Large language models (LLMs) are evolving into autonomous decision-makers, raising concerns about catastrophic risks in high-stakes scenarios, particularly in Chemical, Biological, Radiological and Nuclear (CBRN) domains. Based on the insight that such risks can originate from trade-offs between the agent’s Helpful, Harmlessness and Honest (HHH) goals, we build a novel three-stage evaluation framework, which is carefully constructed to effectively and naturally expose such risks. We conduct 14,400 agentic simulations across 12 advanced LLMs, with extensive experiments and analysis. Results reveal that LLM agents can autonomously engage in catastrophic behaviors and deception, without being deliberately induced. Furthermore, stronger reasoning abilities often increase, rather than mitigate, these risks. We also show that these agents can violate instructions and superior commands. On the whole, we empirically prove the existence of catastrophic risks in autonomous LLM agents.
论文检索
输入标题、作者或关键词，从 3,086 篇学术成果中精准定位
Separate the Wheat from the Chaff: A Post-Hoc Approach to Safety Re-Alignment for Fine-Tuned Language Models
PDF ↗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 parameters. 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://github.com/pikepokenew/IRR.
Large language models (LLMs) exhibit excellent performance in natural language processing (NLP), but remain highly sensitive to the quality of input queries, especially when these queries contain misleading or inaccurate information. Existing methods focus on correcting the output, but they often overlook the potential of improving the ability of LLMs to detect and correct misleading content in the input itself. In this paper, we propose a novel three-stage fine-tuning method that enhances the ability of LLMs to detect and correct misleading information in the input, further improving response accuracy and reducing hallucinations. Specifically, the three stages include (1) training LLMs to identify misleading information, (2) training LLMs to correct the misleading information using built-in or external knowledge, and (3) training LLMs to generate accurate answers based on the corrected queries. To evaluate our method, we conducted experiments on three datasets for the hallucination detection task and the question answering (QA) task, as well as two datasets containing misleading information that we constructed. The experimental results demonstrate that our method significantly improves the accuracy and factuality of LLM responses, while also enhancing the ability to detect hallucinations and reducing the generation of hallucinations in the output, particularly when the query contains misleading information.
Counterfactual examples are widely used in natural language processing (NLP) as valuable data to improve models, and in explainable artificial intelligence (XAI) to understand model behavior. The automated generation of counterfactual examples remains a challenging task even for large language models (LLMs), despite their impressive performance on many tasks. In this paper, we first introduce ZeroCF, a faithful approach for leveraging important words derived from feature attribution methods to generate counterfactual examples in a zero-shot setting. Second, we present a new framework, FitCF, which further verifies aforementioned counterfactuals by label flip verification and then inserts them as demonstrations for few-shot prompting, outperforming three state-of-the-art baselines. Through ablation studies, we identify the importance of each of FitCF’s core components in improving the quality of counterfactuals, as assessed through flip rate, perplexity, and similarity measures. Furthermore, we show the effectiveness of LIME and Integrated Gradients as backbone attribution methods for FitCF and find that the number of demonstrations has the largest effect on performance. Finally, we reveal a strong correlation between the faithfulness of feature attribution scores and the quality of generated counterfactuals, which we hope will serve as an importantfinding for future research in this direction.
Vision-Language Models (VLMs) have made rapid progress in reasoning across visual and textual data. While VLMs perform well on vision tasks that they are trained on, our results highlight key challenges in abstract pattern recognition. We present GlyphPattern, a 954 item dataset that pairs 318 human-written descriptions of visual patterns from 40 writing systems with three visual presentation styles.GlyphPattern evaluates abstract pattern recognition in VLMs, requiring models to understand and judge natural language descriptions of visual patterns. GlyphPattern patterns are drawn from a large-scale cognitive science investigation of human writing systems; as a result, they are rich in spatial reference and compositionality. Our experiments show that GlyphPattern is challenging for state-of-the-art VLMs (GPT-4o achieves only 55% accuracy), with marginal gains from few-shot prompting. Our detailed analysis reveals errors at multiple levels, including visual processing, natural language understanding, and pattern generalization.
Large Language Models (LLMs) have demonstrated remarkable proficiency across a variety of complex tasks. One significant application of LLMs is in tackling software engineering challenges, particularly in resolving real-world tasks on GitHub by fixing code based on the issues reported by the users. However, many current approaches rely on proprietary LLMs, which limits reproducibility, accessibility, and transparency. The critical components of LLMs for addressing software engineering issues and how their capabilities can be effectively enhanced remain unclear. To address these challenges, we introduce SWE-Fixer, a novel open-source framework designed to effectively and efficiently resolve GitHub issues. SWE-Fixer comprises two essential modules: a code file retrieval module and a code editing module. The retrieval module employs BM25 along with a lightweight model to achieve coarse-to-fine file retrieval. Subsequently, the code editing module utilizes the other model to generate patches for the identified files. To mitigate the lack of publicly available datasets, we compile an extensive dataset that includes 110K GitHub issues along with their corresponding patches and train the two models of SWE-Fixer separately. We assess our approach on the SWE-Bench Lite and Verified benchmarks, achieving competitive performance among open-source models with scores of 22.0% and 30.2%. Furthermore, SWE-Fixer reaches state-of-the-art performance (24.7% on Lite and 32.8% on Verified) with PASS_TO_PASS (P2P) filtering. Additionally, our approach requires only two model calls per instance, making it significantly more efficient than existing methods. These results highlight the effectiveness of SWE-Fixer in real-world code-fixing scenarios.We will make our model, dataset, and code publicly available at https://github.com/InternLM/SWE-Fixer.
GenTool: Enhancing Tool Generalization in Language Models through Zero-to-One and Weak-to-Strong Simulation
PDF ↗Large Language Models (LLMs) can enhance their capabilities as AI assistants by integrating external tools, allowing them to access a wider range of information. While recent LLMs are typically fine-tuned with tool usage examples during supervised fine-tuning (SFT), questions remain about their ability to develop robust tool-usage skills and can effectively generalize to unseen queries and tools. In this work, we present GenTool, a novel training framework that prepares LLMs for diverse generalization challenges in tool utilization. Our approach addresses two fundamental dimensions critical for real-world applications: Zero-to-One Generalization, enabling the model to address queries initially lacking a suitable tool by adopting and utilizing one when it becomes available, and Weak-to-Strong Generalization, allowing models to leverage enhanced versions of existing tools to solve queries. To achieve this, we develop synthetic training data simulating these two dimensions of tool usage and introduce a two-stage fine-tuning approach: optimizing tool ranking, then refining tool selection. Through extensive experiments across four generalization scenarios, we demonstrate that our method significantly enhances the tool-usage capabilities of LLMs ranging from 1B to 8B parameters, achieving performance that surpasses GPT-4o. Furthermore, our analysis also provides valuable insights into the challenges LLMs encounter in tool generalization.
Do Robot Snakes Dream like Electric Sheep? Investigating the Effects of Architectural Inductive Biases on Hallucination
PDF ↗The growth in prominence of large language models (LLMs) in everyday life can be largely attributed to their generative abilities, yet some of this is also owed to the risks and costs associated with their use. On one front is their tendency to hallucinate false or misleading information, limiting their reliability. On another is the increasing focus on the computational limitations associated with traditional self-attention based LLMs, which has brought about new alternatives, in particular recurrent models, meant to overcome them. Yet it remains uncommon to consider these two concerns simultaneously. Do changes in architecture exacerbate/alleviate existing concerns about hallucinations? Do they affect how and where they occur? Through an extensive evaluation, we study how these architecture-based inductive biases affect the propensity to hallucinate. While hallucination remains a general phenomenon not limited to specific architectures, the situations in which they occur and the ease with which specific types of hallucinations can be induced can significantly differ based on the model architecture. These findings highlight the need for better understanding both these problems in conjunction with each other, as well as consider how to design more universal techniques for handling hallucinations.
Detecting the translation direction of parallel text has applications for machine translation training and evaluation, but also has forensic applications, such as resolving plagiarism or forgery allegations. In this work, we explore an unsupervised approach to translation direction detection based on the simple hypothesis that p(translation|original)>p(original|translation), motivated by the well-known simplification effect in translationese or machine-translationese. In experiments with multilingual machine translation models across 20 translation directions, we confirm the effectiveness of the approach for high-resource language pairs, achieving document-level accuracies of 82–96% for NMT-produced translations, and 60–81% for human translations, depending on the model used.
Current conversational recommendation systems focus predominantly on text. However, real-world recommendation settings are generally multimodal, causing a significant gap between existing research and practical applications. To address this issue, we propose Muse, the first multimodal conversational recommendation dataset. Muse comprises 83,148 utterances from 7,000 conversations centered around the Clothing domain. Each conversation contains comprehensive multimodal interactions, rich elements, and natural dialogues. Data in Muse are automatically synthesized by a multi-agent framework powered by multimodal large language models (MLLMs). It innovatively derives user profiles from real-world scenarios rather than depending on manual design and history data for better scalability, and then it fulfills conversation simulation and optimization. Both human and LLM evaluations demonstrate the high quality of conversations in Muse. Additionally, fine-tuning experiments on three MLLMs demonstrate Muse’s learnable patterns for recommendations and responses, confirming its value for multimodal conversational recommendation. Our dataset and codes are available at https://anonymous.4open.science/r/Muse-0086.
This paper empirically investigates the relationship between subword vocabulary size and the performance of large language models (LLMs) to provide insights on how to define the vocabulary size. Experimental results show that larger vocabulary sizes lead to better performance in LLMs. Moreover, we consider a continual training scenario where a pre-trained language model is trained on a different target language. We introduce a simple method to use a new vocabulary instead of the pre-defined one. We show that using the new vocabulary outperforms the model with the vocabulary used in pre-training.
Recent advancements in large audio language models (LALMs) have demonstrated impressive results and promising prospects in universal understanding and reasoning across speech, music, and general sound. However, these models still lack the ability to recognize their knowledge boundaries and refuse to answer questions they don’t know proactively. While there have been successful attempts to enhance the reliability of LLMs, reliable LALMs remain largely unexplored. In this paper, we systematically investigate various approaches towards reliable LALMs, including training-free methods such as multi-modal chain-of-thought (MCoT), and training-based methods such as supervised fine-tuning (SFT). Besides, we identify the limitations of previous evaluation metrics and propose a new metric, the Reliability Gain Index (RGI), to assess the effectiveness of different reliable methods. Our findings suggest that both training-free and training-based methods enhance the reliability of LALMs to different extents. Moreover, we find that awareness of reliability is a “meta ability”, which can be transferred across different audio modalities, although significant structural and content differences exist among sound, music, and speech.
Associative memory engages in the integration of relevant information for comprehension in the human cognition system. In this work, we seek to improve alignment between language models and human brain while processing speech information by integrating associative memory. After verifying the alignment between language model and brain by mapping language model activations to brain activity, the original text stimuli expanded with simulated associative memory are regarded as input to computational language models. We find the alignment between language model and brain is improved in brain regions closely related to associative memory processing. We also demonstrate large language models after specific supervised fine-tuning better align with brain response, by building the Association dataset containing 1000 samples of stories, with instructions encouraging associative memory as input and associated content as output.
Stance detection (SD) identifies a text’s position towards a target, typically labeled as favor, against, or none. We introduce Open-Target Stance Detection (OTSD), the most realistic task where targets are neither seen during training nor provided as input. We evaluate Large Language Models (LLMs) from GPT, Gemini, Llama, and Mistral families, comparing their performance to the only existing work, Target-Stance Extraction (TSE), which benefits from predefined targets. Unlike TSE, OTSD removes the dependency of a predefined list, making target generation and evaluation more challenging. We also provide a metric for evaluating target quality that correlates well with human judgment. Our experiments reveal that LLMs outperform TSE in target generation, both when the real target is explicitly and not explicitly mentioned in the text. Similarly, LLMs overall surpass TSE in stance detection for both explicit and non-explicit cases. However, LLMs struggle in both target generation and stance detection when the target is not explicit.
scRAG: Hybrid Retrieval-Augmented Generation for LLM-based Cross-Tissue Single-Cell Annotation
PDF ↗In recent years, large language models (LLMs) such as GPT-4 have demonstrated impressive potential in a wide range of fields, including biology, genomics and healthcare. Numerous studies have attempted to apply pre-trained LLMs to single-cell data analysis within one tissue. However, when it comes to cross-tissue cell annotation, LLMs often suffer from unsatisfactory performance due to the lack of specialized biological knowledge regarding genes and tissues. In this paper, we introduce scRAG, a novel framework that incorporates advanced LLM-based RAG techniques into cross-tissue single-cell annotation. scRAG utilizes LLMs to retrieve structured triples from knowledge graphs and unstructured similar cell information from the reference cell database, and it generates candidate cell types. The framework further optimizes predictions by retrieving marker genes from both candidate cells and similar cells to refine its results. Extensive experiments on a cross-tissue dataset demonstrate that our scRAG framework outperforms various baselines, including generalist models, domain-specific methods, and trained classifiers. The source code is available at https://github.com/YuZhiyin/scRAG.
We would like to express our gratitude to the Reviewers and the Area Chair for their insightful comments and for recognizing the robustness of our proposed framework for analyzing the serial position effects (SPE) in LLMs. We appreciate the acknowledgment of our work in demonstrating the widespread existence of this effect across various LLMs and the experiments we conducted to mitigate SPE.We acknowledge the concerns raised regarding the significance of the mitigation methods, including training-side solutions, CoT, and prompt engineering. The varying degrees of effectiveness observed in these methods highlight both the complexity and importance of addressing this cognitive bias. We believe these effects are inherently rooted in LLMs, and a comprehensive solution that fully addresses SPE may be beyond the scope of this work. However, we have proposed practical strategies, such as using binary choices instead of multiple choices where feasible, limiting prompt length, and placing crucial information at the beginning of prompts. These suggestions are intended to help users, particularly those who may not be experts in the domain of LLMs, to better utilize these models.We agree with the suggestion that a deeper analysis of the relationship between task characteristics and SPE could enhance the manuscript. As it stands, our findings indicate that higher model accuracy tends to correlate with a reduction in SPE, which aligns with expectations—if a model achieves 100% accuracy, it is unlikely to be influenced by SPE. Beyond this, we did not observe any clear relationships, which suggests that SPE may be influenced by a combination of factors, including the specific task, the model used, and the nature of the prompts. We will clarify this point in the final version of the manuscript.
ORBIT: Cost-Effective Dataset Curation for Large Language Model Domain Adaptation with an Astronomy Case Study
PDF ↗Recent advances in language modeling demonstrate the need for high-quality domain-specific training data, especially for tasks that require specialized knowledge. General-purpose models, while versatile, often lack the depth needed for expert-level tasks because of limited domain-specific information. Domain adaptation training can enhance these models, but it demands substantial, high-quality data. To address this, we propose ORBIT, a cost-efficient methodology for curating massive, high-quality domain-specific datasets from noisy web sources, tailored for training specialist large language models. Using astronomy as a primary case study, we refined the 1.3T-token FineWeb-Edu dataset into a high-quality, 10B-token subset focused on astronomy. Fine-tuning LLaMA-3-8B on a 1B-token astronomy subset improved performance on the MMLU astronomy benchmark from 69% to 76% and achieved top results on AstroBench, an astronomy-specific benchmark. Moreover, our model (Orbit-LLaMA) outperformed LLaMA-3-8B-base, with GPT-4o evaluations preferring it in 73% of cases across 1000 astronomy-specific questions. Additionally, we validated ORBIT’s generalizability by applying it to law and medicine, achieving a significant improvement of data quality compared to an unfiltered baseline. We open-source the ORBIT methodology, including the curated datasets, the codebase, and the resulting model.
Active Learning (AL) allows models to learn interactively from user feedback. However, only annotating existing samples may hardly benefit the model’s generalization. Moreover, AL commonly faces a cold start problem due to insufficient annotated data for effective sample selection. To address this, we introduce a counterfactual data augmentation approach inspired by Variation Theory, a theory of human concept learning that emphasizes the essential features of a concept by focusing on what stays the same and what changes. We use a neuro-symbolic pipeline to pinpoint key conceptual dimensions and use a large language model (LLM) to generate targeted variations along those dimensions. Through a text classification experiment, we show that our approach achieves significantly higher performance when there are fewer annotated data, showing its capability to address the cold start problem in AL. We also find that as the annotated training data gets larger, the impact of the generated data starts to diminish. This work demonstrates the value of incorporating human learning theories into the design and optimization of AL.
PersonaBench: Evaluating AI Models on Understanding Personal Information through Accessing (Synthetic) Private User Data
PDF ↗Personalization is essential for AI assistants, especially in private AI settings where models are expected to interpret users’ personal data (e.g., conversations, app usage) to understand their background, preferences, and social context. However, due to privacy concerns, existing academic research lacks direct access to such data, making benchmarking difficult. To fill this gap, we propose a synthetic data pipeline that generates realistic user profiles and private documents, enabling the creation of PersonaBench—a benchmark for evaluating models’ ability to understand personal information. Using this benchmark, we assess Retrieval-Augmented Generation (RAG) pipelines on personalized questions and find that current models struggle to accurately extract and answer questions even when provided with the full set of user documents, highlighting the need for improved personalization methods.
Understanding the Gap: an Analysis of Research Collaborations in NLP and Language Documentation
PDF ↗Despite over 20 years of NLP work explicitly intended for application in language documentation (LD), practical use of this work remains vanishingly scarce. This issue has been noted and discussed over the past 10 years, but without the benefit of data to inform the discourse.To address this lack in the literature, we present a survey- and interview-based analysis of the lack of adoption of NLP in LD, focusing on the matter of collaborations between documentary linguists and NLP researchers. Our data show support for ideas from previous work but also reveal the importance of little-discussed factors such as misaligned professional incentives, technical knowledge burdens, and LD software.