Recent work using Rank-One Model Editing (ROME), a popular model editing method, has shown that there are certain facts that the algorithm is unable to edit without breaking the model. Such edits have previously been called disabling edits. These disabling edits cause immediate model collapse and limits the use of ROME for sequential editing. In this paper, we show that disabling edits are an artifact of irregularities in the implementation of ROME. With this paper, we provide a more stable implementation ROME, which we call r-ROME and show that model collapse is no longer observed when making large scale sequential edits with r-ROME, while further improving generalization and locality of model editing compared to the original implementation of ROME. We also provide a detailed mathematical explanation of the reason behind disabling edits.

Recognizing if LLM output can be grounded in evidence is central to many tasks in NLP: retrieval-augmented generation, summarization, document-grounded dialogue, and more. Current approaches to this kind of "fact-checking" are based on verifying each piece of a model generation against potential evidence using an LLM. However, this process can be very computationally expensive, requiring many calls to LLMs to check a single response. In this work, we show how to build small models that have GPT-4-level performance but for 400x lower cost. We do this by constructing synthetic training data with GPT-4, which involves creating realistic yet challenging instances of factual errors via a structured generation procedure. Training on this data teaches models to check each fact in the claim and recognize synthesis of information across sentences. For evaluation, we unify pre-existing datasets into a benchmark LLM-AggreFact, collected from recent work on fact-checking and grounding LLM generations. Our best system MiniCheck-FT5 (770M parameters) outperforms all systems of comparable size and reaches GPT-4 accuracy. We release LLM-AggreFact, code for data synthesis, and models.

State-of-the-art performance in QA tasks is currently achieved by systems employing Large Language Models (LLMs), however these models tend to hallucinate information in their responses. One approach focuses on enhancing the generation process by incorporating attribution from the given input to the output. However, the challenge of identifying appropriate attributions and verifying their accuracy against a source is a complex task that requires significant improvements in assessing such systems. We introduce an attribution-oriented Chain-of-Thought reasoning method to enhance the accuracy of attributions. This approach focuses the reasoning process on generating an attribution-centric output. Evaluations on two context-enhanced question-answering datasets using GPT-4 demonstrate improved accuracy and correctness of attributions. In addition, the combination of our method with finetuning enhances the response and attribution accuracy of two smaller LLMs, showing their potential to outperform GPT-4 in some cases.

Learning from human feedback plays an important role in aligning generative models, such as large language models (LLM). However, the effectiveness of this approach can be influenced by adversaries, who may intentionally provide misleading preferences to manipulate the output in an undesirable or harmful direction. To tackle this challenge, we study a specific model within this problem domain--contextual dueling bandits with adversarial feedback, where the true preference label can be flipped by an adversary. We propose an algorithm namely robust contextual dueling bandit (\algo), which is based on uncertainty-weighted maximum likelihood estimation. Our algorithm achieves an $\tilde O(d\sqrt{T}+dC)$ regret bound, where $T$ is the number of rounds, $d$ is the dimension of the context, and $ 0 \le C \le T$ is the total number of adversarial feedback. We also prove a lower bound to show that our regret bound is nearly optimal, both in scenarios with and without ($C=0$) adversarial feedback. Additionally, we conduct experiments to evaluate our proposed algorithm against various types of adversarial feedback. Experimental results demonstrate its superiority over the state-of-the-art dueling bandit algorithms in the presence of adversarial feedback.

Relation triple extraction, which outputs a set of triples from long sentences, plays a vital role in knowledge acquisition. Large language models can accurately extract triples from simple sentences through few-shot learning or fine-tuning when given appropriate instructions. However, they often miss out when extracting from complex sentences. In this paper, we design an evaluation-filtering framework that integrates large language models with small models for relational triple extraction tasks. The framework includes an evaluation model that can extract related entity pairs with high precision. We propose a simple labeling principle and a deep neural network to build the model, embedding the outputs as prompts into the extraction process of the large model. We conduct extensive experiments to demonstrate that the proposed method can assist large language models in obtaining more accurate extraction results, especially from complex sentences containing multiple relational triples. Our evaluation model can also be embedded into traditional extraction models to enhance their extraction precision from complex sentences.

In this study, we introduce SpamDam, a SMS spam detection framework designed to overcome key challenges in detecting and understanding SMS spam, such as the lack of public SMS spam datasets, increasing privacy concerns of collecting SMS data, and the need for adversary-resistant detection models. SpamDam comprises four innovative modules: an SMS spam radar that identifies spam messages from online social networks(OSNs); an SMS spam inspector for statistical analysis; SMS spam detectors(SSDs) that enable both central training and federated learning; and an SSD analyzer that evaluates model resistance against adversaries in realistic scenarios. Leveraging SpamDam, we have compiled over 76K SMS spam messages from Twitter and Weibo between 2018 and 2023, forming the largest dataset of its kind. This dataset has enabled new insights into recent spam campaigns and the training of high-performing binary and multi-label classifiers for spam detection. Furthermore, effectiveness of federated learning has been well demonstrated to enable privacy-preserving SMS spam detection. Additionally, we have rigorously tested the adversarial robustness of SMS spam detection models, introducing the novel reverse backdoor attack, which has shown effectiveness and stealthiness in practical tests.

It is assumed that readers are already familiar with Language Models of various flavors such as: Transformer-based Language Models (currently the most promising and studied LMs) [78]; Multimodal Foundation Models such as Blip-2 [48] and CLIP [61]; Auto-regressive Language Models [15, 51]; Recurrent Neural Network Language Models [75]; State space language models [40]; Hybrid Models [24] as well as the current and proposed use cases and the various technologies underlying them [1, 65, 70]. There is growing interest in LM performance and benchmarks [13, 16, 18,46, 47, 64,72, 74, 80] with recent acknowledgement that this is a hard problem [53]. Many suggestions are proposed in the commercial literature [17] and a large number of benchmark-based methods have surfaced (Big Bench [67], GLUE Benchmark, SuperGLUE Benchmark, OpenAI Moderation API, MMLU, EleutherAI LM Eval, OpenAI Evals Adversarial NLI, LIT, ParlAI, CoQA, LAMBADA, HellaSwag, LogiQA, MultiNLI, SQUAD to name a few). A review of existing approaches demonstrates that they are not quantitative or rigorous enough to past muster with respect to accepted testing requirements [3, 55]. In particular, existing use of benchmarks do not investigate the extent to which a benchmark can predict or quantify certain properties on future prompts (that is, statistical soundness of any conclusions) and do not identify factors affecting performance dependence as would be possible with more rigorous experimental design and test execution. LMs can be black box, gray box or white box according to the visibility into the architecture and training data used to create an LM (see Table 1). Remote Black Box LMs typically throttle the number of prompts so sustained access for testing could be difficult unless priority access to an API is given. For example, ChatGPT limits users to a small number of free prompts but allows unlimited prompts on its subscription option. Additionally, reproducability may not be guaranteed because of randomness in the response generation and/or continuous adaptation of the LM platform.

State-of-the-art large language models (LLMs) have become indispensable tools for various tasks. However, training LLMs to serve as effective assistants for humans requires careful consideration. A promising approach is reinforcement learning from human feedback (RLHF), which leverages human feedback to update the model in accordance with human preferences and mitigate issues like toxicity and hallucinations. Yet, an understanding of RLHF for LLMs is largely entangled with initial design choices that popularized the method and current research focuses on augmenting those choices rather than fundamentally improving the framework. In this paper, we analyze RLHF through the lens of reinforcement learning principles to develop an understanding of its fundamentals, dedicating substantial focus to the core component of RLHF -- the reward model. Our study investigates modeling choices, caveats of function approximation, and their implications on RLHF training algorithms, highlighting the underlying assumptions made about the expressivity of reward. Our analysis improves the understanding of the role of reward models and methods for their training, concurrently revealing limitations of the current methodology. We characterize these limitations, including incorrect generalization, model misspecification, and the sparsity of feedback, along with their impact on the performance of a language model. The discussion and analysis are substantiated by a categorical review of current literature, serving as a reference for researchers and practitioners to understand the challenges of RLHF and build upon existing efforts.

Smart cities need the involvement of their residents to enhance quality of life. Conversational query-answering is an emerging approach for user engagement. There is an increasing demand of an advanced conversational question-answering that goes beyond classic systems. Existing approaches have shown that LLMs offer promising capabilities for CQA, but may struggle to capture the nuances of conversational contexts. The new approach involves understanding the content and engaging in a multi-step conversation with the user to fulfill their needs. This paper presents a novel method to elevate the performance of Persian Conversational question-answering (CQA) systems. It combines the strengths of Large Language Models (LLMs) with contextual keyword extraction. Our method extracts keywords specific to the conversational flow, providing the LLM with additional context to understand the user's intent and generate more relevant and coherent responses. We evaluated the effectiveness of this combined approach through various metrics, demonstrating significant improvements in CQA performance compared to an LLM-only baseline. The proposed method effectively handles implicit questions, delivers contextually relevant answers, and tackles complex questions that rely heavily on conversational context. The findings indicate that our method outperformed the evaluation benchmarks up to 8% higher than existing methods and the LLM-only baseline.

Recent progress with LLM-based agents has shown promising results across various tasks. However, their use in answering questions from knowledge bases remains largely unexplored. Implementing a KBQA system using traditional methods is challenging due to the shortage of task-specific training data and the complexity of creating task-focused model structures. In this paper, we present Triad, a unified framework that utilizes an LLM-based agent with three roles for KBQA tasks. The agent is assigned three roles to tackle different KBQA subtasks: agent as a generalist for mastering various subtasks, as a decision maker for the selection of candidates, and as an advisor for answering questions with knowledge. Our KBQA framework is executed in four phases, involving the collaboration of the agent's multiple roles. We evaluated the performance of our framework using three benchmark datasets, and the results show that our framework outperforms state-of-the-art systems on the LC-QuAD and YAGO-QA benchmarks, yielding F1 scores of 11.8% and 20.7%, respectively.