One of the major aspects contributing to the striking performance of large language models (LLMs) is the vast amount of factual knowledge accumulated during pre-training. Yet, many LLMs suffer from self-inconsistency, which raises doubts about their trustworthiness and reliability. This paper focuses on entity type ambiguity, analyzing the proficiency and consistency of state-of-the-art LLMs in applying factual knowledge when prompted with ambiguous entities. To do so, we propose an evaluation protocol that disentangles knowing from applying knowledge, and test state-of-the-art LLMs on 49 ambiguous entities. Our experiments reveal that LLMs struggle with choosing the correct entity reading, achieving an average accuracy of only 85%, and as low as 75% with underspecified prompts. The results also reveal systematic discrepancies in LLM behavior, showing that while the models may possess knowledge, they struggle to apply it consistently, exhibit biases toward preferred readings, and display self-inconsistencies. This highlights the need to address entity ambiguity in the future for more trustworthy LLMs.
论文检索
输入标题、作者或关键词，从 2,271 篇学术成果中精准定位
Large Language Models (LLMs) have emerged as the dominant paradigm in Natural Language Processing owing to their remarkable performance across various target tasks. However, naively fine-tuning them for specific downstream tasks often requires updating a vast number of parameters, resulting in high computational costs and overfitting when training data is limited. In this paper, we propose a novel approach, called *Stochastic Tuning*, that addresses these challenges by selectively updating a small subset of parameters in each step of the tuning process. Our approach is characterized by its customization of updates based on task-specific partial gradients with respect to stochastic sub-networks. The advantage of Stochastic Tuning over existing solutions lies in its ability to consider both parameter weights as well as forward values which guarantees a context-sensitive fine-tuning. Our experiments demonstrate that Stochastic Tuning outperforms existing lightweight fine-tuning methods, improving average performance by over two points on RoBERTa across several tasks in the GLUE benchmark while updating merely **0.08**% of the model’s parameters. The code for our implementation can be found at https://github.com/m-Tajari/StocTuning_LLMs.
There are many emerging quantization methods to resolve the problem that the huge demand on computational and storage costs hinders the deployment of Large language models (LLMs). However, their accuracy performance still can not satisfy the entire academic and industry community. In this work, we propose ATQ, an INT8 weight-activation quantization of LLMs, that can achieve almost lossless accuracy. We employ a mathematically equivalent transformation and a triangle inequality to constrain weight-activation quantization error to the sum of a weight quantization error and an activation quantization error. For the weight part, transformed weights are quantized along the |in-feature| dimension and the quantization error is compensated by optimizing following in-features. For the activation part, transformed activations are in the normal range and can be quantized easily. We provide comparison experiments to demonstrate that our ATQ method can achieve almost lossless in accuracy on OPT and LLaMA families in W8A8 quantization settings. The increase of perplexity is within 1 and the accuracy degradation is within 0.5 percent even in the worst case.
As open-weight large language models (LLMs) achieve ever more impressive performance across a wide range of tasks in English, practitioners aim to adapt these models to different languages. However, such language adaptation is often accompanied by catastrophic forgetting of the base model’s capabilities, severely limiting the usefulness of the resulting model. We address this issue by proposing Branch-and-Merge (BaM), a new adaptation method based on iteratively merging multiple models, fine-tuned on a subset of the available training data. BaM is based on the insight that this yields lower magnitude but higher quality weight changes, reducing forgetting of the source domain while maintaining learning on the target domain. We demonstrate in an extensive empirical study on Bulgarian and German that BaM can significantly reduce forgetting while matching or even improving target domain performance compared to both standard continued pretraining and instruction finetuning across different model architectures.
In Retrieval-Augmented Generation (RAG), retrieval is not always helpful and applying it to every instruction is sub-optimal. Therefore, determining whether to retrieve is crucial for RAG, which is usually referred to as Active Retrieval. However, existing active retrieval methods face two challenges: 1. They usually rely on a single criterion, which struggles with handling various types of instructions. 2. They depend on specialized and highly differentiated procedures, and thus combining them makes the RAG system more complicated and leads to higher response latency. To address these challenges, we propose Unified Active Retrieval (UAR). UAR contains four orthogonal criteria and casts them into plug-and-play classification tasks, which achieves multifaceted retrieval timing judgements with negligible extra inference cost. We further introduce the Unified Active Retrieval Criteria (UAR-Criteria), designed to process diverse active retrieval scenarios through a standardized procedure. Experiments on four representative types of user instructions show that UAR significantly outperforms existing work on the retrieval timing judgement and the performance of downstream tasks, which shows the effectiveness of UAR and its helpfulness to downstream tasks.
Dial BeInfo for Faithfulness: Improving Factuality of Information-Seeking Dialogue via Behavioural Fine-Tuning
PDF ↗Factual faithfulness is a crucial requirement in information-seeking dialogue: the system should respond to the user queries so that the responses are meaningful and aligned with the knowledge provided to the system. However, most modern large language models (LLMs) suffer from hallucinations, that is, they generate responses not supported by or even contradicting the knowledge source. To mitigate the issue and increase faithfulness of information-seeking dialogue systems supported by the LLMs, we introduce BeInfo, a simple yet effective method that applies ‘behavioural tuning’ on the LLMs to aid information-seeking dialogue. Relying on three standard information seeking dialogue datasets, we show that models tuned with BeInfo become considerably more faithful to the knowledge source both for datasets and domains seen during BeInfo-tuning, as well as on unseen domains, when applied in a zero-shot manner. In addition, we present a ‘real-life’ case study on conversations with real users, showcasing that the models with 3B parameters (e.g., Flan-T5) tuned with BeInfo demonstrate strong performance on data from real ‘production’ conversations: when tuned on a limited amount of such realistic in-domain dialogues, they surpass much larger LLMs used ‘off-the-shelf’, both on automatic and human evaluation metrics.
Error type information has been widely used to improve the performance of grammatical error correction (GEC) models, whether for generating corrections, re-ranking them, or combining GEC models. Combining GEC models that have complementary strengths in correcting different error types is very effective in producing better corrections. However, system combination incurs a high computational cost due to the need to run inference on the base systems before running the combination method itself. Therefore, it would be more efficient to have a single model with multiple sub-networks that specialize in correcting different error types. In this paper, we propose a mixture-of-experts model, MoECE, for grammatical error correction. Our model successfully achieves the performance of T5-XL with three times fewer effective parameters. Additionally, our model produces interpretable corrections by also identifying the error type during inference.
Driven by the surge in code generation using large language models (LLMs), numerous benchmarks have emerged to evaluate these LLMs capabilities. We conducted a large-scale human evaluation of *HumanEval* and *MBPP*, two popular benchmarks for Python code generation, analyzing their diversity and difficulty. Our findings unveil a critical bias towards a limited set of programming concepts, neglecting most of the other concepts entirely. Furthermore, we uncover a worrying prevalence of easy tasks that can inflate model performance estimations. To address these limitations, we propose a novel benchmark, *PythonSaga*, featuring 185 hand-crafted prompts in a balanced representation of 38 programming concepts across diverse difficulty levels. The robustness of our benchmark is demonstrated by the poor performance of existing Code-LLMs. The code and data set are openly available to the NLP community at this [URL](https://github.com/PythonSaga/PythonSaga).
Uncovering latent values and opinions embedded in large language models (LLMs) can help identify biases and mitigate potential harm. Recently, this has been approached by prompting LLMs with survey questions and quantifying the stances in the outputs towards morally and politically charged statements. However, the stances generated by LLMs can vary greatly depending on how they are prompted, and there are many ways to argue for or against a given position. In this work, we propose to address this by analysing a large and robust dataset of 156k LLM responses to the 62 propositions of the Political Compass Test (PCT) generated by 6 LLMs using 420 prompt variations. We perform coarse-grained analysis of their generated stances and fine-grained analysis of the plain text justifications for those stances. For fine-grained analysis, we propose to identify tropes in the responses: semantically similar phrases that are recurrent and consistent across different prompts, revealing natural patterns in the text that a given LLM is prone to produce. We find that demographic features added to prompts significantly affect outcomes on the PCT, reflecting bias, as well as disparities between the results of tests when eliciting closed-form vs. open domain responses. Additionally, patterns in the plain text rationales via tropes show that similar justifications are repeatedly generated across models and prompts even with disparate stances.
Low-rank adaptation (LoRA) and its mixture-of-experts (MOE) variants are highly effective parameter-efficient fine-tuning (PEFT) methods. However, they introduce significant latency in multi-tenant settings due to the LoRA modules and MOE routers added to multiple linear modules in the Transformer layer. To address this issue, we propose Mixture of Low-Rank Adaptation (MiLoRA), a novel and efficient LoRA variant. MiLoRA differs from previous MOE-style LoRA methods by considering each LoRA module as an expert and employing a prompt-aware routing mechanism. This mechanism calculates expert routing results once before generating the first new token and reuses these results for subsequent tokens, reducing latency. Extensive experiments and analysis on commonsense reasoning tasks, math reasoning tasks, and widely used LLM evaluation benchmarks demonstrate that MiLoRA consistently outperforms strong PEFT baselines with comparable tunable parameter budgets. Additionally, MiLoRA significantly reduces latency in multi-tenant settings compared to previous LoRA-based methods.
CROWD: Certified Robustness via Weight Distribution for Smoothed Classifiers against Backdoor Attack
PDF ↗Language models are vulnerable to clandestinely modified data and manipulation by attackers. Despite considerable research dedicated to enhancing robustness against adversarial attacks, the realm of provable robustness for backdoor attacks remains relatively unexplored. In this paper, we initiate a pioneering investigation into the certified robustness of NLP models against backdoor triggers.We propose a model-agnostic mechanism for large-scale models that applies to complex model structures without the need for assessing model architecture or internal knowledge. More importantly, we take recent advances in randomized smoothing theory and propose a novel weight-based distribution algorithm to enable semantic similarity and provide theoretical robustness guarantees.Experimentally, we demonstrate the efficacy of our approach across a diverse range of datasets and tasks, highlighting its utility in mitigating backdoor triggers. Our results show strong performance in terms of certified accuracy, scalability, and semantic preservation.
Growing amount and quality of AI-generated texts makes detecting such content more difficult. In most real-world scenarios, the domain (style and topic) of generated data and the generator model are not known in advance. In this work, we focus on the robustness of classifier-based detectors of AI-generated text, namely their ability to transfer to unseen generators or semantic domains. We investigate the geometry of the embedding space of Transformer-based text encoders and show that clearing out harmful linear subspaces helps to train a robust classifier, ignoring domain-specific spurious features. We investigate several subspace decomposition and feature selection strategies and achieve significant improvements over state of the art methods in cross-domain and cross-generator transfer. Our best approaches for head-wise and coordinate-based subspace removal increase the mean out-of-distribution (OOD) classification score by up to 9% and 14% in particular setups for RoBERTa and BERT embeddings respectively. We release our code and data: https://github.com/SilverSolver/RobustATD
The recent development of fact verification systems with natural logic has enhanced their explainability by aligning claims with evidence through set-theoretic operators, providing faithful justifications. Despite these advancements, such systems often rely on a large amount of training data annotated with natural logic. To address this issue, we propose a zero-shot method that utilizes the generalization capabilities of instruction-tuned large language models. To comprehensively assess the zero-shot capabilities of our method and other fact verification systems, we evaluate all models on both artificial and real-world claims, including multilingual datasets. We also compare our method against other fact verification systems in two setups. First, in the zero-shot generalization setup, we demonstrate that our approach outperforms other systems that were not specifically trained on natural logic data, achieving an average accuracy improvement of 8.96 points over the best-performing baseline. Second, in the zero-shot transfer setup, we show that current systems trained on natural logic data do not generalize well to other domains, and our method outperforms these systems across all datasets with real-world claims.
Improving Adversarial Robustness in Vision-Language Models with Architecture and Prompt Design
PDF ↗Vision-Language Models (VLMs) have seen a significant increase in both research interest and real-world applications across various domains, including healthcare, autonomous systems, and security. However, their growing prevalence demands higher reliability and safety including robustness to adversarial attacks. We systematically examine the possibility of incorporating adversarial robustness through various model design choices. We explore the effects of different vision encoders, the resolutions of vision encoders, and the size and type of language models. Additionally, we introduce novel, cost-effective approaches to enhance robustness through prompt engineering. By simply suggesting the possibility of adversarial perturbations or rephrasing questions, we demonstrate substantial improvements in model robustness against strong image-based attacks such as Auto-PGD. Our findings provide important guidelines for developing more robust VLMs, particularly for deployment in safety-critical environments where reliability and security are paramount. These insights are crucial for advancing the field of VLMs, ensuring they can be safely and effectively utilized in a wide range of applications.
In this paper, we introduce BiMediX, the first bilingual medical mixture of experts LLM designed for seamless interaction in both English and Arabic. Our model facilitates a wide range of medical interactions in English and Arabic, including multi-turn chats to inquire about additional details such as patient symptoms and medical history, multiple-choice question answering, and open-ended question answering. We propose a semi-automated English-to-Arabic translation pipeline with human refinement to ensure high-quality translations. We also introduce a comprehensive evaluation benchmark for Arabic medical LLMs. Furthermore, we introduce BiMed1.3M, an extensive Arabic-English bilingual instruction set that covers 1.3 Million diverse medical interactions, including 200k synthesized multi-turn doctor-patient chats, in a 1:2 Arabic-to-English ratio. Our model outperforms state-of-the-art Med42 and Meditron by average absolute gains of 2.5% and 4.1%, respectively, computed across multiple medical evaluation benchmarks in English, while operating at 8-times faster inference. Moreover, our BiMediX outperforms the generic Arabic-English bilingual LLM, Jais-30B, by average absolute gains of 10% on our Arabic and 15% on our bilingual evaluations across multiple datasets. Additionally, BiMediX exceeds the accuracy of GPT4 by 4.4% in open-ended question UPHILL evaluation and largely outperforms state-of-the-art open source medical LLMs in human evaluations of multi-turn conversations. Our trained models, instruction set, and source code are available at https://github.com/mbzuai-oryx/BiMediX.
When Compression Meets Model Compression: Memory-Efficient Double Compression for Large Language Models
PDF ↗Large language models (LLMs) exhibit excellent performance in various tasks. However, the memory requirements of LLMs present a great challenge when deploying on memory-limited devices, even for quantized LLMs. This paper introduces a framework to compress LLM after quantization further, achieving about 2.2x compression ratio. A compression-aware quantization is first proposed to enhance model weight compressibility by re-scaling the model parameters before quantization, followed by a pruning method to improve further. Upon this, we notice that decompression can be a bottleneck during practical scenarios. We then give a detailed analysis of the trade-off between memory usage and latency brought by the proposed method. A speed-adaptive method is proposed to overcome it. The experimental results show inference with the compressed model can achieve a 40% reduction in memory size with negligible loss in accuracy and inference speed.
The primary aim of Knowledge Graph Embeddings (KGE) is to learn low-dimensional representations of entities and relations for predicting missing facts. While rotation-based methods like RotatE and QuatE perform well in KGE, they face two challenges: limited model flexibility requiring proportional increases in relation size with entity dimension, and difficulties in generalizing the model for higher-dimensional rotations. To address these issues, we introduce OrthogonalE, a novel KGE model employing matrices for entities and block-diagonal orthogonal matrices with Riemannian optimization for relations. This approach not only enhances the generality and flexibility of KGE models but also captures several relation patterns that rotation-based methods can identify. Experimental results indicate that our new KGE model, OrthogonalE, offers generality and flexibility, captures several relation patterns, and significantly outperforms state-of-the-art KGE models while substantially reducing the number of relation parameters.
As an essential task in information extraction (IE), Event-Event Causal Relation Extraction (ECRE) aims to identify and classify the causal relationships between event mentions in natural language texts. However, existing research on ECRE has highlighted two critical challenges, including the lack of document-level modeling and causal hallucinations. In this paper, we propose a Knowledge-guided binary Question Answering (KnowQA) method with event structures for ECRE, consisting of two stages: Event Structure Construction and Binary Question Answering. We conduct extensive experiments under both zero-shot and fine-tuning settings with large language models (LLMs) on the MECI and MAVEN-ERE datasets. Experimental results demonstrate the usefulness of event structures on document-level ECRE and the effectiveness of KnowQA by achieving state-of-the-art on the MECI dataset. We observe not only the effectiveness but also the high generalizability and low inconsistency of our method, particularly when with complete event structures after fine-tuning the models.
Most existing Visual Question Answering (VQA) systems are constrained to support domain-specific questions, i.e., to train different models separately for different VQA tasks, thus generalizing poorly to others. For example, models trained on the reasoning-focused dataset GQA struggle to effectively handle samples from the knowledge-emphasizing dataset OKVQA. Meanwhile, in real-world scenarios, it is user-unfriendly to restrict the domain of questions. Therefore, this paper proposes a necessary task: One-to-Many Visual Question Answering, of which the ultimate goal is to enable a single model to answer as many different domains of questions as possible by the effective integration of available VQA resources. To this end, we first investigate into ten common VQA datasets, and break the task of VQA down into the integration of three key abilities.Then, considering assorted questions rely on different VQA abilities, this paper proposes a novel dynamic Mixture of LoRAs (MoL) strategy. MoL mixes three individually trained LoRA adapters (corresponding to each VQA ability) dynamically for different samples demanding various VQA abilities. The proposed MoL strategy is verified to be highly effective by experiments, establishing SOTAs on four datasets. In addition, MoL generalizes well to three extra zero-shot datasets.Data and codes will be released.
The evaluation paradigm of LLM-as-judge gains popularity due to its significant reduction in human labor and time costs. This approach utilizes one or more large language models (LLMs) to assess the quality of outputs from other LLMs. However, existing methods rely on generic scoring rubrics that fail to consider the specificities of each question and its problem-solving process, compromising precision and stability in assessments. Inspired by human examination scoring processes, we propose a new evaluation paradigm based on self-adaptive rubrics. Specifically, we create detailed scoring rubrics for each question, capturing the primary and secondary criteria in a structured format of scoring and deduction points that mimic a human evaluator’s analytical process. Building on this paradigm, we further develop a novel benchmark called SedarEval, which covers a range of domains including long-tail knowledge, mathematics, coding, and logical reasoning. SedarEval consists of 1,000 meticulously crafted questions, each with its own self-adaptive rubric. To further streamline the evaluation, we train a specialized evaluator language model (evaluator LM) to supplant human graders. Using the same training data, our evaluator LM achieves a higher concordance rate with human grading results than other paradigms, including GPT-4, highlighting the superiority and efficiency of our approach.