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.

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 or text encoder independently possesses.Herein, we propose two methods to address the modality gap: (1) a post-hoc embedding standardization method, I0T post that reduces the modality gap approximately to zero and (2) a trainable method, I0T 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, I0T post can serve as an alternative explainable automatic evaluation metric of widely used CLIPScore (CLIP-S).The code is available in https://github.com/xfactlab/I0T.

Zahra Bayramli, Ayhan Suleymanzade, Na Min An, Huzama Ahmad, Eunsu Kim, Junyeong Park, James Thorne, Alice Oh. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Wan Ju Kang, Eunki Kim, Na Min An, Sangryul Kim, Haemin Choi, Ki Hoon Kwak, James Thorne. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Scaling pre-training compute has proven effective for achieving multilinguality, 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 then compare three test-time scaling methods—Outcome Reward Modeling, Process Reward Modeling, and Budget Forcing. Our findings indicate that although “thinking LLMs” have recently garnered significant attention, their performance is comparable to traditional scaling methods like best-of-N once constrained to similar levels of inference FLOPs. More importantly, all tested methods fail to generalize robustly across languages, achieving only modest gains that are smaller than those observed in English, with no improvements in variance or consistency. To foster further research, we release MCLM and MR1-1.5B (a multilingual LLM with reasoning capabilities) and our evaluation results.

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.

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.

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 as one-way classifiers are prone to over-optimization, losing generalizability to unseen inputs. In this paper, we study the cause of over-optimization and its downstream effects on the RLHF procedure, highlighting the importance of robustness in RMs. First, we show that the excessive dispersion of hidden state norms is the main source of over-optimization. Correspondingly, we propose batch-wise sum-to-zero regularization (BSR) that enforces reward sum for each batch to be zero-centered, constraining the rewards with abnormally large magnitudes. We assess the impact of BSR in improving robustness in RMs through four scenarios of over-optimization, where BSR consistently manifests better robustness on unseen inputs. Then, 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, with reducing generation length by 40\% while adding a 7\% increase in win rate, further highlights that robustness in RMs induces robustness in RLHF training.

Minsu Kim, Sangryul Kim, James Thorne. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

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 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 correctly answer the question. 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 language and culture.

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 Multihead 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.

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 revisit SFT in the context of preference alignment, emphasizing that a minor penalty for the disfavored style is sufficient for preference alignment.Building on this foundation, we introduce a straightforward reference model-free monolithic odds ratio preference optimization algorithm, ORPO, eliminating the need 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 diverse sizes from 125M to 7B.Specifically, finetuning 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 including Llama-2 Chat and Zephyr with more than 7B and 13B parameters: achieving up to 12.20% on AlpacaEval 2.0 (Figure 1), and 7.32 in MT-Bench (Table 2).We release code 1 and model checkpoints 2 for Mistral-ORPO-α and Mistral-ORPO-β.

Recently, GPT-4 has become the de facto evaluator for long-form text generated by large language models (LLMs). However, for practitioners and researchers with large and custom evaluation tasks, GPT-4 is unreliable due to its 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. For this purpose, we construct a new dataset – FEEDBACK COLLECTION – that consists of 1K fine-grained score rubrics, 20K instructions, and 100K natural language feedback generated by GPT-4. Using the FEEDBACK COLLECTION, we train PROMETHEUS, a 13B evaluation-specific LLM that can assess any given response based on novel and unseen score rubrics and reference materials provided by the user. Our dataset’s versatility and diversity make our model generalize to challenging real-world criteria, such as prioritizing conciseness, child-readability, or varying levels of formality. We show that PROMETHEUS shows a stronger correlation with GPT-4 evaluation compared to ChatGPT on seven evaluation benchmarks (Two Feedback Collection testsets, MT Bench, Vicuna Bench, Flask Eval, MT Bench Human Judgment, and HHH Alignment), showing the efficacy of our model and dataset design. During human evaluation with hand-crafted score rubrics, PROMETHEUS shows a Pearson correlation of 0.897 with human evaluators, which is on par with GPT-4-0613 (0.882), and greatly outperforms ChatGPT (0.392). Remarkably, when assessing the quality of the generated feedback, PROMETHEUS demonstrates a win rate of 58.62% when compared to GPT-4 evaluation and a win rate of 79.57% when compared to ChatGPT evaluation. Our findings suggests that by adding reference materials and training on GPT-4 feedback, we can obtain effective open-source evaluator LMs.

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 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.

Abstract 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,1 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.

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.1

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. 1 * Equal contribution 1 The source code for the experiments is available at https://github.com/xfactlab/emnlp2023-LLM-Disagreement. ORead the following and determine if the hypothesis can be inferred from the premise.Premise: She smiled back.Hypothesis: She was so happy she couldn't stop smiling.

AI alignment refers to models acting towards human-intended goals, preferences, or ethical principles. 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. In order to evaluate AI-human visual alignment, a dataset should encompass samples with various scenarios 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 (e.g., 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.

Many recent Natural Language Processing (NLP) task formulations, such as question answering and fact verification, are implemented as a two-stage cascading architecture. In the first stage an IR system retrieves "relevant'' documents containing the knowledge, and in the second stage an NLP system performs reasoning to solve the task. Optimizing the IR system for retrieving relevant documents ensures that the NLP system has sufficient information to operate over. These recent NLP task formulations raise interesting and exciting challenges for IR, where the end-user of an IR system is not a human with an information need, but another system exploiting the documents retrieved by the IR system to perform reasoning and address the user information need. Among these challenges, as we will show, is that noise from the IR system, such as retrieving spurious or irrelevant documents, can negatively impact the accuracy of the downstream reasoning module. Hence, there is the need to balance maximizing relevance while minimizing noise in the IR system. This paper presents experimental results on two NLP tasks implemented as a two-stage cascading architecture. We show how spurious or irrelevant retrieved results from the first stage can induce errors in the second stage. We use these results to ground our discussion of the research challenges that the IR community should address in the context of these knowledge-intensive NLP tasks.

James Thorne, Andreas Vlachos. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

James Thorne, Majid Yazdani, Marzieh Saeidi, Fabrizio Silvestri, Sebastian Riedel, Alon Halevy. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

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 finetuning 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.

Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vladimir Karpukhin, Jean Maillard, Vassilis Plachouras, Tim Rocktäschel, Sebastian Riedel. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

In recent years, neural networks have shown impressive performance gains on long-standing AI problems, such as answering queries from text and machine translation. These advances raise the question of whether neural nets can be used at the core of query processing to derive answers from facts, even when the facts are expressed in natural language. If so, it is conceivable that we could relax the fundamental assumption of database management, namely, that our data is represented as fields of a pre-defined schema. Furthermore, such technology would enable combining information from text, images, and structured data seamlessly. This paper introduces neural databases , a class of systems that use NLP transformers as localized answer derivation engines. We ground the vision in NeuralDB, a system for querying facts represented as short natural language sentences. We demonstrate that recent natural language processing models, specifically transformers, can answer select-project-join queries if they are given a set of relevant facts. However, they cannot scale to non-trivial databases nor answer set-based and aggregation queries. Based on these insights, we identify specific research challenges that are needed to build neural databases. Some of the challenges require drawing upon the rich literature in data management, and others pose new research opportunities to the NLP community. Finally, we show that with preliminary solutions, NeuralDB can already answer queries over thousands of sentences with very high accuracy.

James Thorne, Andreas Vlachos, Christos Christodoulopoulos, Arpit Mittal. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

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, Arpit Mittal. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.