Representation Misdirection for Unlearning (RMU), which steers model representation in the intermediate layer to a target random representation, is an effective method for large language model (LLM) unlearning. Despite its high performance, the underlying cause and explanation remain underexplored. In this paper, we theoretically demonstrate that steering forget representations in the intermediate layer reduces token confidence, causing LLMs to generate wrong or nonsense responses. We investigate how the coefficient influences the alignment of forget-sample representations with the random direction and hint at the optimal coefficient values for effective unlearning across different network layers. We show that RMU unlearned models are robust against adversarial jailbreak attacks. Furthermore, our empirical analysis shows that RMU is less effective when applied to the middle and later layers in LLMs. To resolve this drawback, we propose Adaptive RMU -- a simple yet effective alternative method that makes unlearning effective with most layers. Extensive experiments demonstrate that Adaptive RMU significantly improves the unlearning performance compared to prior art while incurring no additional computational cost.
Episodic structures are inherently interpretable and adaptable to evolving large-scale key events. However, state-of-the-art automatic event detection methods overlook event episodes and, therefore, struggle with these crucial characteristics. This paper introduces a novel task, episode detection, aimed at identifying episodes from a news corpus containing key event articles. An episode describes a cohesive cluster of core entities (e.g., "protesters", "police") performing actions at a specific time and location. Furthermore, an episode is a significant part of a larger group of episodes under a particular key event. Automatically detecting episodes is challenging because, unlike key events and atomic actions, we cannot rely on explicit mentions of times and locations to distinguish between episodes or use semantic similarity to merge inconsistent episode co-references. To address these challenges, we introduce EpiMine, an unsupervised episode detection framework that (1) automatically identifies the most salient, key-event-relevant terms and segments, (2) determines candidate episodes in an article based on natural episodic partitions estimated through shifts in discriminative term combinations, and (3) refines and forms final episode clusters using large language model-based reasoning on the candidate episodes. We construct three diverse, real-world event datasets annotated at the episode level. EpiMine outperforms all baselines on these datasets by an average 59.2% increase across all metrics.
Large Language Models (LLMs) offer significant potential for clinical symptom extraction, but their deployment in healthcare settings is constrained by privacy concerns, computational limitations, and operational costs. This study investigates the optimization of compact LLMs for cancer toxicity symptom extraction using a novel iterative refinement approach. We employ a student-teacher architecture, utilizing Zephyr-7b-beta and Phi3-mini-128 as student models and GPT-4o as the teacher, to dynamically select between prompt refinement, Retrieval-Augmented Generation (RAG), and fine-tuning strategies. Our experiments on 294 clinical notes covering 12 post-radiotherapy toxicity symptoms demonstrate the effectiveness of this approach. The RAG method proved most efficient, improving average accuracy scores from 0.32 to 0.73 for Zephyr-7b-beta and from 0.40 to 0.87 for Phi3-mini-128 during refinement. In the test set, both models showed an approximate 0.20 increase in accuracy across symptoms. Notably, this improvement was achieved at a cost 45 times lower than GPT-4o for Zephyr and 79 times lower for Phi-3. These results highlight the potential of iterative refinement techniques in enhancing the capabilities of compact LLMs for clinical applications, offering a balance between performance, cost-effectiveness, and privacy preservation in healthcare settings.
This study utilizes the advanced capabilities of the GPT-4 multimodal Large Language Model (LLM) to explore its potential in iris recognition - a field less common and more specialized than face recognition. By focusing on this niche yet crucial area, we investigate how well AI tools like ChatGPT can understand and analyze iris images. Through a series of meticulously designed experiments employing a zero-shot learning approach, the capabilities of ChatGPT-4 was assessed across various challenging conditions including diverse datasets, presentation attacks, occlusions such as glasses, and other real-world variations. The findings convey ChatGPT-4's remarkable adaptability and precision, revealing its proficiency in identifying distinctive iris features, while also detecting subtle effects like makeup on iris recognition. A comparative analysis with Gemini Advanced - Google's AI model - highlighted ChatGPT-4's better performance and user experience in complex iris analysis tasks. This research not only validates the use of LLMs for specialized biometric applications but also emphasizes the importance of nuanced query framing and interaction design in extracting significant insights from biometric data. Our findings suggest a promising path for future research and the development of more adaptable, efficient, robust and interactive biometric security solutions.
This study examines the impact of AI on human false memories -- recollections of events that did not occur or deviate from actual occurrences. It explores false memory induction through suggestive questioning in Human-AI interactions, simulating crime witness interviews. Four conditions were tested: control, survey-based, pre-scripted chatbot, and generative chatbot using a large language model (LLM). Participants (N=200) watched a crime video, then interacted with their assigned AI interviewer or survey, answering questions including five misleading ones. False memories were assessed immediately and after one week. Results show the generative chatbot condition significantly increased false memory formation, inducing over 3 times more immediate false memories than the control and 1.7 times more than the survey method. 36.4% of users' responses to the generative chatbot were misled through the interaction. After one week, the number of false memories induced by generative chatbots remained constant. However, confidence in these false memories remained higher than the control after one week. Moderating factors were explored: users who were less familiar with chatbots but more familiar with AI technology, and more interested in crime investigations, were more susceptible to false memories. These findings highlight the potential risks of using advanced AI in sensitive contexts, like police interviews, emphasizing the need for ethical considerations.
Customers reach out to online live chat agents with various intents, such as asking about product details or requesting a return. In this paper, we propose the problem of predicting user intent from browsing history and address it through a two-stage approach. The first stage classifies a user's browsing history into high-level intent categories. Here, we represent each browsing history as a text sequence of page attributes and use the ground-truth class labels to fine-tune pretrained Transformers. The second stage provides a large language model (LLM) with the browsing history and predicted intent class to generate fine-grained intents. For automatic evaluation, we use a separate LLM to judge the similarity between generated and ground-truth intents, which closely aligns with human judgments. Our two-stage approach yields significant performance gains compared to generating intents without the classification stage.
LLM (large language model) practitioners commonly notice that outputs can vary for the same inputs under settings expected to be deterministic. Yet the questions of how pervasive this is, and with what impact on results, have not to our knowledge been systematically investigated. We investigate non-determinism in five LLMs configured to be deterministic when applied to eight common tasks in across 10 runs, in both zero-shot and few-shot settings. We see accuracy variations up to 15% across naturally occurring runs with a gap of best possible performance to worst possible performance up to 70%. In fact, none of the LLMs consistently delivers repeatable accuracy across all tasks, much less identical output strings. Sharing preliminary results with insiders has revealed that non-determinism perhaps essential to the efficient use of compute resources via co-mingled data in input buffers so this issue is not going away anytime soon. To better quantify our observations, we introduce metrics focused on quantifying determinism, TARr@N for the total agreement rate at N runs over raw output, and TARa@N for total agreement rate of parsed-out answers. Our code and data are publicly available at https://github.com/breckbaldwin/llm-stability.
Enabling Large Language Models (LLMs) to generate citations in Question-Answering (QA) tasks is an emerging paradigm aimed at enhancing the verifiability of their responses when LLMs are utilizing external references to generate an answer. However, there is currently no unified framework to standardize and fairly compare different citation generation methods, leading to difficulties in reproducing different methods and a comprehensive assessment. To cope with the problems above, we introduce \name, an open-source and modular toolkit designed to facilitate the implementation and evaluation of existing citation generation methods, while also fostering the development of new approaches to improve citation quality in LLM outputs. This tool is highly extensible, allowing users to utilize 4 main modules and 14 components to construct a pipeline, evaluating an existing method or innovative designs. Our experiments with two state-of-the-art LLMs and 11 citation generation baselines demonstrate varying strengths of different modules in answer accuracy and citation quality improvement, as well as the challenge of enhancing granularity. Based on our analysis of the effectiveness of components, we propose a new method, self-RAG \snippet, obtaining a balanced answer accuracy and citation quality. Citekit is released at https://github.com/SjJ1017/Citekit.
Mainframe operating systems, despite their inception in the 1940s, continue to support critical sectors like finance and government. However, these systems are often viewed as outdated, requiring extensive maintenance and modernization. Addressing this challenge necessitates innovative tools that can understand and interact with legacy codebases. To this end, we introduce XMainframe, a state-of-the-art large language model (LLM) specifically designed with knowledge of mainframe legacy systems and COBOL codebases. Our solution involves the creation of an extensive data collection pipeline to produce high-quality training datasets, enhancing XMainframe's performance in this specialized domain. Additionally, we present MainframeBench, a comprehensive benchmark for assessing mainframe knowledge, including multiple-choice questions, question answering, and COBOL code summarization. Our empirical evaluations demonstrate that XMainframe consistently outperforms existing state-of-the-art LLMs across these tasks. Specifically, XMainframe achieves 30% higher accuracy than DeepSeek-Coder on multiple-choice questions, doubles the BLEU score of Mixtral-Instruct 8x7B on question answering, and scores six times higher than GPT-3.5 on COBOL summarization. Our work highlights the potential of XMainframe to drive significant advancements in managing and modernizing legacy systems, thereby enhancing productivity and saving time for software developers.
The remarkable multimodal capabilities and interactive experience of GPT-4o underscore their necessity in practical applications, yet open-source models rarely excel in both areas. In this paper, we introduce VITA, the first-ever open-source Multimodal Large Language Model (MLLM) adept at simultaneous processing and analysis of Video, Image, Text, and Audio modalities, and meanwhile has an advanced multimodal interactive experience. Starting from Mixtral 8x7B as a language foundation, we expand its Chinese vocabulary followed by bilingual instruction tuning. We further endow the language model with visual and audio capabilities through two-stage multi-task learning of multimodal alignment and instruction tuning. VITA demonstrates robust foundational capabilities of multilingual, vision, and audio understanding, as evidenced by its strong performance across a range of both unimodal and multimodal benchmarks. Beyond foundational capabilities, we have made considerable progress in enhancing the natural multimodal human-computer interaction experience. VITA is the first step for the open-source community to explore the seamless integration of multimodal understanding and interaction. While there is still lots of work to be done on VITA to get close to close-source counterparts, we hope that its role as a pioneer can serve as a cornerstone for subsequent research. Project Page: https://vita-home.github.io.
Accurate survival prediction in radiotherapy (RT) is critical for optimizing treatment decisions. This study developed and validated the RT-Surv framework, which integrates general-domain, open-source large language models (LLMs) to structure unstructured electronic health records alongside structured clinical data. Using data from 34,276 patients and an external cohort of 852, the framework successfully transformed unstructured clinical information into structured formats. Incorporating LLM-structured clinical features improved the concordance index from 0.779 to 0.842 during external validation, demonstrating a significant performance enhancement. Key LLM-structured features, such as disease extent, general condition, and RT purpose, showed high predictive importance and aligned closely with statistically significant predictors identified through conventional statistical analyses, thereby improving model interpretability. Furthermore, the framework enhanced risk stratification, enabling more distinct differentiation among low-, intermediate-, and high-risk groups (p < 0.001) using LLM-structured clinical features. These findings highlight the potential of LLMs to convert unstructured data into actionable insights, improving predictive modeling and patient outcomes in clinics.
Visual Spatial Description (VSD) aims to generate texts that describe the spatial relationships between objects within images. Traditional visual spatial relationship classification (VSRC) methods typically output the spatial relationship between two objects in an image, often neglecting world knowledge and lacking general language capabilities. In this paper, we propose a Large Language-and-Vision Assistant for Visual Spatial Description, named LLaVA-VSD, which is designed for the classification, description, and open-ended description of visual spatial relationships. Specifically, the model first constructs a VSD instruction-following dataset using given figure-caption pairs for the three tasks. It then employs LoRA to fine-tune a Large Language and Vision Assistant for VSD, which has 13 billion parameters and supports high-resolution images. Finally, a large language model (Qwen-2) is used to refine the generated sentences, enhancing their diversity and accuracy. LLaVA-VSD demonstrates excellent multimodal conversational capabilities and can follow open-ended instructions to assist with inquiries about object relationships in images.
Retrieval augmented generation (RAG) is a process where a large language model (LLM) retrieves useful information from a database and then generates the responses. It is becoming popular in enterprise settings for daily business operations. For example, Copilot for Microsoft 365 has accumulated millions of businesses. However, the security implications of adopting such RAG-based systems are unclear. In this paper, we introduce ConfusedPilot, a class of security vulnerabilities of RAG systems that confuse Copilot and cause integrity and confidentiality violations in its responses. First, we investigate a vulnerability that embeds malicious text in the modified prompt in RAG, corrupting the responses generated by the LLM. Second, we demonstrate a vulnerability that leaks secret data, which leverages the caching mechanism during retrieval. Third, we investigate how both vulnerabilities can be exploited to propagate misinformation within the enterprise and ultimately impact its operations, such as sales and manufacturing. We also discuss the root cause of these attacks by investigating the architecture of a RAG-based system. This study highlights the security vulnerabilities in today's RAG-based systems and proposes design guidelines to secure future RAG-based systems.
Multi-modal Large Language Models (MLLMs) have demonstrated remarkable capabilities in executing instructions for a variety of single-image tasks. Despite this progress, significant challenges remain in modeling long image sequences. In this work, we introduce the versatile multi-modal large language model, mPLUG-Owl3, which enhances the capability for long image-sequence understanding in scenarios that incorporate retrieved image-text knowledge, interleaved image-text, and lengthy videos. Specifically, we propose novel hyper attention blocks to efficiently integrate vision and language into a common language-guided semantic space, thereby facilitating the processing of extended multi-image scenarios. Extensive experimental results suggest that mPLUG-Owl3 achieves state-of-the-art performance among models with a similar size on single-image, multi-image, and video benchmarks. Moreover, we propose a challenging long visual sequence evaluation named Distractor Resistance to assess the ability of models to maintain focus amidst distractions. Finally, with the proposed architecture, mPLUG-Owl3 demonstrates outstanding performance on ultra-long visual sequence inputs. We hope that mPLUG-Owl3 can contribute to the development of more efficient and powerful multimodal large language models.
Controlling the format of outputs generated by large language models (LLMs) is a critical functionality in various applications. Current methods typically employ constrained decoding with rule-based automata or fine-tuning with manually crafted format instructions, both of which struggle with open-domain format requirements. To address this limitation, we introduce a novel framework for controlled generation in LLMs, leveraging user-provided, one-shot QA pairs. This study investigates LLMs' capabilities to follow open-domain, one-shot constraints and replicate the format of the example answers. We observe that this is a non-trivial problem for current LLMs. We also develop a dataset collection methodology for supervised fine-tuning that enhances the open-domain format control of LLMs without degrading output quality, as well as a benchmark on which we evaluate both the helpfulness and format correctness of LLM outputs. The resulting datasets, named OIFC-SFT, along with the related code, will be made publicly available at https://github.com/cofe-ai/OIFC.
This study investigates how BERT processes and represents Argument Structure Constructions (ASCs), extending previous LSTM analyses. Using a dataset of 2000 sentences across four ASC types (transitive, ditransitive, caused-motion, resultative), we analyzed BERT's token embeddings across 12 layers. Visualizations with MDS and t-SNE and clustering quantified by Generalized Discrimination Value (GDV) were used. Feedforward classifiers (probes) predicted construction categories from embeddings. CLS token embeddings clustered best in layers 2-4, decreased in intermediate layers, and slightly increased in final layers. DET and SUBJ embeddings showed consistent clustering in intermediate layers, VERB embeddings increased in clustering from layer 1 to 12, and OBJ embeddings peaked in layer 10. Probe accuracies indicated low construction information in layer 1, with over 90 percent accuracy from layer 2 onward, revealing latent construction information beyond GDV clustering. Fisher Discriminant Ratio (FDR) analysis of attention weights showed OBJ tokens were crucial for differentiating ASCs, followed by VERB and DET tokens. SUBJ, CLS, and SEP tokens had insignificant FDR scores. This study highlights BERT's layered processing of linguistic constructions and its differences from LSTMs. Future research will compare these findings with neuroimaging data to understand the neural correlates of ASC processing. This research underscores neural language models' potential to mirror linguistic processing in the human brain, offering insights into the computational and neural mechanisms underlying language understanding.
We introduce a novel graph-based Retrieval-Augmented Generation (RAG) framework specifically designed for the medical domain, called \textbf{MedGraphRAG}, aimed at enhancing Large Language Model (LLM) capabilities for generating evidence-based medical responses, thereby improving safety and reliability when handling private medical data. Graph-based RAG (GraphRAG) leverages LLMs to organize RAG data into graphs, showing strong potential for gaining holistic insights from long-form documents. However, its standard implementation is overly complex for general use and lacks the ability to generate evidence-based responses, limiting its effectiveness in the medical field. To extend the capabilities of GraphRAG to the medical domain, we propose unique Triple Graph Construction and U-Retrieval techniques over it. In our graph construction, we create a triple-linked structure that connects user documents to credible medical sources and controlled vocabularies. In the retrieval process, we propose U-Retrieval which combines Top-down Precise Retrieval with Bottom-up Response Refinement to balance global context awareness with precise indexing. These effort enable both source information retrieval and comprehensive response generation. Our approach is validated on 9 medical Q\&A benchmarks, 2 health fact-checking benchmarks, and one collected dataset testing long-form generation. The results show that MedGraphRAG consistently outperforms state-of-the-art models across all benchmarks, while also ensuring that responses include credible source documentation and definitions. Our code is released at: https://github.com/MedicineToken/Medical-Graph-RAG.
We present a novel prompting strategy for artificial intelligence driven digital avatars. To better quantify how our prompting strategy affects anthropomorphic features like humor, authenticity, and favorability we present Crowd Vote - an adaptation of Crowd Score that allows for judges to elect a large language model (LLM) candidate over competitors answering the same or similar prompts. To visualize the responses of our LLM, and the effectiveness of our prompting strategy we propose an end-to-end framework for creating high-fidelity artificial intelligence (AI) driven digital avatars. This pipeline effectively captures an individual's essence for interaction and our streaming algorithm delivers a high-quality digital avatar with real-time audio-video streaming from server to mobile device. Both our visualization tool, and our Crowd Vote metrics demonstrate our AI driven digital avatars have state-of-the-art humor, authenticity, and favorability outperforming all competitors and baselines. In the case of our Donald Trump and Joe Biden avatars, their authenticity and favorability are rated higher than even their real-world equivalents.
As Large Language Models (LLMs) become increasingly sophisticated and ubiquitous in natural language processing (NLP) applications, ensuring their robustness, trustworthiness, and alignment with human values has become a critical challenge. This paper presents a novel framework for contextual grounding in textual models, with a particular emphasis on the Context Representation stage. Our approach aims to enhance the reliability and ethical alignment of these models through a comprehensive, context-aware methodology. By explicitly capturing and representing relevant situational, cultural, and ethical contexts in a machine-readable format, we lay the foundation for anchoring a model's behavior within these contexts. Our approach leverages techniques from knowledge representation and reasoning, such as ontologies, semantic web technologies, and logic-based formalisms. We evaluate our framework on real-world textual datasets, demonstrating its effectiveness in improving model performance, fairness, and alignment with human expectations, while maintaining high accuracy. Furthermore, we discuss the other key components of the framework, including context-aware encoding, context-aware learning, interpretability and explainability, and continuous monitoring and adaptation. This research contributes to the growing body of work on responsible AI, offering a practical approach to developing more reliable, trustworthy, and ethically-aligned language models. Our findings have significant implications for the deployment of LLMs in sensitive domains such as healthcare, legal systems, and social services, where contextual understanding is paramount.
Power dispatch is essential for providing stable, cost-effective, and eco-friendly electricity to society. However, traditional methods falter as power systems grow in scale and complexity, struggling with multitasking, swift problem-solving, and human-machine collaboration. This paper introduces GAIA, the pioneering Large Language Model (LLM) tailored for power dispatch tasks. We have developed a novel dataset construction technique that harnesses a range of data sources to fine-tune GAIA for optimal performance in this domain. This approach streamlines LLM training, allowing for the seamless integration of multidimensional data in power system management. Additionally, we have crafted specialized prompt strategies to boost GAIA's input-output efficiency in dispatch scenarios. When evaluated on the ElecBench benchmark, GAIA surpasses the baseline model LLaMA2 on multiple metrics. In practical applications, GAIA has demonstrated its ability to enhance decision-making processes, improve operational efficiency, and facilitate better human-machine interactions in power dispatch operations. This paper expands the application of LLMs to power dispatch and validates their practical utility, paving the way for future innovations in this field.