Proprietary LMs such as GPT-4 are often employed to assess the quality of responses from various LMs. However, concerns including transparency, controllability, and affordability strongly motivate the development of open-source LMs specialized in evaluations. On the other hand, existing open evaluator LMs exhibit critical shortcomings: 1) they issue scores that significantly diverge from those assigned by humans, and 2) they lack the flexibility to perform both direct assessment and pairwise ranking, the two most prevalent forms of assessment. Additionally, they do not possess the ability to evaluate based on custom evaluation criteria, focusing instead on general attributes like helpfulness and harmlessness. To address these issues, we introduce Prometheus 2, a more powerful evaluator LM than its predecessor that closely mirrors human and GPT-4 judgements. Moreover, it is capable of processing both direct assessment and pair-wise ranking formats grouped with a user-defined evaluation criteria. On four direct assessment benchmarks and four pairwise ranking benchmarks, Prometheus 2 scores the highest correlation and agreement with humans and proprietary LM judges among all tested open evaluator LMs. Our models, code, and data are all publicly available at https://github.com/prometheus-eval/prometheus-eval.
Owing to their powerful semantic reasoning capabilities, Large Language Models (LLMs) have been effectively utilized as recommenders, achieving impressive performance. However, the high inference latency of LLMs significantly restricts their practical deployment. To address this issue, this work investigates knowledge distillation from cumbersome LLM-based recommendation models to lightweight conventional sequential models. It encounters three challenges: 1) the teacher's knowledge may not always be reliable; 2) the capacity gap between the teacher and student makes it difficult for the student to assimilate the teacher's knowledge; 3) divergence in semantic space poses a challenge to distill the knowledge from embeddings. To tackle these challenges, this work proposes a novel distillation strategy, DLLM2Rec, specifically tailored for knowledge distillation from LLM-based recommendation models to conventional sequential models. DLLM2Rec comprises: 1) Importance-aware ranking distillation, which filters reliable and student-friendly knowledge by weighting instances according to teacher confidence and student-teacher consistency; 2) Collaborative embedding distillation integrates knowledge from teacher embeddings with collaborative signals mined from the data. Extensive experiments demonstrate the effectiveness of the proposed DLLM2Rec, boosting three typical sequential models with an average improvement of 47.97%, even enabling them to surpass LLM-based recommenders in some cases.
Large Language Models (LLMs) have swiftly emerged as vital resources for different applications in the biomedical and healthcare domains; however, these models encounter issues such as generating inaccurate information or hallucinations. Retrieval-augmented generation provided a solution for these models to update knowledge and enhance their performance. In contrast to previous retrieval-augmented LMs, which utilize specialized cross-attention mechanisms to help LLM encode retrieved text, BiomedRAG adopts a simpler approach by directly inputting the retrieved chunk-based documents into the LLM. This straightforward design is easily applicable to existing retrieval and language models, effectively bypassing noise information in retrieved documents, particularly in noise-intensive tasks. Moreover, we demonstrate the potential for utilizing the LLM to supervise the retrieval model in the biomedical domain, enabling it to retrieve the document that assists the LM in improving its predictions. Our experiments reveal that with the tuned scorer,\textsc{ BiomedRAG} attains superior performance across 5 biomedical NLP tasks, encompassing information extraction (triple extraction, relation extraction), text classification, link prediction, and question-answering, leveraging over 9 datasets. For instance, in the triple extraction task, \textsc{BiomedRAG} outperforms other triple extraction systems with micro-F1 scores of 81.42 and 88.83 on GIT and ChemProt corpora, respectively.
Social media abounds with multimodal sarcasm, and identifying sarcasm targets is particularly challenging due to the implicit incongruity not directly evident in the text and image modalities. Current methods for Multimodal Sarcasm Target Identification (MSTI) predominantly focus on superficial indicators in an end-to-end manner, overlooking the nuanced understanding of multimodal sarcasm conveyed through both the text and image. This paper proposes a versatile MSTI framework with a coarse-to-fine paradigm, by augmenting sarcasm explainability with reasoning and pre-training knowledge. Inspired by the powerful capacity of Large Multimodal Models (LMMs) on multimodal reasoning, we first engage LMMs to generate competing rationales for coarser-grained pre-training of a small language model on multimodal sarcasm detection. We then propose fine-tuning the model for finer-grained sarcasm target identification. Our framework is thus empowered to adeptly unveil the intricate targets within multimodal sarcasm and mitigate the negative impact posed by potential noise inherently in LMMs. Experimental results demonstrate that our model far outperforms state-of-the-art MSTI methods, and markedly exhibits explainability in deciphering sarcasm as well.
Large language models (LLMs) can generate long-form and coherent text, yet they often hallucinate facts, which undermines their reliability. To mitigate this issue, inference-time methods steer LLM representations toward the "truthful directions" previously learned for truth elicitation. However, applying these truthful directions with the same intensity fails to generalize across different query contexts. We propose LITO, a Learnable Intervention method for Truthfulness Optimization that automatically identifies the optimal intervention intensity tailored to each specific context. LITO explores a sequence of model generations based on increasing levels of intervention intensities. It selects the most accurate response or refuses to answer when the predictions are highly uncertain. Experiments on multiple LLMs and question-answering datasets demonstrate that LITO improves truthfulness while preserving task accuracy. The adaptive nature of LITO counters the limitations of one-size-fits-all intervention methods, maximizing truthfulness by reflecting the model's internal knowledge only when it is confident. Our code is available at https://github.com/launchnlp/LITO.
Learning never ends, and there is no age limit to grow yourself. However, the educational landscape may face challenges in effectively catering to students' inclusion and diverse learning needs. These students should have access to state-of-the-art methods for lecture delivery, online resources, and technology needs. However, with all the diverse learning sources, it becomes harder for students to comprehend a large amount of knowledge in a short period of time. Traditional assistive technologies and learning aids often lack the dynamic adaptability required for individualized education plans. Large Language Models (LLM) have been used in language translation, text summarization, and content generation applications. With rapid growth in AI over the past years, AI-powered chatbots and virtual assistants have been developed. This research aims to bridge this gap by introducing an innovative study buddy we will be calling the 'SAMCares'. The system leverages a Large Language Model (LLM) (in our case, LLaMa-2 70B as the base model) and Retriever-Augmented Generation (RAG) to offer real-time, context-aware, and adaptive educational support. The context of the model will be limited to the knowledge base of Sam Houston State University (SHSU) course notes. The LLM component enables a chat-like environment to interact with it to meet the unique learning requirements of each student. For this, we will build a custom web-based GUI. At the same time, RAG enhances real-time information retrieval and text generation, in turn providing more accurate and context-specific assistance. An option to upload additional study materials in the web GUI is added in case additional knowledge support is required. The system's efficacy will be evaluated through controlled trials and iterative feedback mechanisms.
Graphical User Interfaces (GUIs) are central to app development projects. App developers may use the GUIs of other apps as a means of requirements refinement and rapid prototyping or as a source of inspiration for designing and improving their own apps. Recent research has thus suggested retrieving relevant GUI designs that match a certain text query from screenshot datasets acquired through crowdsourced or automated exploration of GUIs. However, such text-to-GUI retrieval approaches only leverage the textual information of the GUI elements, neglecting visual information such as icons or background images. In addition, retrieved screenshots are not steered by app developers and lack app features that require particular input data. To overcome these limitations, this paper proposes GUing, a GUI search engine based on a vision-language model called GUIClip, which we trained specifically for the problem of designing app GUIs. For this, we first collected from Google Play app introduction images which display the most representative screenshots and are often captioned (i.e.~labelled) by app vendors. Then, we developed an automated pipeline to classify, crop, and extract the captions from these images. This resulted in a large dataset which we share with this paper: including 303k app screenshots, out of which 135k have captions. We used this dataset to train a novel vision-language model, which is, to the best of our knowledge, the first of its kind for GUI retrieval. We evaluated our approach on various datasets from related work and in a manual experiment. The results demonstrate that our model outperforms previous approaches in text-to-GUI retrieval achieving a Recall@10 of up to 0.69 and a HIT@10 of 0.91. We also explored the performance of GUIClip for other GUI tasks including GUI classification and sketch-to-GUI retrieval with encouraging results.
Standard reinforcement learning from human feedback (RLHF) approaches relying on parametric models like the Bradley-Terry model fall short in capturing the intransitivity and irrationality in human preferences. Recent advancements suggest that directly working with preference probabilities can yield a more accurate reflection of human preferences, enabling more flexible and accurate language model alignment. In this paper, we propose a self-play-based method for language model alignment, which treats the problem as a constant-sum two-player game aimed at identifying the Nash equilibrium policy. Our approach, dubbed Self-Play Preference Optimization (SPPO), utilizes iterative policy updates to provably approximate the Nash equilibrium. Additionally, we propose a new SPPO objective which is both strongly motivated by theory and is simple and effective in practice. In our experiments, using only 60k prompts (without responses) from the UltraFeedback dataset and without any prompt augmentation, by leveraging a pre-trained preference model PairRM with only 0.4B parameters, SPPO can obtain a model from fine-tuning Mistral-7B-Instruct-v0.2 that achieves the state-of-the-art length-controlled win-rate of 28.53% against GPT-4-Turbo on AlpacaEval 2.0. It also outperforms the (iterative) DPO and IPO on MT-Bench, Arena-Hard, and the Open LLM Leaderboard. Starting from a stronger base model Llama-3-8B-Instruct, we are able to achieve a length-controlled win rate of 38.77%. Notably, the strong performance of SPPO is achieved without additional external supervision (e.g., responses, preferences, etc.) from GPT-4 or other stronger language models. Codes are available at https://github.com/uclaml/SPPO.
This study presents a targeted model editing analysis focused on the latest large language model, Llama-3. We explore the efficacy of popular model editing techniques - ROME, MEMIT, and EMMET, which are designed for precise layer interventions. We identify the most effective layers for targeted edits through an evaluation that encompasses up to 4096 edits across three distinct strategies: sequential editing, batch editing, and a hybrid approach we call as sequential-batch editing. Our findings indicate that increasing edit batch-sizes may degrade model performance more significantly than using smaller edit batches sequentially for equal number of edits. With this, we argue that sequential model editing is an important component for scaling model editing methods and future research should focus on methods that combine both batched and sequential editing. This observation suggests a potential limitation in current model editing methods which push towards bigger edit batch sizes, and we hope it paves way for future investigations into optimizing batch sizes and model editing performance.
Recent studies introduced effective compression techniques for Large Language Models (LLMs) via post-training quantization or low-bit weight representation. Although quantized weights offer storage efficiency and allow for faster inference, existing works have indicated that quantization might compromise performance and exacerbate biases in LLMs. This study investigates the confidence and calibration of quantized models, considering factors such as language model type and scale as contributors to quantization loss. Firstly, we reveal that quantization with GPTQ to 4-bit results in a decrease in confidence regarding true labels, with varying impacts observed among different language models. Secondly, we observe fluctuations in the impact on confidence across different scales. Finally, we propose an explanation for quantization loss based on confidence levels, indicating that quantization disproportionately affects samples where the full model exhibited low confidence levels in the first place.
Causal reasoning is viewed as crucial for achieving human-level machine intelligence. Recent advances in language models have expanded the horizons of artificial intelligence across various domains, sparking inquiries into their potential for causal reasoning. In this work, we introduce Causal evaluation of Language Models (CaLM), which, to the best of our knowledge, is the first comprehensive benchmark for evaluating the causal reasoning capabilities of language models. First, we propose the CaLM framework, which establishes a foundational taxonomy consisting of four modules: causal target (i.e., what to evaluate), adaptation (i.e., how to obtain the results), metric (i.e., how to measure the results), and error (i.e., how to analyze the bad results). This taxonomy defines a broad evaluation design space while systematically selecting criteria and priorities. Second, we compose the CaLM dataset, comprising 126,334 data samples, to provide curated sets of causal targets, adaptations, metrics, and errors, offering extensive coverage for diverse research pursuits. Third, we conduct an extensive evaluation of 28 leading language models on a core set of 92 causal targets, 9 adaptations, 7 metrics, and 12 error types. Fourth, we perform detailed analyses of the evaluation results across various dimensions (e.g., adaptation, scale). Fifth, we present 50 high-level empirical findings across 9 dimensions (e.g., model), providing valuable guidance for future language model development. Finally, we develop a multifaceted platform, including a website, leaderboards, datasets, and toolkits, to support scalable and adaptable assessments. We envision CaLM as an ever-evolving benchmark for the community, systematically updated with new causal targets, adaptations, models, metrics, and error types to reflect ongoing research advancements. Project website is at https://opencausalab.github.io/CaLM.
The growth of social networks makes toxic content spread rapidly. Hate speech detection is a task to help decrease the number of harmful comments. With the diversity in the hate speech created by users, it is necessary to interpret the hate speech besides detecting it. Hence, we propose a methodology to construct a system for targeted hate speech detection from online streaming texts from social media. We first introduce the ViTHSD - a targeted hate speech detection dataset for Vietnamese Social Media Texts. The dataset contains 10K comments, each comment is labeled to specific targets with three levels: clean, offensive, and hate. There are 5 targets in the dataset, and each target is labeled with the corresponding level manually by humans with strict annotation guidelines. The inter-annotator agreement obtained from the dataset is 0.45 by Cohen's Kappa index, which is indicated as a moderate level. Then, we construct a baseline for this task by combining the Bi-GRU-LSTM-CNN with the pre-trained language model to leverage the power of text representation of BERTology. Finally, we suggest a methodology to integrate the baseline model for targeted hate speech detection into the online streaming system for practical application in preventing hateful and offensive content on social media.
Chain-of-Thought (CoT) reasoning often fails to faithfully reflect a language model's underlying decision process. We address this by making CoT text causally essential in a "Markovian" language model, factoring next-token prediction through an intermediate CoT and training it to predict future tokens independently of the original prompt. We formalize this via an "informativeness" objective that quantifies how much a trained CoT improves next-token predictions over a baseline. Using policy gradient, we show that Llama 3.1 8B achieves a 33.2% absolute accuracy improvement on GSM8K. Perturbation tests confirm stronger reliance on the CoT, while cross-model transfers indicate these reasoning traces generalize across interpreters. Our approach enhances both accuracy and interpretability, potentially extending CoT reasoning to arbitrarily long contexts and diverse tasks.
Out-of-distribution (OOD) problems in few-shot classification (FSC) occur when novel classes sampled from testing distributions differ from base classes drawn from training distributions, which considerably degrades the performance of deep learning models deployed in real-world applications. Recent studies suggest that the OOD problems in FSC mainly including: (a) cross-domain few-shot classification (CD-FSC) and (b) spurious-correlation few-shot classification (SC-FSC). Specifically, CD-FSC occurs when a classifier learns transferring knowledge from base classes drawn from seen training distributions but recognizes novel classes sampled from unseen testing distributions. In contrast, SC-FSC arises when a classifier relies on non-causal features (or contexts) that happen to be correlated with the labels (or concepts) in base classes but such relationships no longer hold during the model deployment. Despite CD-FSC has been extensively studied, SC-FSC remains understudied due to lack of the corresponding evaluation benchmarks. To this end, we present Meta Concept Context (MetaCoCo), a benchmark with spurious-correlation shifts collected from real-world scenarios. Moreover, to quantify the extent of spurious-correlation shifts of the presented MetaCoCo, we further propose a metric by using CLIP as a pre-trained vision-language model. Extensive experiments on the proposed benchmark are performed to evaluate the state-of-the-art methods in FSC, cross-domain shifts, and self-supervised learning. The experimental results show that the performance of the existing methods degrades significantly in the presence of spurious-correlation shifts. We open-source all codes of our benchmark and hope that the proposed MetaCoCo can facilitate future research on spurious-correlation shifts problems in FSC. The code is available at: https://github.com/remiMZ/MetaCoCo-ICLR24.
Large Language Models (LLMs) have catalyzed significant advancements in Natural Language Processing (NLP), yet they encounter challenges such as hallucination and the need for domain-specific knowledge. To mitigate these, recent methodologies have integrated information retrieved from external resources with LLMs, substantially enhancing their performance across NLP tasks. This survey paper addresses the absence of a comprehensive overview on Retrieval-Augmented Language Models (RALMs), both Retrieval-Augmented Generation (RAG) and Retrieval-Augmented Understanding (RAU), providing an in-depth examination of their paradigm, evolution, taxonomy, and applications. The paper discusses the essential components of RALMs, including Retrievers, Language Models, and Augmentations, and how their interactions lead to diverse model structures and applications. RALMs demonstrate utility in a spectrum of tasks, from translation and dialogue systems to knowledge-intensive applications. The survey includes several evaluation methods of RALMs, emphasizing the importance of robustness, accuracy, and relevance in their assessment. It also acknowledges the limitations of RALMs, particularly in retrieval quality and computational efficiency, offering directions for future research. In conclusion, this survey aims to offer a structured insight into RALMs, their potential, and the avenues for their future development in NLP. The paper is supplemented with a Github Repository containing the surveyed works and resources for further study: https://github.com/2471023025/RALM_Survey.
Retrieval-Augmented Generation (RAG) systems are widely used across various industries for querying closed-domain and in-house knowledge bases. However, evaluating these systems presents significant challenges due to the private nature of closed-domain data and a scarcity of queries with verifiable ground truths. Moreover, there is a lack of analytical methods to diagnose problematic modules and identify types of failure, such as those caused by knowledge deficits or issues with robustness. To address these challenges, we introduce GRAMMAR (GRounded And Modular Methodology for Assessment of RAG), an evaluation framework comprising a grounded data generation process and an evaluation protocol that effectively pinpoints defective modules. Our validation experiments reveal that GRAMMAR provides a reliable approach for identifying vulnerable modules and supports hypothesis testing for textual form vulnerabilities. An open-source tool accompanying this framework is available in our GitHub repository (see https://github.com/xinzhel/grammar), allowing for easy reproduction of our results and enabling reliable and modular evaluation in closed-domain settings.
In this paper, we establish a benchmark for table visual question answering, referred to as the TableVQA-Bench, derived from pre-existing table question-answering (QA) and table structure recognition datasets. It is important to note that existing datasets have not incorporated images or QA pairs, which are two crucial components of TableVQA. As such, the primary objective of this paper is to obtain these necessary components. Specifically, images are sourced either through the application of a \textit{stylesheet} or by employing the proposed table rendering system. QA pairs are generated by exploiting the large language model (LLM) where the input is a text-formatted table. Ultimately, the completed TableVQA-Bench comprises 1,500 QA pairs. We comprehensively compare the performance of various multi-modal large language models (MLLMs) on TableVQA-Bench. GPT-4V achieves the highest accuracy among commercial and open-sourced MLLMs from our experiments. Moreover, we discover that the number of vision queries plays a significant role in TableVQA performance. To further analyze the capabilities of MLLMs in comparison to their LLM backbones, we investigate by presenting image-formatted tables to MLLMs and text-formatted tables to LLMs, respectively. Our findings suggest that processing visual inputs is more challenging than text inputs, as evidenced by the lower performance of MLLMs, despite generally requiring higher computational costs than LLMs. The proposed TableVQA-Bench and evaluation codes are available at \href{https://github.com/naver-ai/tablevqabench}{https://github.com/naver-ai/tablevqabench}.
Cross-lingual transfer has become an effective way of transferring knowledge between languages. In this paper, we explore an often overlooked aspect in this domain: the influence of the source language of a language model on language transfer performance. We consider a case where the target language and its script are not part of the pre-trained model. We conduct a series of experiments on monolingual and multilingual models that are pre-trained on different tokenization methods to determine factors that affect cross-lingual transfer to a new language with a unique script. Our findings reveal the importance of the tokenizer as a stronger factor than the shared script, language similarity, and model size.
The representation of events in text plays a significant role in various NLP tasks. Recent research demonstrates that contrastive learning has the ability to improve event comprehension capabilities of Pre-trained Language Models (PLMs) and enhance the performance of event representation learning. However, the efficacy of event representation learning based on contrastive learning and PLMs is limited by the short length of event texts. The length of event texts differs significantly from the text length used in the pre-training of PLMs. As a result, there is inconsistency in the distribution of text length between pre-training and event representation learning, which may undermine the learning process of event representation based on PLMs. In this study, we present PromptCL, a novel framework for event representation learning that effectively elicits the capabilities of PLMs to comprehensively capture the semantics of short event texts. PromptCL utilizes a Prompt template borrowed from prompt learning to expand the input text during Contrastive Learning. This helps in enhancing the event representation learning by providing a structured outline of the event components. Moreover, we propose Subject-Predicate-Object (SPO) word order and Event-oriented Masked Language Modeling (EventMLM) to train PLMs to understand the relationships between event components. Our experimental results demonstrate that PromptCL outperforms state-of-the-art baselines on event related tasks. Additionally, we conduct a thorough analysis and demonstrate that using a prompt results in improved generalization capabilities for event representations. Our code will be available at https://github.com/YuboFeng2023/PromptCL.
We introduce WellPlay, a reasoning dataset for multi-agent conversational inference in Murder Mystery Games (MMGs). WellPlay comprises 1,482 inferential questions across 12 games, spanning objectives, reasoning, and relationship understanding, and establishes a systematic benchmark for evaluating agent reasoning abilities in complex social settings. Building on this foundation, we present PLAYER*, a novel framework for Large Language Model (LLM)-based agents in MMGs. MMGs pose unique challenges, including undefined state spaces, absent intermediate rewards, and the need for strategic reasoning through natural language. PLAYER* addresses these challenges with a sensor-based state representation and an information-driven strategy that optimises questioning and suspect pruning. Experiments show that PLAYER* outperforms existing methods in reasoning accuracy, efficiency, and agent-human interaction, advancing reasoning agents for complex social scenarios.