James Thorne

Seonghyeon Ye, Doyoung Kim, Sungdong Kim et al. · cmu

Evaluation of Large Language Models (LLMs) is challenging because instruction-following necessitates alignment with human values and the required set of skills varies depending on the instruction. However, previous studies have mainly focused on coarse-grained evaluation (i.e. overall preference-based evaluation), which limits interpretability since it does not consider the nature of user instructions that require instance-wise skill composition. In this paper, we introduce FLASK (Fine-grained Language Model Evaluation based on Alignment Skill Sets), a fine-grained evaluation protocol for both human-based and model-based evaluation which decomposes coarse-level scoring to a skill set-level scoring for each instruction. We experimentally observe that the fine-graininess of evaluation is crucial for attaining a holistic view of model performance and increasing the reliability of the evaluation. Using FLASK, we compare multiple open-source and proprietary LLMs and observe a high correlation between model-based and human-based evaluations. We publicly release the evaluation data and code implementation at https://github.com/kaistAI/FLASK.

Seungone Kim, Jamin Shin, Yejin Cho et al. · cmu, uw

Recently, using a powerful proprietary Large Language Model (LLM) (e.g., GPT-4) as an evaluator for long-form responses has become the de facto standard. However, for practitioners with large-scale evaluation tasks and custom criteria in consideration (e.g., child-readability), using proprietary LLMs as an evaluator is unreliable due to the closed-source nature, uncontrolled versioning, and prohibitive costs. In this work, we propose Prometheus, a fully open-source LLM that is on par with GPT-4's evaluation capabilities when the appropriate reference materials (reference answer, score rubric) are accompanied. We first construct the Feedback Collection, a new dataset that consists of 1K fine-grained score rubrics, 20K instructions, and 100K responses and language feedback generated by GPT-4. Using the Feedback Collection, we train Prometheus, a 13B evaluator LLM that can assess any given long-form text based on customized score rubric provided by the user. Experimental results show that Prometheus scores a Pearson correlation of 0.897 with human evaluators when evaluating with 45 customized score rubrics, which is on par with GPT-4 (0.882), and greatly outperforms ChatGPT (0.392). Furthermore, measuring correlation with GPT-4 with 1222 customized score rubrics across four benchmarks (MT Bench, Vicuna Bench, Feedback Bench, Flask Eval) shows similar trends, bolstering Prometheus's capability as an evaluator LLM. Lastly, Prometheus achieves the highest accuracy on two human preference benchmarks (HHH Alignment & MT Bench Human Judgment) compared to open-sourced reward models explicitly trained on human preference datasets, highlighting its potential as an universal reward model. We open-source our code, dataset, and model at https://kaistai.github.io/prometheus/.

Jiyoung Lee, Seungho Kim, Seunghyun Won et al. · nvidia, utoronto

AI alignment refers to models acting towards human-intended goals, preferences, or ethical principles. Given that most large-scale deep learning models act as black boxes and cannot be manually controlled, analyzing the similarity between models and humans can be a proxy measure for ensuring AI safety. In this paper, we focus on the models' visual perception alignment with humans, further referred to as AI-human visual alignment. Specifically, we propose a new dataset for measuring AI-human visual alignment in terms of image classification, a fundamental task in machine perception. In order to evaluate AI-human visual alignment, a dataset should encompass samples with various scenarios that may arise in the real world and have gold human perception labels. Our dataset consists of three groups of samples, namely Must-Act (i.e., Must-Classify), Must-Abstain, and Uncertain, based on the quantity and clarity of visual information in an image and further divided into eight categories. All samples have a gold human perception label; even Uncertain (severely blurry) sample labels were obtained via crowd-sourcing. The validity of our dataset is verified by sampling theory, statistical theories related to survey design, and experts in the related fields. Using our dataset, we analyze the visual alignment and reliability of five popular visual perception models and seven abstention methods. Our code and data is available at https://github.com/jiyounglee-0523/VisAlign.

Jiwoo Hong, Yejin Cho, Jaemin Jung et al.

We address an important gap in detecting political bias in news articles. Previous works that perform document classification can be influenced by the writing style of each news outlet, leading to overfitting and limited generalizability. Our approach overcomes this limitation by considering both the sentence-level semantics and the document-level rhetorical structure, resulting in a more robust and style-agnostic approach to detecting political bias in news articles. We introduce a novel multi-head hierarchical attention model that effectively encodes the structure of long documents through a diverse ensemble of attention heads. While journalism follows a formalized rhetorical structure, the writing style may vary by news outlet. We demonstrate that our method overcomes this domain dependency and outperforms previous approaches for robustness and accuracy. Further analysis and human evaluation demonstrate the ability of our model to capture common discourse structures in journalism. Our code is available at: https://github.com/xfactlab/emnlp2023-Document-Hierarchy

Yongjin Yang, Joonkee Kim, Yujin Kim et al.

With the proliferation of social media, accurate detection of hate speech has become critical to ensure safety online. To combat nuanced forms of hate speech, it is important to identify and thoroughly explain hate speech to help users understand its harmful effects. Recent benchmarks have attempted to tackle this issue by training generative models on free-text annotations of implications in hateful text. However, we find significant reasoning gaps in the existing annotations schemes, which may hinder the supervision of detection models. In this paper, we introduce a hate speech detection framework, HARE, which harnesses the reasoning capabilities of large language models (LLMs) to fill these gaps in explanations of hate speech, thus enabling effective supervision of detection models. Experiments on SBIC and Implicit Hate benchmarks show that our method, using model-generated data, consistently outperforms baselines, using existing free-text human annotations. Analysis demonstrates that our method enhances the explanation quality of trained models and improves generalization to unseen datasets. Our code is available at https://github.com/joonkeekim/hare-hate-speech.git.

Philhoon Oh, James Thorne

For knowledge intensive NLP tasks, it has been widely accepted that accessing more information is a contributing factor to improvements in the model's end-to-end performance. However, counter-intuitively, too much context can have a negative impact on the model when evaluated on common question answering (QA) datasets. In this paper, we analyze how passages can have a detrimental effect on retrieve-then-read architectures used in question answering. Our empirical evidence indicates that the current read architecture does not fully leverage the retrieved passages and significantly degrades its performance when using the whole passages compared to utilizing subsets of them. Our findings demonstrate that model accuracy can be improved by 10% on two popular QA datasets by filtering out detrimental passages. Additionally, these outcomes are attained by utilizing existing retrieval methods without further training or data. We further highlight the challenges associated with identifying the detrimental passages. First, even with the correct context, the model can make an incorrect prediction, posing a challenge in determining which passages are most influential. Second, evaluation typically considers lexical matching, which is not robust to variations of correct answers. Despite these limitations, our experimental results underscore the pivotal role of identifying and removing these detrimental passages for the context-efficient retrieve-then-read pipeline. Code and data are available at https://github.com/xfactlab/emnlp2023-damaging-retrieval

Yejoon Lee, Philhoon Oh, James Thorne

Recent works in open-domain question answering (QA) have explored generating context passages from large language models (LLMs), replacing the traditional retrieval step in the QA pipeline. However, it is not well understood why generated passages can be more effective than retrieved ones. This study revisits the conventional formulation of QA and introduces the concept of knowledge corpus error. This error arises when the knowledge corpus used for retrieval is only a subset of the entire string space, potentially excluding more helpful passages that exist outside the corpus. LLMs may mitigate this shortcoming by generating passages in a larger space. We come up with an experiment of paraphrasing human-annotated gold context using LLMs to observe knowledge corpus error empirically. Our results across three QA benchmarks reveal an increased performance (10% - 13%) when using paraphrased passage, indicating a signal for the existence of knowledge corpus error. Our code is available at https://github.com/xfactlab/emnlp2023-knowledge-corpus-error

James Thorne

Document retrieval is a core component of many knowledge-intensive natural language processing task formulations such as fact verification and question answering. Sources of textual knowledge, such as Wikipedia articles, condition the generation of answers from the models. Recent advances in retrieval use sequence-to-sequence models to incrementally predict the title of the appropriate Wikipedia page given a query. However, this method requires supervision in the form of human annotation to label which Wikipedia pages contain appropriate context. This paper introduces a distant-supervision method that does not require any annotation to train autoregressive retrievers that attain competitive R-Precision and Recall in a zero-shot setting. Furthermore we show that with task-specific supervised fine-tuning, autoregressive retrieval performance for two Wikipedia-based fact verification tasks can approach or even exceed full supervision using less than $1/4$ of the annotated data indicating possible directions for data-efficient autoregressive retrieval.

Noah Lee, Jiwoo Hong, James Thorne

Large language models (LLMs) have shown potential as general evaluators along with the evident benefits of speed and cost. While their correlation against human annotators has been widely studied, consistency as evaluators is still understudied, raising concerns about the reliability of LLM evaluators. In this paper, we conduct extensive studies on the two aspects of consistency in LLM evaluations, Self-Consistency (SC) and Inter-scale Consistency (IC), on different scoring scales and criterion granularity with open-source and proprietary models. Our comprehensive analysis demonstrates that strong proprietary models are not necessarily consistent evaluators, highlighting the importance of considering consistency in assessing the capability of LLM evaluators.

Woojin Chung, Jiwoo Hong, Na Min An et al.

Stable pre-training is essential for achieving better-performing language models. However, tracking pre-training stability by calculating gradient variance at every step is impractical due to the significant computational costs. We explore Token Embedding Variability (TEV) as a simple and efficient proxy for assessing pre-training stability in language models with pre-layer normalization, given that shallower layers are more prone to gradient explosion (section 2.2). Moreover, we propose Multi-head Low-Rank Attention (MLRA) as an architecture to alleviate such instability by limiting the exponential growth of output embedding variance, thereby preventing the gradient explosion (section 3.2). Empirical results on GPT-2 with MLRA demonstrate increased stability and lower perplexity, particularly in deeper models.

Jiwoo Hong, Noah Lee, Eunki Kim et al.

The Bradley-Terry (BT) model is widely practiced in reward modeling for reinforcement learning with human feedback (RLHF). Despite its effectiveness, reward models (RMs) trained with BT model loss are prone to over-optimization, losing generalizability to unseen input distributions. In this paper, we study the cause of over-optimization in RM training and its downstream effects on the RLHF procedure, accentuating the importance of distributional robustness of RMs in unseen data. First, we show that the excessive dispersion of hidden state norms is the main source of over-optimization. Then, we propose batch-wise sum-to-zero regularization (BSR) to enforce zero-centered reward sum per batch, constraining the rewards with extreme magnitudes. We assess the impact of BSR in improving robustness in RMs through four scenarios of over-optimization, where BSR consistently manifests better robustness. Subsequently, we compare the plain BT model and BSR on RLHF training and empirically show that robust RMs better align the policy to the gold preference model. Finally, we apply BSR to high-quality data and models, which surpasses state-of-the-art RMs in the 8B scale by adding more than 5% in complex preference prediction tasks. By conducting RLOO training with 8B RMs, AlpacaEval 2.0 reduces generation length by 40% while adding a 7% increase in win rate, further highlighting that robustness in RMs induces robustness in RLHF training. We release the code, data, and models: https://github.com/LinkedIn-XFACT/RM-Robustness.

Jiwoo Hong, Noah Lee, James Thorne

While recent preference alignment algorithms for language models have demonstrated promising results, supervised fine-tuning (SFT) remains imperative for achieving successful convergence. In this paper, we study the crucial role of SFT within the context of preference alignment, emphasizing that a minor penalty for the disfavored generation style is sufficient for preference-aligned SFT. Building on this foundation, we introduce a straightforward and innovative reference model-free monolithic odds ratio preference optimization algorithm, ORPO, eliminating the necessity for an additional preference alignment phase. We demonstrate, both empirically and theoretically, that the odds ratio is a sensible choice for contrasting favored and disfavored styles during SFT across the diverse sizes from 125M to 7B. Specifically, fine-tuning Phi-2 (2.7B), Llama-2 (7B), and Mistral (7B) with ORPO on the UltraFeedback alone surpasses the performance of state-of-the-art language models with more than 7B and 13B parameters: achieving up to 12.20% on $\text{AlpacaEval}_{2.0}$ (Figure 1), 66.19% on IFEval (instruction-level loose, Table 6), and 7.32 in MT-Bench (Figure 12). We release code and model checkpoints for Mistral-ORPO-$α$ (7B) and Mistral-ORPO-$β$ (7B).

Jun Seong Kim, Kyaw Ye Thu, Javad Ismayilzada et al.

In a highly globalized world, it is important for multi-modal large language models (MLLMs) to recognize and respond correctly to mixed-cultural inputs. For example, a model should correctly identify kimchi (Korean food) in an image both when an Asian woman is eating it, as well as an African man is eating it. However, current MLLMs show an over-reliance on the visual features of the person, leading to misclassification of the entities. To examine the robustness of MLLMs to different ethnicity, we introduce MixCuBe, a cross-cultural bias benchmark, and study elements from five countries and four ethnicities. Our findings reveal that MLLMs achieve both higher accuracy and lower sensitivity to such perturbation for high-resource cultures, but not for low-resource cultures. GPT-4o, the best-performing model overall, shows up to 58% difference in accuracy between the original and perturbed cultural settings in low-resource cultures. Our dataset is publicly available at: https://huggingface.co/datasets/kyawyethu/MixCuBe.

Namgyu Ho, Sangmin Bae, Taehyeon Kim et al.

We introduce the Block Transformer which adopts hierarchical global-to-local modeling to autoregressive transformers to mitigate the inference bottlenecks associated with self-attention. Self-attention requires the key-value (KV) cache of all previous sequences to be retrieved from memory at every decoding step to retrieve context information, leading to two primary bottlenecks during batch inference. First, there is a significant delay in obtaining the first token, as the information of the entire prompt must first be processed to prefill the KV cache. Second, computation of subsequent tokens is bottlenecked by the high memory I/O demand of fetching the entire KV cache, which grows linearly with sequence length, incurring quadratic memory reads overall. We design the Block Transformer to strategically mitigate these costs, by incorporating coarsity and locality into an integrated global-to-local architecture. At the lower layers, we aggregate tokens into fixed size blocks to apply attention across the entire sequence at coarse-grained detail, to capture the global context while minimizing KV cache overhead. At upper layers, we apply attention within each block to decode individual tokens, to model fine-grained details with a lightweight local KV cache. We pretrain vanilla and Block Transformers from scratch and demonstrate that Block Transformers reach 10--20x inference throughput compared to vanilla transformers with equivalent perplexity and zero-shot task performance. Code is available at https://github.com/itsnamgyu/block-transformer.

Noah Lee, Na Min An, James Thorne

Large language models (LLMs) have shown impressive achievements in solving a broad range of tasks. Augmented by instruction fine-tuning, LLMs have also been shown to generalize in zero-shot settings as well. However, whether LLMs closely align with the human disagreement distribution has not been well-studied, especially within the scope of natural language inference (NLI). In this paper, we evaluate the performance and alignment of LLM distribution with humans using two different techniques to estimate the multinomial distribution: Monte Carlo Estimation (MCE) and Log Probability Estimation (LPE). As a result, we show LLMs exhibit limited ability in solving NLI tasks and simultaneously fail to capture human disagreement distribution. The inference and human alignment performances plunge even further on data samples with high human disagreement levels, raising concerns about their natural language understanding (NLU) ability and their representativeness to a larger human population. The source code for the experiments is available at https://github.com/xfactlab/emnlp2023-LLM-Disagreement

Fabio Petroni, Aleksandra Piktus, Angela Fan et al.

Challenging problems such as open-domain question answering, fact checking, slot filling and entity linking require access to large, external knowledge sources. While some models do well on individual tasks, developing general models is difficult as each task might require computationally expensive indexing of custom knowledge sources, in addition to dedicated infrastructure. To catalyze research on models that condition on specific information in large textual resources, we present a benchmark for knowledge-intensive language tasks (KILT). All tasks in KILT are grounded in the same snapshot of Wikipedia, reducing engineering turnaround through the re-use of components, as well as accelerating research into task-agnostic memory architectures. We test both task-specific and general baselines, evaluating downstream performance in addition to the ability of the models to provide provenance. We find that a shared dense vector index coupled with a seq2seq model is a strong baseline, outperforming more tailor-made approaches for fact checking, open-domain question answering and dialogue, and yielding competitive results on entity linking and slot filling, by generating disambiguated text. KILT data and code are available at https://github.com/facebookresearch/KILT.

Eunsu Kim, Juyoung Suk, Philhoon Oh et al.

Despite the rapid development of large language models (LLMs) for the Korean language, there remains an obvious lack of benchmark datasets that test the requisite Korean cultural and linguistic knowledge. Because many existing Korean benchmark datasets are derived from the English counterparts through translation, they often overlook the different cultural contexts. For the few benchmark datasets that are sourced from Korean data capturing cultural knowledge, only narrow tasks such as bias and hate speech detection are offered. To address this gap, we introduce a benchmark of Cultural and Linguistic Intelligence in Korean (CLIcK), a dataset comprising 1,995 QA pairs. CLIcK sources its data from official Korean exams and textbooks, partitioning the questions into eleven categories under the two main categories of language and culture. For each instance in CLIcK, we provide fine-grained annotation of which cultural and linguistic knowledge is required to answer the question correctly. Using CLIcK, we test 13 language models to assess their performance. Our evaluation uncovers insights into their performances across the categories, as well as the diverse factors affecting their comprehension. CLIcK offers the first large-scale comprehensive Korean-centric analysis of LLMs' proficiency in Korean culture and language.

Michael Schlichtkrull, Yulong Chen, Chenxi Whitehouse et al. · amazon-science

The Automated Verification of Textual Claims (AVeriTeC) shared task asks participants to retrieve evidence and predict veracity for real-world claims checked by fact-checkers. Evidence can be found either via a search engine, or via a knowledge store provided by the organisers. Submissions are evaluated using AVeriTeC score, which considers a claim to be accurately verified if and only if both the verdict is correct and retrieved evidence is considered to meet a certain quality threshold. The shared task received 21 submissions, 18 of which surpassed our baseline. The winning team was TUDA_MAI with an AVeriTeC score of 63%. In this paper we describe the shared task, present the full results, and highlight key takeaways from the shared task.

Guijin Son, Jiwoo Hong, Hyunwoo Ko et al.

Scaling pre-training compute has proven effective for achieving mulitlinguality, but does the same hold for test-time scaling? In this work, we introduce MCLM, a multilingual math benchmark featuring competition-level problems in 55 languages. We test three test-time scaling methods-Outcome Reward Modeling (ORM), Process Reward Modeling (ORM), and Budget Forcing (BF)-on both Qwen2.5-1.5B Math and MR1-1.5B, a multilingual LLM we trained for extended reasoning. Our experiments show that using Qwen2.5-1.5B Math with ORM achieves a score of 35.8 on MCLM, while BF on MR1-1.5B attains 35.2. Although "thinking LLMs" have recently garnered significant attention, we find that their performance is comparable to traditional scaling methods like best-of-N once constrained to similar levels of inference FLOPs. Moreover, while BF yields a 20-point improvement on English AIME, it provides only a 1.94-point average gain across other languages-a pattern consistent across the other test-time scaling methods we studied-higlighting that test-time scaling may not generalize as effectively to multilingual tasks. To foster further research, we release MCLM, MR1-1.5B, and evaluation results.

Minju Seo, Jinheon Baek, James Thorne et al.

Despite large successes of recent language models on diverse tasks, they suffer from severe performance degeneration in low-resource settings with limited training data available. Many existing works tackle this problem by generating synthetic data from the training data and then training models on them, recently using Large Language Models (LLMs). However, in low-resource settings, the amount of seed data samples to use for data augmentation is very small, which makes generated samples suboptimal and less diverse. To tackle this challenge, we propose a novel method that augments training data by incorporating a wealth of examples from other datasets, along with the given training data. Specifically, we first retrieve the relevant instances from other datasets, such as their input-output pairs or contexts, based on their similarities with the given seed data, and then prompt LLMs to generate new samples with the contextual information within and across the original and retrieved samples. This approach can ensure that the generated data is not only relevant but also more diverse than what could be achieved using the limited seed data alone. We validate our proposed Retrieval-Augmented Data Augmentation (RADA) framework on multiple datasets under low-resource settings of training and test-time data augmentation scenarios, on which it outperforms existing LLM-powered data augmentation baselines.

Jiwoo Hong, Noah Lee, Rodrigo Martínez-Castaño et al.

Reinforcement learning with human feedback (RLHF) is shown to largely benefit from precise reward models (RMs). However, recent studies in reward modeling schemes are skewed towards English, limiting the applicability of RLHF in multilingual alignments. In this work, we investigate the cross-lingual transfer of RMs trained in diverse languages, primarily from English. Our experimental results demonstrate the strong cross-lingual transfer of English RMs, exceeding target language RMs by 3~4% average increase in Multilingual RewardBench. Furthermore, we analyze the cross-lingual transfer of RMs through the representation shifts. Finally, we perform multilingual alignment to exemplify how cross-lingual transfer in RM propagates to enhanced multilingual instruction-following capability, along with extensive analyses on off-the-shelf RMs. We release the code, model, and data.

Zahra Bayramli, Ayhan Suleymanzade, Na Min An et al.

Text-to-image diffusion models have recently enabled the creation of visually compelling, detailed images from textual prompts. However, their ability to accurately represent various cultural nuances remains an open question. In our work, we introduce CultDiff benchmark, evaluating state-of-the-art diffusion models whether they can generate culturally specific images spanning ten countries. We show that these models often fail to generate cultural artifacts in architecture, clothing, and food, especially for underrepresented country regions, by conducting a fine-grained analysis of different similarity aspects, revealing significant disparities in cultural relevance, description fidelity, and realism compared to real-world reference images. With the collected human evaluations, we develop a neural-based image-image similarity metric, namely, CultDiff-S, to predict human judgment on real and generated images with cultural artifacts. Our work highlights the need for more inclusive generative AI systems and equitable dataset representation over a wide range of cultures.

Eunki Kim, Sangryul Kim, James Thorne

To enhance reasoning capabilities, previous works have explored incorporating special-purpose tokens into the training process. These strategies strengthen the learning mechanism of transformer-based large language models (LLMs). Building on prior research, in which inserting dummy tokens consecutively just before reasoning steps can enhance effectiveness, we introduce a novel approach termed Dynamic Inserting Tokens Training (DIT). Our method identifies positions within sequences where model confidence is lowest according to token log-likelihood. Strategically inserting [PAUSE] tokens on these positions bolsters the model's predictive capabilities for subsequent tokens. Experimental results across diverse datasets and models, from the 2.7B model to the 8B model, demonstrate that DIT consistently outperforms traditional fine-tuning and previous token insertion methods. With this simple yet effective method, we achieve accuracy gains of up to 4.7%p on GSM8K, 3.23%p on AQUA-RAT, and pass@1 improvements of up to 3.4%p on MBPP datasets. Our work shows a model-based, dynamic approach rather than a heuristic one, thereby broadening the scope of research in reasoning.

Sheikh Shafayat, H M Quamran Hasan, Minhajur Rahman Chowdhury Mahim et al. · allen-ai

In this study, we introduce BEnQA, a dataset comprising parallel Bengali and English exam questions for middle and high school levels in Bangladesh. Our dataset consists of approximately 5K questions covering several subjects in science with different types of questions, including factual, application, and reasoning-based questions. We benchmark several Large Language Models (LLMs) with our parallel dataset and observe a notable performance disparity between the models in Bengali and English. We also investigate some prompting methods, and find that Chain-of-Thought prompting is beneficial mostly on reasoning questions, but not so much on factual ones. We also find that appending English translation helps to answer questions in Bengali. Our findings point to promising future research directions for improving the performance of LLMs in Bengali and more generally in low-resource languages.

Na Min An, Eunki Kim, James Thorne et al.

Contrastive Language-Image Pretraining (CLIP) enables zero-shot inference in downstream tasks such as image-text retrieval and classification. However, recent works extending CLIP suffer from the issue of modality gap, which arises when the image and text embeddings are projected to disparate manifolds, deviating from the intended objective of image-text contrastive learning. We discover that this phenomenon is linked to the modality-specific characteristic that each image/text encoder independently possesses and propose two methods to address the modality gap: (1) a post-hoc embedding standardization method, $\text{I0T}_{\text{post}}$ that reduces the modality gap approximately to zero and (2) a trainable method, $\text{I0T}_{\text{async}}$, to alleviate the modality gap problem by adding two normalization layers for each encoder. Our I0T framework can significantly reduce the modality gap while preserving the original embedding representations of trained models with their locked parameters. In practice, $\text{I0T}_{\text{post}}$ can serve as an alternative explainable automatic evaluation metric of widely used CLIPScore (CLIP-S).

Wan Ju Kang, Eunki Kim, Na Min An et al.

Often, the needs and visual abilities differ between the annotator group and the end user group. Generating detailed diagram descriptions for blind and low-vision (BLV) users is one such challenging domain. Sighted annotators could describe visuals with ease, but existing studies have shown that direct generations by them are costly, bias-prone, and somewhat lacking by BLV standards. In this study, we ask sighted individuals to assess -- rather than produce -- diagram descriptions generated by vision-language models (VLM) that have been guided with latent supervision via a multi-pass inference. The sighted assessments prove effective and useful to professional educators who are themselves BLV and teach visually impaired learners. We release Sightation, a collection of diagram description datasets spanning 5k diagrams and 137k samples for completion, preference, retrieval, question answering, and reasoning training purposes and demonstrate their fine-tuning potential in various downstream tasks.

Na Min An, Eunki Kim, Wan Ju Kang et al.

For individuals with blindness or low vision (BLV), navigating complex environments can pose serious risks. Large Vision-Language Models (LVLMs) show promise for generating scene descriptions, but their effectiveness for BLV users remains underexplored. To address this gap, we conducted a user study with eight BLV participants to systematically evaluate preferences for six types of LVLM descriptions. While they helped to reduce fear and improve actionability, user ratings showed wide variation in sufficiency and conciseness. Furthermore, GPT-4o--despite its strong potential to refine descriptions--was not consistently preferred by participants. We use the insights obtained from the user study to build training data for building our new automatic evaluation metric that can capture BLV preferences effectively. Our findings underscore the urgent need for BLV-centered evaluation metrics and human-in-the-loop feedback to advance LVLM description quality for accessibility.

Sangryul Kim, James Thorne

Question Answering (QA) in NLP is the task of finding answers to a query within a relevant context retrieved by a retrieval system. Yet, the mix of relevant and irrelevant information in these contexts can hinder performance enhancements in QA tasks. To address this, we introduce a context filtering approach that removes non-essential details, summarizing crucial content through Reward Modeling. This method emphasizes keeping vital data while omitting the extraneous during summarization model training. We offer a framework for developing efficient QA models by discerning useful information from dataset pairs, bypassing the need for costly human evaluation. Furthermore, we show that our approach can significantly outperform the baseline, as evidenced by a 6.8-fold increase in the EM Per Token (EPT) metric, which we propose as a measure of token efficiency, indicating a notable token-efficiency boost for low-resource settings.

Eunki Kim, Na Min An, James Thorne et al.

Evaluating image-text alignment while reflecting human preferences across multiple aspects is a significant issue for the development of reliable vision-language applications. It becomes especially crucial in real-world scenarios where multiple valid descriptions exist depending on contexts or user needs. However, research progress is hindered by the lack of comprehensive benchmarks and existing evaluation predictors lacking at least one of these key properties: (1) Alignment with human judgments, (2) Long-sequence processing, (3) Inference efficiency, and (4) Applicability to multi-objective scoring. To address these challenges, we propose a plug-and-play architecture to build a robust predictor, MULTI-TAP (Multi-Objective Task-Aware Predictor), capable of both multi and single-objective scoring. MULTI-TAP can produce a single overall score, utilizing a reward head built on top of a large vision-language model (LVLMs). We show that MULTI-TAP is robust in terms of application to different LVLM architectures, achieving significantly higher performance than existing metrics and even on par with the GPT-4o-based predictor, G-VEval, with a smaller size (7-8B). By training a lightweight ridge regression layer on the frozen hidden states of a pre-trained LVLM, MULTI-TAP can produce fine-grained scores for multiple human-interpretable objectives. MULTI-TAP performs better than VisionREWARD, a high-performing multi-objective reward model, in both performance and efficiency on multi-objective benchmarks and our newly released text-image-to-text dataset, EYE4ALL. Our new dataset, consisting of chosen/rejected human preferences (EYE4ALLPref) and human-annotated fine-grained scores across seven dimensions (EYE4ALLMulti), can serve as a foundation for developing more accessible AI systems by capturing the underlying preferences of users, including blind and low-vision (BLV) individuals.

Minsu Kim, Sangryul Kim, James Thorne

This paper investigates the knowledge of language models from the perspective of Bayesian epistemology. We explore how language models adjust their confidence and responses when presented with evidence with varying levels of informativeness and reliability. To study these properties, we create a dataset with various types of evidence and analyze language models' responses and confidence using verbalized confidence, token probability, and sampling. We observed that language models do not consistently follow Bayesian epistemology: language models follow the Bayesian confirmation assumption well with true evidence but fail to adhere to other Bayesian assumptions when encountering different evidence types. Also, we demonstrated that language models can exhibit high confidence when given strong evidence, but this does not always guarantee high accuracy. Our analysis also reveals that language models are biased toward golden evidence and show varying performance depending on the degree of irrelevance, helping explain why they deviate from Bayesian assumptions.

Philhoon Oh, Jinwoo Shin, James Thorne

Recent advancements in Large Language Models (LLMs) underscore the necessity of Retrieval Augmented Generation (RAG) to leverage external information. However, LLMs are sensitive to the position of relevant information within contexts and tend to generate incorrect responses when such information is placed in the middle, known as `Lost in the Middle' phenomenon. In this paper, we introduce a framework that generates consistent outputs for decoder-only models, irrespective of the input context order. Experimental results for three open domain question answering tasks demonstrate position invariance, where the model is not sensitive to input context order, and superior robustness to irrelevent passages compared to prevailing approaches for RAG pipelines.

Jiwoo Hong, Sayak Paul, Noah Lee et al.

Modern preference alignment methods, such as DPO, rely on divergence regularization to a reference model for training stability-but this creates a fundamental problem we call "reference mismatch." In this paper, we investigate the negative impacts of reference mismatch in aligning text-to-image (T2I) diffusion models, showing that larger reference mismatch hinders effective adaptation given the same amount of data, e.g., as when learning new artistic styles, or personalizing to specific objects. We demonstrate this phenomenon across text-to-image (T2I) diffusion models and introduce margin-aware preference optimization (MaPO), a reference-agnostic approach that breaks free from this constraint. By directly optimizing the likelihood margin between preferred and dispreferred outputs under the Bradley-Terry model without anchoring to a reference, MaPO transforms diverse T2I tasks into unified pairwise preference optimization. We validate MaPO's versatility across five challenging domains: (1) safe generation, (2) style adaptation, (3) cultural representation, (4) personalization, and (5) general preference alignment. Our results reveal that MaPO's advantage grows dramatically with reference mismatch severity, outperforming both DPO and specialized methods like DreamBooth while reducing training time by 15%. MaPO thus emerges as a versatile and memory-efficient method for generic T2I adaptation tasks.

Minsu Kim, James Thorne

This paper investigates the inherent knowledge in language models from the perspective of epistemological holism. The purpose of this paper is to explore whether LLMs exhibit characteristics consistent with epistemological holism. These characteristics suggest that core knowledge, such as general scientific knowledge, each plays a specific role, serving as the foundation of our knowledge system and being difficult to revise. To assess these traits related to holism, we created a scientific reasoning dataset and examined the epistemology of language models through three tasks: Abduction, Revision, and Argument Generation. In the abduction task, the language models explained situations while avoiding revising the core knowledge. However, in other tasks, the language models were revealed not to distinguish between core and peripheral knowledge, showing an incomplete alignment with holistic knowledge principles.

EuiYul Song, Philhoon Oh, Sangryul Kim et al.

Modern deterministic retrieval pipelines prioritize achieving state-of-the-art performance but often lack interpretability in decision-making. These models face challenges in assessing uncertainty, leading to overconfident predictions. To overcome these limitations, we integrate uncertainty calibration and interpretability into a retrieval pipeline. Specifically, we introduce Bayesian methodologies and multi-perspective retrieval to calibrate uncertainty within a retrieval pipeline. We incorporate techniques such as LIME and SHAP to analyze the behavior of a black-box reranker model. The importance scores derived from these explanation methodologies serve as supplementary relevance scores to enhance the base reranker model. We evaluate the resulting performance enhancements achieved through uncertainty calibration and interpretable reranking on Question Answering and Fact Checking tasks. Our methods demonstrate substantial performance improvements across three KILT datasets.

Jiho Kim, Sungjin Park, Yeonsu Kwon et al.

In real world applications, knowledge graphs (KG) are widely used in various domains (e.g. medical applications and dialogue agents). However, for fact verification, KGs have not been adequately utilized as a knowledge source. KGs can be a valuable knowledge source in fact verification due to their reliability and broad applicability. A KG consists of nodes and edges which makes it clear how concepts are linked together, allowing machines to reason over chains of topics. However, there are many challenges in understanding how these machine-readable concepts map to information in text. To enable the community to better use KGs, we introduce a new dataset, FactKG: Fact Verification via Reasoning on Knowledge Graphs. It consists of 108k natural language claims with five types of reasoning: One-hop, Conjunction, Existence, Multi-hop, and Negation. Furthermore, FactKG contains various linguistic patterns, including colloquial style claims as well as written style claims to increase practicality. Lastly, we develop a baseline approach and analyze FactKG over these reasoning types. We believe FactKG can advance both reliability and practicality in KG-based fact verification.

Rami Aly, Zhijiang Guo, Michael Schlichtkrull et al.

Fact verification has attracted a lot of attention in the machine learning and natural language processing communities, as it is one of the key methods for detecting misinformation. Existing large-scale benchmarks for this task have focused mostly on textual sources, i.e. unstructured information, and thus ignored the wealth of information available in structured formats, such as tables. In this paper we introduce a novel dataset and benchmark, Fact Extraction and VERification Over Unstructured and Structured information (FEVEROUS), which consists of 87,026 verified claims. Each claim is annotated with evidence in the form of sentences and/or cells from tables in Wikipedia, as well as a label indicating whether this evidence supports, refutes, or does not provide enough information to reach a verdict. Furthermore, we detail our efforts to track and minimize the biases present in the dataset and could be exploited by models, e.g. being able to predict the label without using evidence. Finally, we develop a baseline for verifying claims against text and tables which predicts both the correct evidence and verdict for 18% of the claims.

James Thorne, Majid Yazdani, Marzieh Saeidi et al.

Neural models have shown impressive performance gains in answering queries from natural language text. However, existing works are unable to support database queries, such as "List/Count all female athletes who were born in 20th century", which require reasoning over sets of relevant facts with operations such as join, filtering and aggregation. We show that while state-of-the-art transformer models perform very well for small databases, they exhibit limitations in processing noisy data, numerical operations, and queries that aggregate facts. We propose a modular architecture to answer these database-style queries over multiple spans from text and aggregating these at scale. We evaluate the architecture using WikiNLDB, a novel dataset for exploring such queries. Our architecture scales to databases containing thousands of facts whereas contemporary models are limited by how many facts can be encoded. In direct comparison on small databases, our approach increases overall answer accuracy from 85% to 90%. On larger databases, our approach retains its accuracy whereas transformer baselines could not encode the context.

James Thorne, Andreas Vlachos

This paper introduces the task of factual error correction: performing edits to a claim so that the generated rewrite is better supported by evidence. This extends the well-studied task of fact verification by providing a mechanism to correct written texts that are refuted or only partially supported by evidence. We demonstrate that it is feasible to train factual error correction systems from existing fact checking datasets which only contain labeled claims accompanied by evidence, but not the correction. We achieve this by employing a two-stage distant supervision approach that incorporates evidence into masked claims when generating corrections. Our approach, based on the T5 transformer and using retrieved evidence, achieved better results than existing work which used a pointer copy network and gold evidence, producing accurate factual error corrections for 5x more instances in human evaluation and a .125 increase in SARI score. The evaluation is conducted on a dataset of 65,000 instances based on a recent fact verification shared task and we release it to enable further work on the task.

Max Glockner, Ieva Staliūnaitė, James Thorne et al.

Automated fact-checking systems verify claims against evidence to predict their veracity. In real-world scenarios, the retrieved evidence may not unambiguously support or refute the claim and yield conflicting but valid interpretations. Existing fact-checking datasets assume that the models developed with them predict a single veracity label for each claim, thus discouraging the handling of such ambiguity. To address this issue we present AmbiFC, a fact-checking dataset with 10k claims derived from real-world information needs. It contains fine-grained evidence annotations of 50k passages from 5k Wikipedia pages. We analyze the disagreements arising from ambiguity when comparing claims against evidence in AmbiFC, observing a strong correlation of annotator disagreement with linguistic phenomena such as underspecification and probabilistic reasoning. We develop models for predicting veracity handling this ambiguity via soft labels and find that a pipeline that learns the label distribution for sentence-level evidence selection and veracity prediction yields the best performance. We compare models trained on different subsets of AmbiFC and show that models trained on the ambiguous instances perform better when faced with the identified linguistic phenomena.

James Thorne, Andreas Vlachos

This paper introduces the task of factual error correction: performing edits to a claim so that the generated rewrite is better supported by evidence. This extends the well-studied task of fact verification by providing a mechanism to correct written texts that are refuted or only partially supported by evidence. We demonstrate that it is feasible to train factual error correction systems from existing fact checking datasets which only contain labeled claims accompanied by evidence, but not the correction. We achieve this by employing a two-stage distant supervision approach that incorporates evidence into masked claims when generating corrections. Our approach, based on the T5 transformer and using retrieved evidence, achieved better results than existing work which used a pointer copy network and gold evidence, producing accurate factual error corrections for 5x more instances in human evaluation and a .125 increase in SARI score. The evaluation is conducted on a dataset of 65,000 instances based on a recent fact verification shared task and we release it to enable further work on the task.

James Thorne, Majid Yazdani, Marzieh Saeidi et al.

In recent years, neural networks have shown impressive performance gains on long-standing AI problems, and in particular, answering queries from natural language text. These advances raise the question of whether they can be extended to a point where we can relax the fundamental assumption of database management, namely, that our data is represented as fields of a pre-defined schema. This paper presents a first step in answering that question. We describe NeuralDB, a database system with no pre-defined schema, in which updates and queries are given in natural language. We develop query processing techniques that build on the primitives offered by the state of the art Natural Language Processing methods. We begin by demonstrating that at the core, recent NLP transformers, powered by pre-trained language models, can answer select-project-join queries if they are given the exact set of relevant facts. However, they cannot scale to non-trivial databases and cannot perform aggregation queries. Based on these findings, we describe a NeuralDB architecture that runs multiple Neural SPJ operators in parallel, each with a set of database sentences that can produce one of the answers to the query. The result of these operators is fed to an aggregation operator if needed. We describe an algorithm that learns how to create the appropriate sets of facts to be fed into each of the Neural SPJ operators. Importantly, this algorithm can be trained by the Neural SPJ operator itself. We experimentally validate the accuracy of NeuralDB and its components, showing that we can answer queries over thousands of sentences with very high accuracy.

James Thorne, Andreas Vlachos

The biases present in training datasets have been shown to affect models for sentence pair classification tasks such as natural language inference (NLI) and fact verification. While fine-tuning models on additional data has been used to mitigate them, a common issue is that of catastrophic forgetting of the original training dataset. In this paper, we show that elastic weight consolidation (EWC) allows fine-tuning of models to mitigate biases while being less susceptible to catastrophic forgetting. In our evaluation on fact verification and NLI stress tests, we show that fine-tuning with EWC dominates standard fine-tuning, yielding models with lower levels of forgetting on the original (biased) dataset for equivalent gains in accuracy on the fine-tuning (unbiased) dataset.

James Thorne, Andreas Vlachos, Christos Christodoulopoulos et al.

The task of Natural Language Inference (NLI) is widely modeled as supervised sentence pair classification. While there has been a lot of work recently on generating explanations of the predictions of classifiers on a single piece of text, there have been no attempts to generate explanations of classifiers operating on pairs of sentences. In this paper, we show that it is possible to generate token-level explanations for NLI without the need for training data explicitly annotated for this purpose. We use a simple LSTM architecture and evaluate both LIME and Anchor explanations for this task. We compare these to a Multiple Instance Learning (MIL) method that uses thresholded attention make token-level predictions. The approach we present in this paper is a novel extension of zero-shot single-sentence tagging to sentence pairs for NLI. We conduct our experiments on the well-studied SNLI dataset that was recently augmented with manually annotation of the tokens that explain the entailment relation. We find that our white-box MIL-based method, while orders of magnitude faster, does not reach the same accuracy as the black-box methods.

James Thorne, Andreas Vlachos

This paper describes a baseline for the second iteration of the Fact Extraction and VERification shared task (FEVER2.0) which explores the resilience of systems through adversarial evaluation. We present a collection of simple adversarial attacks against systems that participated in the first FEVER shared task. FEVER modeled the assessment of truthfulness of written claims as a joint information retrieval and natural language inference task using evidence from Wikipedia. A large number of participants made use of deep neural networks in their submissions to the shared task. The extent as to whether such models understand language has been the subject of a number of recent investigations and discussion in literature. In this paper, we present a simple method of generating entailment-preserving and entailment-altering perturbations of instances by common patterns within the training data. We find that a number of systems are greatly affected with absolute losses in classification accuracy of up to $29\%$ on the newly perturbed instances. Using these newly generated instances, we construct a sample submission for the FEVER2.0 shared task. Addressing these types of attacks will aid in building more robust fact-checking models, as well as suggest directions to expand the datasets.

James Thorne, Andreas Vlachos, Oana Cocarascu et al.

We present the results of the first Fact Extraction and VERification (FEVER) Shared Task. The task challenged participants to classify whether human-written factoid claims could be Supported or Refuted using evidence retrieved from Wikipedia. We received entries from 23 competing teams, 19 of which scored higher than the previously published baseline. The best performing system achieved a FEVER score of 64.21%. In this paper, we present the results of the shared task and a summary of the systems, highlighting commonalities and innovations among participating systems.

James Thorne, Andreas Vlachos

The recently increased focus on misinformation has stimulated research in fact checking, the task of assessing the truthfulness of a claim. Research in automating this task has been conducted in a variety of disciplines including natural language processing, machine learning, knowledge representation, databases, and journalism. While there has been substantial progress, relevant papers and articles have been published in research communities that are often unaware of each other and use inconsistent terminology, thus impeding understanding and further progress. In this paper we survey automated fact checking research stemming from natural language processing and related disciplines, unifying the task formulations and methodologies across papers and authors. Furthermore, we highlight the use of evidence as an important distinguishing factor among them cutting across task formulations and methods. We conclude with proposing avenues for future NLP research on automated fact checking.

James Thorne, Andreas Vlachos, Christos Christodoulopoulos et al.

In this paper we introduce a new publicly available dataset for verification against textual sources, FEVER: Fact Extraction and VERification. It consists of 185,445 claims generated by altering sentences extracted from Wikipedia and subsequently verified without knowledge of the sentence they were derived from. The claims are classified as Supported, Refuted or NotEnoughInfo by annotators achieving 0.6841 in Fleiss $κ$. For the first two classes, the annotators also recorded the sentence(s) forming the necessary evidence for their judgment. To characterize the challenge of the dataset presented, we develop a pipeline approach and compare it to suitably designed oracles. The best accuracy we achieve on labeling a claim accompanied by the correct evidence is 31.87%, while if we ignore the evidence we achieve 50.91%. Thus we believe that FEVER is a challenging testbed that will help stimulate progress on claim verification against textual sources.