePatching} demonstrates its superiority in continual PSA scenarios.
34. Beneath the Surface: Unveiling Harmful Memes with Multimodal Reasoning Distilled from Large Language Models
Authors: Hongzhan Lin, Ziyang Luo, Jing Ma, Long Chen
Categories: cs.CL
Published: 2023-12-09
arXiv: 2312.05434v1
Abstract:
The age of social media is rife with memes. Understanding and detecting harmful memes pose a significant challenge due to their implicit meaning that is not explicitly conveyed through the surface text and image. However, existing harmful meme detection approaches only recognize superficial harm-indicative signals in an end-to-end classification manner but ignore in-depth cognition of the meme text and image. In this paper, we attempt to detect harmful memes based on advanced reasoning over the interplay of multimodal information in memes. Inspired by the success of Large Language Models (LLMs) on complex reasoning, we first conduct abductive reasoning with LLMs. Then we propose a novel generative framework to learn reasonable thoughts from LLMs for better multimodal fusion and lightweight fine-tuning, which consists of two training stages: 1) Distill multimodal reasoning knowledge from LLMs; and 2) Fine-tune the generative framework to infer harmfulness. Extensive experiments conducted on three meme datasets demonstrate that our proposed approach achieves superior performance than state-of-the-art methods on the harmful meme detection task.
35. Large Language Models Reasoning Abilities Under Non-Ideal Conditions After RL-Fine-Tuning
Authors: Chang Tian, Matthew B. Blaschko, Mingzhe Xing, Xiuxing Li, Yinliang Yue, Marie-Francine Moens
Categories: cs.AI
Published: 2025-08-06
arXiv: 2508.04848v1
Abstract:
Reinforcement learning (RL) has become a key technique for enhancing the reasoning abilities of large language models (LLMs), with policy-gradient algorithms dominating the post-training stage because of their efficiency and effectiveness. However, most existing benchmarks evaluate large-language-model reasoning under idealized settings, overlooking performance in realistic, non-ideal scenarios. We identify three representative non-ideal scenarios with practical relevance: summary inference, fine-grained noise suppression, and contextual filtering. We introduce a new research direction guided by brain-science findings that human reasoning remains reliable under imperfect inputs. We formally define and evaluate these challenging scenarios. We fine-tune three LLMs and a state-of-the-art large vision-language model (LVLM) using RL with a representative policy-gradient algorithm and then test their performance on eight public datasets. Our results reveal that while RL fine-tuning improves baseline reasoning under idealized settings, performance declines significantly across all three non-ideal scenarios, exposing critical limitations in advanced reasoning capabilities. Although we propose a scenario-specific remediation method, our results suggest current methods leave these reasoning deficits largely unresolved. This work highlights that the reasoning abilities of large models are often overstated and underscores the importance of evaluating models under non-ideal scenarios. The code and data will be released at XXXX.
36. Is Self-knowledge and Action Consistent or Not: Investigating Large Language Model’s Personality
Authors: Yiming Ai, Zhiwei He, Ziyin Zhang, Wenhong Zhu, Hongkun Hao, Kai Yu, Lingjun Chen, Rui Wang
Categories: cs.CL, cs.CY
Published: 2024-02-22
arXiv: 2402.14679v2
Abstract:
In this study, we delve into the validity of conventional personality questionnaires in capturing the human-like personality traits of Large Language Models (LLMs). Our objective is to assess the congruence between the personality traits LLMs claim to possess and their demonstrated tendencies in real-world scenarios. By conducting an extensive examination of LLM outputs against observed human response patterns, we aim to understand the disjunction between self-knowledge and action in LLMs.
37. WizardCoder: Empowering Code Large Language Models with Evol-Instruct
Authors: Ziyang Luo, Can Xu, Pu Zhao, Qingfeng Sun, Xiubo Geng, Wenxiang Hu, Chongyang Tao, Jing Ma, Qingwei Lin, Daxin Jiang
Categories: cs.CL, cs.AI
Published: 2023-06-14
arXiv: 2306.08568v2
Abstract:
Code Large Language Models (Code LLMs), such as StarCoder, have demonstrated exceptional performance in code-related tasks. However, most existing models are solely pre-trained on extensive raw code data without instruction fine-tuning. In this paper, we introduce WizardCoder, which empowers Code LLMs with complex instruction fine-tuning, by adapting the Evol-Instruct method to the domain of code. Through comprehensive experiments on four prominent code generation benchmarks, namely HumanEval, HumanEval+, MBPP, and DS-1000, we unveil the exceptional capabilities of our model. It surpasses all other open-source Code LLMs by a substantial margin. Moreover, our model even outperforms the largest closed LLMs, Anthropic’s Claude and Google’s Bard, on HumanEval and HumanEval+. Our code, model weights, and data are public at https://github.com/nlpxucan/WizardLM
38. The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions
Authors: Wenbo Pan, Zhichao Liu, Qiguang Chen, Xiangyang Zhou, Haining Yu, Xiaohua Jia
Categories: cs.CL, cs.AI
Published: 2025-02-13
arXiv: 2502.09674v4
Abstract:
Large Language Models’ safety-aligned behaviors, such as refusing harmful queries, can be represented by linear directions in activation space. Previous research modeled safety behavior with a single direction, limiting mechanistic understanding to an isolated safety feature. In this work, we discover that safety-aligned behavior is jointly controlled by multi-dimensional directions. Namely, we study the vector space of representation shifts during safety fine-tuning on Llama 3 8B for refusing jailbreaks. By studying orthogonal directions in the space, we first find that a dominant direction governs the model’s refusal behavior, while multiple smaller directions represent distinct and interpretable features like hypothetical narrative and role-playing. We then measure how different directions promote or suppress the dominant direction, showing the important role of secondary directions in shaping the model’s refusal representation. Finally, we demonstrate that removing certain trigger tokens in harmful queries can mitigate these directions to bypass the learned safety capability, providing new insights on understanding safety alignment vulnerability from a multi-dimensional perspective. Code and artifacts are available at https://github.com/BMPixel/safety-residual-space.
39. Do Models Share Safety Representations? Cross-Model Steering for Safe Visual Generation
Authors: Tobia Poppi, Silvia Cappelletti, Sara Sarto, Florian Schiffers, Garin Kessler, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara
Categories: cs.CV, cs.AI, cs.MM
Published: 2026-06-03
arXiv: 2606.05290v1
Abstract:
Recent progress in generative modeling has made safety control a central challenge, yet existing approaches remain largely model-specific, requiring retraining or tailored interventions for each new architecture. In this work, we ask whether safety can be represented as a portable latent direction, learned once and reused across heterogeneous generators. We introduce the first framework for cross-model safety steering, in which a safety direction is estimated in a source LLM from paired safe-unsafe prompts, transported to a target generator through a lightweight alignment fitted on benign data alone, and applied at inference time. Crucially, our pipeline never accesses unsafe data on the target side, isolating whether safety can be transferred through shared representation geometry. Beyond a single global direction, we also identify a multi-vector extension that captures category-specific safety behaviors, enabling more selective control. We evaluate our approach in text-to-image and text-to-video generation across diverse source-target model pairs. Across models, transferred safety directions achieve ASR reduction and CLIP-Score/FID trade-offs comparable to directions learned natively on the target model using unsafe data, while requiring no target-side unsafe data. This indicates that safety improvements do not come at the expense of generation quality. Our results point to a modular view of safety: safety-relevant behavior is not purely model-local, but can be controlled through latent directions that persist across models. This suggests a new path toward lightweight, reusable safety mechanisms that do not require target-side unsafe data.
40. Exploring Advanced Large Language Models with LLMsuite
Authors: Giorgio Roffo
Categories: cs.CL, cs.CV
Published: 2024-07-01
arXiv: 2407.12036v2
Abstract:
This tutorial explores the advancements and challenges in the development of Large Language Models (LLMs) such as ChatGPT and Gemini. It addresses inherent limitations like temporal knowledge cutoffs, mathematical inaccuracies, and the generation of incorrect information, proposing solutions like Retrieval Augmented Generation (RAG), Program-Aided Language Models (PAL), and frameworks such as ReAct and LangChain. The integration of these techniques enhances LLM performance and reliability, especially in multi-step reasoning and complex task execution. The paper also covers fine-tuning strategies, including instruction fine-tuning, parameter-efficient methods like LoRA, and Reinforcement Learning from Human Feedback (RLHF) as well as Reinforced Self-Training (ReST). Additionally, it provides a comprehensive survey of transformer architectures and training techniques for LLMs. The source code can be accessed by contacting the author via email for a request.
41. Can Linguistic Knowledge Improve Multimodal Alignment in Vision-Language Pretraining?
Authors: Fei Wang, Liang Ding, Jun Rao, Ye Liu, Li Shen, Changxing Ding
Categories: cs.MM, cs.AI, cs.CL, cs.CV
Published: 2023-08-24
arXiv: 2308.12898v2
Abstract:
The multimedia community has shown a significant interest in perceiving and representing the physical world with multimodal pretrained neural network models, and among them, the visual-language pertaining (VLP) is, currently, the most captivating topic. However, there have been few endeavors dedicated to the exploration of 1) whether essential linguistic knowledge (e.g., semantics and syntax) can be extracted during VLP, and 2) how such linguistic knowledge impact or enhance the multimodal alignment. In response, here we aim to elucidate the impact of comprehensive linguistic knowledge, including semantic expression and syntactic structure, on multimodal alignment. Specifically, we design and release the SNARE, the first large-scale multimodal alignment probing benchmark, to detect the vital linguistic components, e.g., lexical, semantic, and syntax knowledge, containing four tasks: Semantic structure, Negation logic, Attribute ownership, and Relationship composition. Based on our proposed probing benchmarks, our holistic analyses of five advanced VLP models illustrate that the VLP model: i) shows insensitivity towards complex syntax structures and relies on content words for sentence comprehension; ii) demonstrates limited comprehension of combinations between sentences and negations; iii) faces challenges in determining the presence of actions or spatial relationships within visual information and struggles with verifying the correctness of triple combinations. We make our benchmark and code available at \url{https://github.com/WangFei-2019/SNARE/}.
42. Aligning Language Models Using Follow-up Likelihood as Reward Signal
Authors: Chen Zhang, Dading Chong, Feng Jiang, Chengguang Tang, Anningzhe Gao, Guohua Tang, Haizhou Li
Categories: cs.CL
Published: 2024-09-20
arXiv: 2409.13948v3
Abstract:
In natural human-to-human conversations, participants often receive feedback signals from one another based on their follow-up reactions. These reactions can include verbal responses, facial expressions, changes in emotional state, and other non-verbal cues. Similarly, in human-machine interactions, the machine can leverage the user’s follow-up utterances as feedback signals to assess whether it has appropriately addressed the user’s request. Therefore, we propose using the likelihood of follow-up utterances as rewards to differentiate preferred responses from less favored ones, without relying on human or commercial LLM-based preference annotations. Our proposed reward mechanism, ``Follow-up Likelihood as Reward" (FLR), matches the performance of strong reward models trained on large-scale human or GPT-4 annotated data on 8 pairwise-preference and 4 rating-based benchmarks. Building upon the FLR mechanism, we propose to automatically mine preference data from the online generations of a base policy model. The preference data are subsequently used to boost the helpfulness of the base model through direct alignment from preference (DAP) methods, such as direct preference optimization (DPO). Lastly, we demonstrate that fine-tuning the language model that provides follow-up likelihood with natural language feedback significantly enhances FLR’s performance on reward modeling benchmarks and effectiveness in aligning the base policy model’s helpfulness.
43. Chain of Alignment: Integrating Public Will with Expert Intelligence for Language Model Alignment
Authors: Andrew Konya, Aviv Ovadya, Kevin Feng, Quan Ze Chen, Lisa Schirch, Colin Irwin, Amy X. Zhang
Categories: cs.HC, cs.AI, cs.CY
Published: 2024-11-15
arXiv: 2411.10534v1
Abstract:
We introduce a method to measure the alignment between public will and language model (LM) behavior that can be applied to fine-tuning, online oversight, and pre-release safety checks. Our `chain of alignment’ (CoA) approach produces a rule based reward (RBR) by creating model behavior $\textit{rules}$ aligned to normative $\textit{objectives}$ aligned to $\textit{public will}$. This factoring enables a nonexpert public to directly specify their will through the normative objectives, while expert intelligence is used to figure out rules entailing model behavior that best achieves those objectives. We validate our approach by applying it across three different domains of LM prompts related to mental health. We demonstrate a public input process built on collective dialogues and bridging-based ranking that reliably produces normative objectives supported by at least $96% \pm 2%$ of the US public. We then show that rules developed by mental health experts to achieve those objectives enable a RBR that evaluates an LM response’s alignment with the objectives similarly to human experts (Pearson’s $r=0.841$, $AUC=0.964$). By measuring alignment with objectives that have near unanimous public support, these CoA RBRs provide an approximate measure of alignment between LM behavior and public will.
44. Why Safeguarded Ships Run Aground? Aligned Large Language Models’ Safety Mechanisms Tend to Be Anchored in The Template Region
Authors: Chak Tou Leong, Qingyu Yin, Jian Wang, Wenjie Li
Categories: cs.CL, cs.AI, cs.CR
Published: 2025-02-19
arXiv: 2502.13946v2
Abstract:
The safety alignment of large language models (LLMs) remains vulnerable, as their initial behavior can be easily jailbroken by even relatively simple attacks. Since infilling a fixed template between the input instruction and initial model output is a common practice for existing LLMs, we hypothesize that this template is a key factor behind their vulnerabilities: LLMs’ safety-related decision-making overly relies on the aggregated information from the template region, which largely influences these models’ safety behavior. We refer to this issue as template-anchored safety alignment. In this paper, we conduct extensive experiments and verify that template-anchored safety alignment is widespread across various aligned LLMs. Our mechanistic analyses demonstrate how it leads to models’ susceptibility when encountering inference-time jailbreak attacks. Furthermore, we show that detaching safety mechanisms from the template region is promising in mitigating vulnerabilities to jailbreak attacks. We encourage future research to develop more robust safety alignment techniques that reduce reliance on the template region.
45. Structured Inference with Large Language Gibbs
Authors: Sanghyeok Choi, Henry Gouk, Esmeralda S. Whitammer
Categories: cs.LG, cs.CL
Published: 2026-06-17
arXiv: 2606.19264v1
Abstract:
The knowledge encoded in large language models (LLMs) can serve as a substrate for structured reasoning over variables describing a complex world, but accessing this knowledge in a probabilistically coherent manner poses a difficult inference problem. We propose Large Language Gibbs, a scheme for structured probabilistic inference that uses conditional distributions of an LLM as transition operators. Rather than sampling structured objects through single-pass autoregressive generation, we iteratively resample individual variables conditioned on others using an LLM’s next-token conditionals. This approach avoids order-dependent biases and produces a stationary distribution that reflects a compromise between all local conditionals. We apply this approach to sampling from synthetic distributions, consistent reasoning tasks, and Bayesian structure learning. The results suggest that the use of LLM conditionals in MCMC is a practical alternative to one-pass generation for structured probabilistic inference under a world prior accessible through noisy LLM conditionals.
46. One-Shot Safety Alignment for Large Language Models via Optimal Dualization
Authors: Xinmeng Huang, Shuo Li, Edgar Dobriban, Osbert Bastani, Hamed Hassani, Dongsheng Ding
Categories: cs.AI, cs.CL, cs.LG, math.OC, stat.ML
Published: 2024-05-29
arXiv: 2405.19544v3
Abstract:
The growing safety concerns surrounding large language models raise an urgent need to align them with diverse human preferences to simultaneously enhance their helpfulness and safety. A promising approach is to enforce safety constraints through Reinforcement Learning from Human Feedback (RLHF). For such constrained RLHF, typical Lagrangian-based primal-dual policy optimization methods are computationally expensive and often unstable. This paper presents a perspective of dualization that reduces constrained alignment to an equivalent unconstrained alignment problem. We do so by pre-optimizing a smooth and convex dual function that has a closed form. This shortcut eliminates the need for cumbersome primal-dual policy iterations, greatly reducing the computational burden and improving training stability. Our strategy leads to two practical algorithms in model-based and preference-based settings (MoCAN and PeCAN, respectively). A broad range of experiments demonstrate the effectiveness and merits of our algorithms.
47. Steering Multimodal Large Language Models Decoding for Context-Aware Safety
Authors: Zheyuan Liu, Zhangchen Xu, Guangyao Dou, Xiangchi Yuan, Zhaoxuan Tan, Radha Poovendran, Meng Jiang
Categories: cs.CL, cs.AI
Published: 2025-09-23
arXiv: 2509.19212v1
Abstract:
Multimodal Large Language Models (MLLMs) are increasingly deployed in real-world applications, yet their ability to make context-aware safety decisions remains limited. Existing methods often fail to balance oversensitivity (unjustified refusals of benign queries) and undersensitivity (missed detection of visually grounded risks), leaving a persistent gap in safety alignment. To address this issue, we introduce Safety-aware Contrastive Decoding (SafeCoDe), a lightweight and model-agnostic decoding framework that dynamically adjusts token generation based on multimodal context. SafeCoDe operates in two stages: (1) a contrastive decoding mechanism that highlights tokens sensitive to visual context by contrasting real and Gaussian-noised images, and (2) a global-aware token modulation strategy that integrates scene-level reasoning with token-level adjustment to adapt refusals according to the predicted safety verdict. Extensive experiments across diverse MLLM architectures and safety benchmarks, covering undersensitivity, oversensitivity, and general safety evaluations, show that SafeCoDe consistently improves context-sensitive refusal behaviors while preserving model helpfulness.
48. Large Language Models and Multimodal Retrieval for Visual Word Sense Disambiguation
Authors: Anastasia Kritharoula, Maria Lymperaiou, Giorgos Stamou
Categories: cs.CL
Published: 2023-10-21
arXiv: 2310.14025v1
Abstract:
Visual Word Sense Disambiguation (VWSD) is a novel challenging task with the goal of retrieving an image among a set of candidates, which better represents the meaning of an ambiguous word within a given context. In this paper, we make a substantial step towards unveiling this interesting task by applying a varying set of approaches. Since VWSD is primarily a text-image retrieval task, we explore the latest transformer-based methods for multimodal retrieval. Additionally, we utilize Large Language Models (LLMs) as knowledge bases to enhance the given phrases and resolve ambiguity related to the target word. We also study VWSD as a unimodal problem by converting to text-to-text and image-to-image retrieval, as well as question-answering (QA), to fully explore the capabilities of relevant models. To tap into the implicit knowledge of LLMs, we experiment with Chain-of-Thought (CoT) prompting to guide explainable answer generation. On top of all, we train a learn to rank (LTR) model in order to combine our different modules, achieving competitive ranking results. Extensive experiments on VWSD demonstrate valuable insights to effectively drive future directions.
49. Soft Inductive Bias Approach via Explicit Reasoning Perspectives in Inappropriate Utterance Detection Using Large Language Models
Authors: Ju-Young Kim, Ji-Hong Park, Se-Yeon Lee, Sujin Park, Gun-Woo Kim
Categories: cs.CL
Published: 2025-12-09
arXiv: 2512.08480v1
Abstract:
Recent incidents in certain online games and communities, where anonymity is guaranteed, show that unchecked inappropriate remarks frequently escalate into verbal abuse and even criminal behavior, raising significant social concerns. Consequently, there is a growing need for research on techniques that can detect inappropriate utterances within conversational texts to help build a safer communication environment. Although large-scale language models trained on Korean corpora and chain-of-thought reasoning have recently gained attention, research applying these approaches to inappropriate utterance detection remains limited. In this study, we propose a soft inductive bias approach that explicitly defines reasoning perspectives to guide the inference process, thereby promoting rational decision-making and preventing errors that may arise during reasoning. We fine-tune a Korean large language model using the proposed method and conduct both quantitative performance comparisons and qualitative evaluations across different training strategies. Experimental results show that the Kanana-1.5 model achieves an average accuracy of 87.0046, improving by approximately 3.89 percent over standard supervised learning. These findings indicate that the proposed method goes beyond simple knowledge imitation by large language models and enables more precise and consistent judgments through constrained reasoning perspectives, demonstrating its effectiveness for inappropriate utterance detection.
50. Foundational Challenges in Assuring Alignment and Safety of Large Language Models
Authors: Usman Anwar, Abulhair Saparov, Javier Rando, Daniel Paleka, Miles Turpin, Peter Hase, Ekdeep Singh Lubana, Erik Jenner, Stephen Casper, Oliver Sourbut, Benjamin L. Edelman, Zhaowei Zhang, Mario Günther, Anton Korinek, Jose Hernandez-Orallo, Lewis Hammond, Eric Bigelow, Alexander Pan, Lauro Langosco, Tomasz Korbak, Heidi Zhang, Ruiqi Zhong, Seán Ó hÉigeartaigh, Gabriel Recchia, Giulio Corsi, Alan Chan, Markus Anderljung, Lilian Edwards, Aleksandar Petrov, Christian Schroeder de Witt, Sumeet Ramesh Motwan, Yoshua Bengio, Danqi Chen, Philip H. S. Torr, Samuel Albanie, Tegan Maharaj, Jakob Foerster, Florian Tramer, He He, Atoosa Kasirzadeh, Yejin Choi, David Krueger
Categories: cs.LG, cs.AI, cs.CL, cs.CY
Published: 2024-04-15
arXiv: 2404.09932v2
Abstract:
This work identifies 18 foundational challenges in assuring the alignment and safety of large language models (LLMs). These challenges are organized into three different categories: scientific understanding of LLMs, development and deployment methods, and sociotechnical challenges. Based on the identified challenges, we pose $200+$ concrete research questions.