Recent studies have shown that generative language models often reflect and amplify societal biases in their outputs. However, these studies frequently conflate observed biases with other task-specific shortcomings, such as comprehension failure. For example, when a model misinterprets a text and produces a response that reinforces a stereotype, it becomes difficult to determine whether the issue arises from inherent bias or from a misunderstanding of the given content. In this paper, we conduct a multi-faceted evaluation that distinctly disentangles bias from flaws within the reading comprehension task. We propose a targeted stereotype mitigation framework that implicitly mitigates observed stereotypes in generative models through instruction-tuning on general-purpose datasets. We reduce stereotypical outputs by over 60% across multiple dimensions -- including nationality, age, gender, disability, and physical appearance -- by addressing comprehension-based failures, and without relying on explicit debiasing techniques. We evaluate several state-of-the-art generative models to demonstrate the effectiveness of our approach while maintaining the overall utility. Our findings highlight the need to critically disentangle the concept of `bias' from other types of errors to build more targeted and effective mitigation strategies. CONTENT WARNING: Some examples contain offensive stereotypes.
This study examines whether large language models (LLMs), such as ChatGPT, specifically the latest GPT-4o-mini, exhibit sensitivity to repeated prompts and whether repeating a question can improve response accuracy. We hypothesize that reiterating a question within a single prompt might enhance the model's focus on key elements of the query. To test this, we evaluate ChatGPT's performance on a large sample of two reading comprehension datasets under both open-book and closed-book settings, varying the repetition of each question to 1, 3, or 5 times per prompt. Our findings indicate that the model does not demonstrate sensitivity to repeated questions, highlighting its robustness and consistency in this context.
Scientific literature is typically dense, requiring significant background knowledge and deep comprehension for effective engagement. We introduce SciDQA, a new dataset for reading comprehension that challenges LLMs for a deep understanding of scientific articles, consisting of 2,937 QA pairs. Unlike other scientific QA datasets, SciDQA sources questions from peer reviews by domain experts and answers by paper authors, ensuring a thorough examination of the literature. We enhance the dataset's quality through a process that carefully filters out lower quality questions, decontextualizes the content, tracks the source document across different versions, and incorporates a bibliography for multi-document question-answering. Questions in SciDQA necessitate reasoning across figures, tables, equations, appendices, and supplementary materials, and require multi-document reasoning. We evaluate several open-source and proprietary LLMs across various configurations to explore their capabilities in generating relevant and factual responses. Our comprehensive evaluation, based on metrics for surface-level similarity and LLM judgements, highlights notable performance discrepancies. SciDQA represents a rigorously curated, naturally derived scientific QA dataset, designed to facilitate research on complex scientific text understanding.
This study explores the potential of using training dynamics as an automated alternative to human annotation for evaluating the quality of training data. The framework used is Data Maps, which classifies data points into categories such as easy-to-learn, hard-to-learn, and ambiguous (Swayamdipta et al., 2020). Swayamdipta et al. (2020) highlight that difficult-to-learn examples often contain errors, and ambiguous cases significantly impact model training. To confirm the reliability of these findings, we replicated the experiments using a challenging dataset, with a focus on medical question answering. In addition to text comprehension, this field requires the acquisition of detailed medical knowledge, which further complicates the task. A comprehensive evaluation was conducted to assess the feasibility and transferability of the Data Maps framework to the medical domain. The evaluation indicates that the framework is unsuitable for addressing datasets' unique challenges in answering medical questions.
Objective: Scientific publications play a crucial role in uncovering insights, testing novel drugs, and shaping healthcare policies. Accessing the quality of publications requires evaluating their Risk of Bias (RoB), a process typically conducted by human reviewers. In this study, we introduce a new dataset for machine reading comprehension and RoB assessment and present RoBIn (Risk of Bias Inference), an innovative model crafted to automate such evaluation. The model employs a dual-task approach, extracting evidence from a given context and assessing the RoB based on the gathered evidence. Methods: We use data from the Cochrane Database of Systematic Reviews (CDSR) as ground truth to label open-access clinical trial publications from PubMed. This process enabled us to develop training and test datasets specifically for machine reading comprehension and RoB inference. Additionally, we created extractive (RoBInExt) and generative (RoBInGen) Transformer-based approaches to extract relevant evidence and classify the RoB effectively. Results: RoBIn is evaluated across various settings and benchmarked against state-of-the-art methods for RoB inference, including large language models in multiple scenarios. In most cases, the best-performing RoBIn variant surpasses traditional machine learning and LLM-based approaches, achieving an ROC AUC of 0.83. Conclusion: Based on the evidence extracted from clinical trial reports, RoBIn performs a binary classification to decide whether the trial is at a low RoB or a high/unclear RoB. We found that both RoBInGen and RoBInExt are robust and have the best results in many settings.
One useful application of NLP models is to support people in reading complex text from unfamiliar domains (e.g., scientific articles). Simplifying the entire text makes it understandable but sometimes removes important details. On the contrary, helping adult readers understand difficult concepts in context can enhance their vocabulary and knowledge. In a preliminary human study, we first identify that lack of context and unfamiliarity with difficult concepts is a major reason for adult readers' difficulty with domain-specific text. We then introduce "targeted concept simplification," a simplification task for rewriting text to help readers comprehend text containing unfamiliar concepts. We also introduce WikiDomains, a new dataset of 22k definitions from 13 academic domains paired with a difficult concept within each definition. We benchmark the performance of open-source and commercial LLMs and a simple dictionary baseline on this task across human judgments of ease of understanding and meaning preservation. Interestingly, our human judges preferred explanations about the difficult concept more than simplification of the concept phrase. Further, no single model achieved superior performance across all quality dimensions, and automated metrics also show low correlations with human evaluations of concept simplification ($\sim0.2$), opening up rich avenues for research on personalized human reading comprehension support.
Recent studies have demonstrated that large language models (LLMs) exhibit significant biases in evaluation tasks, particularly in preferentially rating and favoring self-generated content. However, the extent to which this bias manifests in fact-oriented tasks, especially within retrieval-augmented generation (RAG) frameworks-where keyword extraction and factual accuracy take precedence over stylistic elements-remains unclear. Our study addresses this knowledge gap by simulating two critical phases of the RAG framework. In the first phase, we access the suitability of human-authored versus model-generated passages, emulating the pointwise reranking process. The second phase involves conducting pairwise reading comprehension tests to simulate the generation process. Contrary to previous findings indicating a self-preference in rating tasks, our results reveal no significant self-preference effect in RAG frameworks. Instead, we observe that factual accuracy significantly influences LLMs' output, even in the absence of prior knowledge. Our research contributes to the ongoing discourse on LLM biases and their implications for RAG-based system, offering insights that may inform the development of more robust and unbiased LLM systems.
Can human reading comprehension be assessed from eye movements in reading? In this work, we address this longstanding question using large-scale eyetracking data over textual materials that are geared towards behavioral analyses of reading comprehension. We focus on a fine-grained and largely unaddressed task of predicting reading comprehension from eye movements at the level of a single question over a passage. We tackle this task using three new multimodal language models, as well as a battery of prior models from the literature. We evaluate the models' ability to generalize to new textual items, new participants, and the combination of both, in two different reading regimes, ordinary reading and information seeking. The evaluations suggest that although the task is highly challenging, eye movements contain useful signals for fine-grained prediction of reading comprehension. Code and data will be made publicly available.
LLM text decoding is key component for perceived LLM quality. We demonstrate two experiments showing that decoding methods could be improved by manipulation of token probabilities. First, we test few LLM on SummEval summary scoring dataset, to measure reading comprehension. We compare scores from greedy decoding to expected values over the next token distribution. We scale logits by large temperature to increase the entropy of scores. This allows strong improvement of performance on SummEval (in terms of correlations to human judgement). We see improvement from 6-8% to 13-28% for 7B Mistral and from 20%-46% to 37%-56% for Mixtral, beating GPT 4 0314 result on two metrics. Part of the gain seems related to positional bias. Secondly, we use probability-based tree sampling algorithm, to examine all most probable generations for given prompt.
We test the robustness of debate as a method of scalable oversight by training models to debate with data generated via self-play. In a long-context reading comprehension task, we find that language model based evaluators answer questions more accurately when judging models optimized to win debates. By contrast, we find no such relationship for consultancy models trained to persuade a judge without an opposing debater present. In quantitative and qualitative comparisons between our debate models and novel consultancy baselines, we find evidence that debate training encourages stronger and more informative arguments, showing promise that it can help provide high-quality supervision for tasks that are difficult to directly evaluate.
Learning performance data describe correct and incorrect answers or problem-solving attempts in adaptive learning, such as in intelligent tutoring systems (ITSs). Learning performance data tend to be highly sparse (80\%\(\sim\)90\% missing observations) in most real-world applications due to adaptive item selection. This data sparsity presents challenges to using learner models to effectively predict future performance explore new hypotheses about learning. This article proposes a systematic framework for augmenting learner data to address data sparsity in learning performance data. First, learning performance is represented as a three-dimensional tensor of learners' questions, answers, and attempts, capturing longitudinal knowledge states during learning. Second, a tensor factorization method is used to impute missing values in sparse tensors of collected learner data, thereby grounding the imputation on knowledge tracing tasks that predict missing performance values based on real observations. Third, a module for generating patterns of learning is used. This study contrasts two forms of generative Artificial Intelligence (AI), including Generative Adversarial Networks (GANs) and Generate Pre-Trained Transformers (GPT) to generate data associated with different clusters of learner data. We tested this approach on an adult literacy dataset from AutoTutor lessons developed for Adult Reading Comprehension (ARC). We found that: (1) tensor factorization improved the performance in tracing and predicting knowledge mastery compared with other knowledge tracing techniques without data augmentation, showing higher relative fidelity for this imputation method, and (2) the GAN-based simulation showed greater overall stability and less statistical bias based on a divergence evaluation with varying simulation sample sizes compared to GPT.
Logical reading comprehension is a challenging task that entails grasping the underlying semantics of text and applying reasoning to deduce the correct answer. Prior researches have primarily focused on enhancing logical reasoning capabilities through Chain-of-Thought (CoT) or data augmentation. However, previous work constructing chain-of-thought rationales concentrates solely on analyzing correct options, neglecting the incorrect alternatives. Addtionally, earlier efforts on data augmentation by altering contexts rely on rule-based methods, which result in generated contexts that lack diversity and coherence. To address these issues, we propose a Premise-Oriented Data Augmentation (PODA) framework. This framework can generate CoT rationales including analyses for both correct and incorrect options, while constructing diverse and high-quality counterfactual contexts from incorrect candidate options. We integrate summarizing premises and identifying premises for each option into rationales. Subsequently, we employ multi-step prompts with identified premises to construct counterfactual context. To facilitate the model's capabilities to better differentiate the reasoning process associated with each option, we introduce a novel thought-path contrastive learning method that compares reasoning paths between the original and counterfactual samples. Experimental results on three representative LLMs demonstrate that our method can improve the baselines substantially across two challenging logical reasoning benchmarks (ReClor and LogiQA 2.0). The data and code are released at https://github.com/lalalamdbf/TPReasoner.
This study explores replacing Transformers in Visual Language Models (VLMs) with Mamba, a recent structured state space model (SSM) that demonstrates promising performance in sequence modeling. We test models up to 3B parameters under controlled conditions, showing that Mamba-based VLMs outperforms Transformers-based VLMs in captioning, question answering, and reading comprehension. However, we find that Transformers achieve greater performance in visual grounding and the performance gap widens with scale. We explore two hypotheses to explain this phenomenon: 1) the effect of task-agnostic visual encoding on the updates of the hidden states, and 2) the difficulty in performing visual grounding from the perspective of in-context multimodal retrieval. Our results indicate that a task-aware encoding yields minimal performance gains on grounding, however, Transformers significantly outperform Mamba at in-context multimodal retrieval. Overall, Mamba shows promising performance on tasks where the correct output relies on a summary of the image but struggles when retrieval of explicit information from the context is required.
State-of-the-art Large Language Models (LLMs) are accredited with an increasing number of different capabilities, ranging from reading comprehension, over advanced mathematical and reasoning skills to possessing scientific knowledge. In this paper we focus on their multi-hop reasoning capability: the ability to identify and integrate information from multiple textual sources. Given the concerns with the presence of simplifying cues in existing multi-hop reasoning benchmarks, which allow models to circumvent the reasoning requirement, we set out to investigate, whether LLMs are prone to exploiting such simplifying cues. We find evidence that they indeed circumvent the requirement to perform multi-hop reasoning, but they do so in more subtle ways than what was reported about their fine-tuned pre-trained language model (PLM) predecessors. Motivated by this finding, we propose a challenging multi-hop reasoning benchmark, by generating seemingly plausible multi-hop reasoning chains, which ultimately lead to incorrect answers. We evaluate multiple open and proprietary state-of-the-art LLMs, and find that their performance to perform multi-hop reasoning is affected, as indicated by up to 45% relative decrease in F1 score when presented with such seemingly plausible alternatives. We conduct a deeper analysis and find evidence that while LLMs tend to ignore misleading lexical cues, misleading reasoning paths indeed present a significant challenge.
In recent years, Large Language Models (LLMs) have demonstrated significant improvements across a variety of tasks, one of which is the long-context capability. The key to improving long-context performance lies in effective data organization and management strategies that integrate data from multiple domains and optimize the context window during training. Through extensive experimental analysis, we identified three key challenges in designing effective data management strategies that enable the model to achieve long-context capability without sacrificing performance in other tasks: (1) a shortage of long documents across multiple domains, (2) effective construction of context windows, and (3) efficient organization of large-scale datasets. To address these challenges, we introduce DataSculpt, a novel data management framework designed for long-context training. We first formulate the organization of training data as a multi-objective combinatorial optimization problem, focusing on attributes including relevance, homogeneity, integrity, and efficiency. Specifically, our approach utilizes a coarse-to-fine methodology to optimize training data organization both efficiently and effectively. We begin by clustering the data based on semantic similarity (coarse), followed by a multi-objective greedy search within each cluster to score and concatenate documents into various context windows (fine). Our comprehensive evaluations demonstrate that DataSculpt significantly enhances long-context training performance, resulting in improvements of 18.09% in retrieval augmentation, 21.23% in summarization, 21.27% in reading comprehension, and a 3.81% increase in code completion, while also maintaining overall model proficiency with a 4.88% improvement.
Advancements in AI and natural language processing have revolutionized machine-human language interactions, with question answering (QA) systems playing a pivotal role. The knowledge base question answering (KBQA) task, utilizing structured knowledge graphs (KG), allows for handling extensive knowledge-intensive questions. However, a significant gap exists in KBQA datasets, especially for low-resource languages. Many existing construction pipelines for these datasets are outdated and inefficient in human labor, and modern assisting tools like Large Language Models (LLM) are not utilized to reduce the workload. To address this, we have designed and implemented a modern, semi-automated approach for creating datasets, encompassing tasks such as KBQA, Machine Reading Comprehension (MRC), and Information Retrieval (IR), tailored explicitly for low-resource environments. We executed this pipeline and introduced the PUGG dataset, the first Polish KBQA dataset, and novel datasets for MRC and IR. Additionally, we provide a comprehensive implementation, insightful findings, detailed statistics, and evaluation of baseline models.
Grounded Multimodal Named Entity Recognition (GMNER) is an emerging information extraction (IE) task, aiming to simultaneously extract entity spans, types, and corresponding visual regions of entities from given sentence-image pairs data. Recent unified methods employing machine reading comprehension or sequence generation-based frameworks show limitations in this difficult task. The former, utilizing human-designed type queries, struggles to differentiate ambiguous entities, such as Jordan (Person) and off-White x Jordan (Shoes). The latter, following the one-by-one decoding order, suffers from exposure bias issues. We maintain that these works misunderstand the relationships of multimodal entities. To tackle these, we propose a novel unified framework named Multi-grained Query-guided Set Prediction Network (MQSPN) to learn appropriate relationships at intra-entity and inter-entity levels. Specifically, MQSPN explicitly aligns textual entities with visual regions by employing a set of learnable queries to strengthen intra-entity connections. Based on distinct intra-entity modeling, MQSPN reformulates GMNER as a set prediction, guiding models to establish appropriate inter-entity relationships from a optimal global matching perspective. Additionally, we incorporate a query-guided Fusion Net (QFNet) as a glue network to boost better alignment of two-level relationships. Extensive experiments demonstrate that our approach achieves state-of-the-art performances in widely used benchmarks.
In this work, we present TextHarmony, a unified and versatile multimodal generative model proficient in comprehending and generating visual text. Simultaneously generating images and texts typically results in performance degradation due to the inherent inconsistency between vision and language modalities. To overcome this challenge, existing approaches resort to modality-specific data for supervised fine-tuning, necessitating distinct model instances. We propose Slide-LoRA, which dynamically aggregates modality-specific and modality-agnostic LoRA experts, partially decoupling the multimodal generation space. Slide-LoRA harmonizes the generation of vision and language within a singular model instance, thereby facilitating a more unified generative process. Additionally, we develop a high-quality image caption dataset, DetailedTextCaps-100K, synthesized with a sophisticated closed-source MLLM to enhance visual text generation capabilities further. Comprehensive experiments across various benchmarks demonstrate the effectiveness of the proposed approach. Empowered by Slide-LoRA, TextHarmony achieves comparable performance to modality-specific fine-tuning results with only a 2% increase in parameters and shows an average improvement of 2.5% in visual text comprehension tasks and 4.0% in visual text generation tasks. Our work delineates the viability of an integrated approach to multimodal generation within the visual text domain, setting a foundation for subsequent inquiries. Code is available at https://github.com/bytedance/TextHarmony.
Using questions in written text is an effective strategy to enhance readability. However, what makes an active reading question good, what the linguistic role of these questions is, and what is their impact on human reading remains understudied. We introduce GuidingQ, a dataset of 10K in-text questions from textbooks and scientific articles. By analyzing the dataset, we present a comprehensive understanding of the use, distribution, and linguistic characteristics of these questions. Then, we explore various approaches to generate such questions using language models. Our results highlight the importance of capturing inter-question relationships and the challenge of question position identification in generating these questions. Finally, we conduct a human study to understand the implication of such questions on reading comprehension. We find that the generated questions are of high quality and are almost as effective as human-written questions in terms of improving readers' memorization and comprehension.
Recently, there has been a growing interest among large language model (LLM) developers in LLM-based document reading systems, which enable users to upload their own documents and pose questions related to the document contents, going beyond simple reading comprehension tasks. Consequently, these systems have been carefully designed to tackle challenges such as file parsing, metadata extraction, multi-modal information understanding and long-context reading. However, no current benchmark exists to evaluate their performance in such scenarios, where a raw file and questions are provided as input, and a corresponding response is expected as output. In this paper, we introduce DocBench, a new benchmark designed to evaluate LLM-based document reading systems. Our benchmark involves a meticulously crafted process, including the recruitment of human annotators and the generation of synthetic questions. It includes 229 real documents and 1,102 questions, spanning across five different domains and four major types of questions. We evaluate both proprietary LLM-based systems accessible via web interfaces or APIs, and a parse-then-read pipeline employing open-source LLMs. Our evaluations reveal noticeable gaps between existing LLM-based document reading systems and human performance, underscoring the challenges of developing proficient systems. To summarize, DocBench aims to establish a standardized benchmark for evaluating LLM-based document reading systems under diverse real-world scenarios, thereby guiding future advancements in this research area.