Keisuke Sakaguchi

Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras et al. · allen-ai, uw

Many language generation models are now available for a wide range of generation tasks, including machine translation and summarization. Combining such diverse models may lead to further progress, but ensembling generation models is challenging during inference: conventional ensembling methods (e.g., shallow fusion) require that the models share vocabulary/tokenization schemes. We introduce Twist decoding, a simple and general text generation algorithm that benefits from diverse models at inference time. Our method does not assume the vocabulary, tokenization or even generation order is shared. Our extensive evaluations on machine translation and scientific paper summarization demonstrate that Twist decoding substantially outperforms each model decoded in isolation over various scenarios, including cases where domain-specific and general-purpose models are both available. Twist decoding also consistently outperforms the popular reranking heuristic where output candidates from one model are rescored by another. We hope that our work will encourage researchers and practitioners to examine generation models collectively, not just independently, and to seek out models with complementary strengths to the currently available models. Our code is available at https://github.com/jungokasai/twist_decoding.

Jungo Kasai, Yuhei Kasai, Keisuke Sakaguchi et al.

As large language models (LLMs) gain popularity among speakers of diverse languages, we believe that it is crucial to benchmark them to better understand model behaviors, failures, and limitations in languages beyond English. In this work, we evaluate LLM APIs (ChatGPT, GPT-3, and GPT-4) on the Japanese national medical licensing examinations from the past five years, including the current year. Our team comprises native Japanese-speaking NLP researchers and a practicing cardiologist based in Japan. Our experiments show that GPT-4 outperforms ChatGPT and GPT-3 and passes all six years of the exams, highlighting LLMs' potential in a language that is typologically distant from English. However, our evaluation also exposes critical limitations of the current LLM APIs. First, LLMs sometimes select prohibited choices that should be strictly avoided in medical practice in Japan, such as suggesting euthanasia. Further, our analysis shows that the API costs are generally higher and the maximum context size is smaller for Japanese because of the way non-Latin scripts are currently tokenized in the pipeline. We release our benchmark as Igaku QA as well as all model outputs and exam metadata. We hope that our results and benchmark will spur progress on more diverse applications of LLMs. Our benchmark is available at https://github.com/jungokasai/IgakuQA.

Jungo Kasai, Keisuke Sakaguchi, Yoichi Takahashi et al. · allen-ai, uw

We introduce REALTIME QA, a dynamic question answering (QA) platform that announces questions and evaluates systems on a regular basis (weekly in this version). REALTIME QA inquires about the current world, and QA systems need to answer questions about novel events or information. It therefore challenges static, conventional assumptions in open-domain QA datasets and pursues instantaneous applications. We build strong baseline models upon large pretrained language models, including GPT-3 and T5. Our benchmark is an ongoing effort, and this paper presents real-time evaluation results over the past year. Our experimental results show that GPT-3 can often properly update its generation results, based on newly-retrieved documents, highlighting the importance of up-to-date information retrieval. Nonetheless, we find that GPT-3 tends to return outdated answers when retrieved documents do not provide sufficient information to find an answer. This suggests an important avenue for future research: can an open-domain QA system identify such unanswerable cases and communicate with the user or even the retrieval module to modify the retrieval results? We hope that REALTIME QA will spur progress in instantaneous applications of question answering and beyond.

Chandra Bhagavatula, Jena D. Hwang, Doug Downey et al. · allen-ai, uw

Commonsense capabilities of pre-trained language models dramatically improve with scale, leading many to believe that scale is the only winning recipe. But is it? Here, we investigate an alternative that a priori seems impossible: can smaller language models (e.g., GPT-2) win over models that are orders of magnitude larger and better (e.g., GPT-3), if powered with novel commonsense distillation algorithms? The key intellectual challenge is to design a learning algorithm that achieve a competitive level of commonsense acquisition, without relying on the benefits of scale. In particular, we study generative models of commonsense knowledge, focusing on the task of generating generics, statements of commonsense facts about everyday concepts, e.g., birds can fly. We introduce I2D2, a novel commonsense distillation framework that loosely follows the Symbolic Knowledge Distillation of West et al. but breaks the dependence on the extreme-scale teacher model with two innovations: (1) the novel adaptation of NeuroLogic Decoding to enhance the generation quality of the weak, off-the-shelf language models, and (2) self-imitation learning to iteratively learn from the model's own enhanced commonsense acquisition capabilities. Empirical results suggest that scale is not the only way, as novel algorithms can be a promising alternative. Moreover, our study leads to a new corpus of generics, Gen-A-tomic, that is the largest and highest quality available to date.

LLM-jp, Akiko Aizawa, Eiji Aramaki et al.

This paper introduces LLM-jp, a cross-organizational project for the research and development of Japanese large language models (LLMs). LLM-jp aims to develop open-source and strong Japanese LLMs, and as of this writing, more than 1,500 participants from academia and industry are working together for this purpose. This paper presents the background of the establishment of LLM-jp, summaries of its activities, and technical reports on the LLMs developed by LLM-jp. For the latest activities, visit https://llm-jp.nii.ac.jp/en/.

Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras et al. · allen-ai, uw

Text generation with beam search has proven successful in a wide range of applications. We point out that, though largely overlooked in the literature, the commonly-used implementation of beam decoding (e.g., Hugging Face Transformers and fairseq) uses a first come, first served heuristic: it keeps a set of already completed sequences over time steps and stops when the size of this set reaches the beam size. Based on this finding, we introduce a patience factor, a simple modification to this beam decoding implementation, that generalizes the stopping criterion and provides flexibility to the depth of search. Empirical results demonstrate that adjusting this patience factor improves decoding performance of strong pretrained models on news text summarization and machine translation over diverse language pairs, with a negligible inference slowdown. Our approach only modifies one line of code and can be thus readily incorporated in any implementation. Further, we find that different versions of beam decoding result in large performance differences in summarization, demonstrating the need for clarity in specifying the beam search implementation in research work. Our code will be available upon publication.

Mutsumi Sasaki, Go kamoda, Ryosuke Takahashi et al.

Quantities with measurement units, such as 110 cm and 1.2 m, require language models (LMs) to combine a numeral with a symbolic unit scale. Here, we study how LMs compare such quantities in controlled settings spanning several unit systems. We find that accuracy degrades near the comparison boundary, where small changes in value determine the correct answer. The resulting errors are systematic: linear surrogate models predict LM preferences from numerical-difference and unit-scale-difference cues, and causal interventions on subspaces aligned with these variables shift model's output. The results suggest that LMs compare quantities through a bag of heuristics over numerals and units, rather than first converting both expressions to an exact shared-scale representation.

Michael Regan, Jena D. Hwang, Keisuke Sakaguchi et al. · uw

Making sense of familiar yet new situations typically involves making generalizations about causal schemas, stories that help humans reason about event sequences. Reasoning about events includes identifying cause and effect relations shared across event instances, a process we refer to as causal schema induction. Statistical schema induction systems may leverage structural knowledge encoded in discourse or the causal graphs associated with event meaning, however resources to study such causal structure are few in number and limited in size. In this work, we investigate how to apply schema induction models to the task of knowledge discovery for enhanced search of English-language news texts. To tackle the problem of data scarcity, we present Torquestra, a manually curated dataset of text-graph-schema units integrating temporal, event, and causal structures. We benchmark our dataset on three knowledge discovery tasks, building and evaluating models for each. Results show that systems that harness causal structure are effective at identifying texts sharing similar causal meaning components rather than relying on lexical cues alone. We make our dataset and models available for research purposes.

Masato Mita, Keisuke Sakaguchi, Masato Hagiwara et al.

Natural language processing technology has rapidly improved automated grammatical error correction tasks, and the community begins to explore document-level revision as one of the next challenges. To go beyond sentence-level automated grammatical error correction to NLP-based document-level revision assistant, there are two major obstacles: (1) there are few public corpora with document-level revisions being annotated by professional editors, and (2) it is not feasible to elicit all possible references and evaluate the quality of revision with such references because there are infinite possibilities of revision. This paper tackles these challenges. First, we introduce a new document-revision corpus, TETRA, where professional editors revised academic papers sampled from the ACL anthology which contain few trivial grammatical errors that enable us to focus more on document- and paragraph-level edits such as coherence and consistency. Second, we explore reference-less and interpretable methods for meta-evaluation that can detect quality improvements by document revision. We show the uniqueness of TETRA compared with existing document revision corpora and demonstrate that a fine-tuned pre-trained language model can discriminate the quality of documents after revision even when the difference is subtle. This promising result will encourage the community to further explore automated document revision models and metrics in future.

Steven Coyne, Keisuke Sakaguchi, Diana Galvan-Sosa et al.

GPT-3 and GPT-4 models are powerful, achieving high performance on a variety of Natural Language Processing tasks. However, there is a relative lack of detailed published analysis of their performance on the task of grammatical error correction (GEC). To address this, we perform experiments testing the capabilities of a GPT-3.5 model (text-davinci-003) and a GPT-4 model (gpt-4-0314) on major GEC benchmarks. We compare the performance of different prompts in both zero-shot and few-shot settings, analyzing intriguing or problematic outputs encountered with different prompt formats. We report the performance of our best prompt on the BEA-2019 and JFLEG datasets, finding that the GPT models can perform well in a sentence-level revision setting, with GPT-4 achieving a new high score on the JFLEG benchmark. Through human evaluation experiments, we compare the GPT models' corrections to source, human reference, and baseline GEC system sentences and observe differences in editing strategies and how they are scored by human raters.

Keito Kudo, Yoichi Aoki, Tatsuki Kuribayashi et al.

Compositionality is a pivotal property of symbolic reasoning. However, how well recent neural models capture compositionality remains underexplored in the symbolic reasoning tasks. This study empirically addresses this question by systematically examining recently published pre-trained seq2seq models with a carefully controlled dataset of multi-hop arithmetic symbolic reasoning. We introduce a skill tree on compositionality in arithmetic symbolic reasoning that defines the hierarchical levels of complexity along with three compositionality dimensions: systematicity, productivity, and substitutivity. Our experiments revealed that among the three types of composition, the models struggled most with systematicity, performing poorly even with relatively simple compositions. That difficulty was not resolved even after training the models with intermediate reasoning steps.

Shabnam Behzad, Keisuke Sakaguchi, Nathan Schneider et al.

We present ELQA, a corpus of questions and answers in and about the English language. Collected from two online forums, the >70k questions (from English learners and others) cover wide-ranging topics including grammar, meaning, fluency, and etymology. The answers include descriptions of general properties of English vocabulary and grammar as well as explanations about specific (correct and incorrect) usage examples. Unlike most NLP datasets, this corpus is metalinguistic -- it consists of language about language. As such, it can facilitate investigations of the metalinguistic capabilities of NLU models, as well as educational applications in the language learning domain. To study this, we define a free-form question answering task on our dataset and conduct evaluations on multiple LLMs (Large Language Models) to analyze their capacity to generate metalinguistic answers.

Yoichi Aoki, Keito Kudo, Tatsuki Kuribayashi et al.

Neural reasoning accuracy improves when generating intermediate reasoning steps. However, the source of this improvement is yet unclear. Here, we investigate and factorize the benefit of generating intermediate steps for symbolic reasoning. Specifically, we decompose the reasoning strategy w.r.t. step granularity and chaining strategy. With a purely symbolic numerical reasoning dataset (e.g., A=1, B=3, C=A+3, C?), we found that the choice of reasoning strategies significantly affects the performance, with the gap becoming even larger as the extrapolation length becomes longer. Surprisingly, we also found that certain configurations lead to nearly perfect performance, even in the case of length extrapolation. Our results indicate the importance of further exploring effective strategies for neural reasoning models.

Go Kamoda, Benjamin Heinzerling, Keisuke Sakaguchi et al.

Factual probing is a method that uses prompts to test if a language model "knows" certain world knowledge facts. A problem in factual probing is that small changes to the prompt can lead to large changes in model output. Previous work aimed to alleviate this problem by optimizing prompts via text mining or fine-tuning. However, such approaches are relation-specific and do not generalize to unseen relation types. Here, we propose to use test-time augmentation (TTA) as a relation-agnostic method for reducing sensitivity to prompt variations by automatically augmenting and ensembling prompts at test time. Experiments show improved model calibration, i.e., with TTA, model confidence better reflects prediction accuracy. Improvements in prediction accuracy are observed for some models, but for other models, TTA leads to degradation. Error analysis identifies the difficulty of producing high-quality prompt variations as the main challenge for TTA.

YiFeng Wang, Zhun Sun, Keisuke Sakaguchi

We present Activation Residual Hessian Quantization (ARHQ), a post-training weight splitting method designed to mitigate error propagation in low-bit activation-weight quantization. By constructing an input-side residual Hessian from activation quantization residuals (G_x), ARHQ analytically identifies and isolates error-sensitive weight directions into a high-precision low-rank branch. This is achieved via a closed-form truncated SVD on the scaled weight matrix W G^{1/2}_x . Experimental results on Qwen3-4B-Thinking-2507 demonstrate that ARHQ significantly improves layer-wise SNR and preserves downstream reasoning performance on ZebraLogic even under aggressive quantization. The code is available at https://github.com/BeautMoonQ/ARHQ.

Subaru Kimura, Ryota Tanaka, Shumpei Miyawaki et al.

We explore visual prompt injection (VPI) that maliciously exploits the ability of large vision-language models (LVLMs) to follow instructions drawn onto the input image. We propose a new VPI method, "goal hijacking via visual prompt injection" (GHVPI), that swaps the execution task of LVLMs from an original task to an alternative task designated by an attacker. The quantitative analysis indicates that GPT-4V is vulnerable to the GHVPI and demonstrates a notable attack success rate of 15.8%, which is an unignorable security risk. Our analysis also shows that successful GHVPI requires high character recognition capability and instruction-following ability in LVLMs.

Yoshinari Fujinuma, Keisuke Sakaguchi

Masked diffusion language models (dLMs) generate text through bidirectional denoising, yet this capability remains locked for infilling prompts. This limitation is an artifact of the current supervised finetuning (SFT) convention of applying response-only masking. To unlock this capability, we extend full-sequence masking during SFT, where both prompts and responses are masked jointly. Once unlocked, the model infills masked portions of a prompt template conditioned on few-shot examples. We show that such model-infilled prompts match or surpass manually designed templates, transfer effectively across models, and are complementary to existing prompt optimization methods. Our results suggest that training practices, not architectural limitations, are the primary bottleneck preventing masked diffusion language models from infilling effective prompts

Diana Galvan-Sosa, Gabrielle Gaudeau, Pride Kavumba et al.

The performance and usability of Large-Language Models (LLMs) are driving their use in explanation generation tasks. However, despite their widespread adoption, LLM explanations have been found to be unreliable, making it difficult for users to distinguish good from bad explanations. To address this issue, we present Rubrik's CUBE, an education-inspired rubric and a dataset of 26k explanations, written and later quality-annotated using the rubric by both humans and six open- and closed-source LLMs. The CUBE dataset focuses on two reasoning and two language tasks, providing the necessary diversity for us to effectively test our proposed rubric. Using Rubrik, we find that explanations are influenced by both task and perceived difficulty. Low quality stems primarily from a lack of conciseness in LLM-generated explanations, rather than cohesion and word choice. The full dataset, rubric, and code are available at https://github.com/RubriksCube/rubriks_cube.

Ana Brassard, Benjamin Heinzerling, Keito Kudo et al.

Evaluating the quality of free-text explanations is a multifaceted, subjective, and labor-intensive task. Large language models (LLMs) present an appealing alternative due to their potential for consistency, scalability, and cost-efficiency. In this work, we present ACORN, a new dataset of 3,500 free-text explanations and aspect-wise quality ratings, and use it to evaluate how LLMs rate explanations. We observed that larger models outputted labels that maintained or increased the inter-annotator agreement, suggesting that they are within the expected variance between human raters. However, their correlation with majority-voted human ratings varied across different quality aspects, indicating that they are not a complete replacement. In turn, using LLMs as a supplement to a smaller group of human raters in some cases improved the correlation with the original majority labels. However, the effect was limited to cases where human raters were scarce, and an additional human rater had a more pronounced effect in all cases. Overall, we recommend against using LLMs as a complete replacement for human raters but encourage using them in configurations that end with targeted human involvement. Data available here: https://github.com/a-brassard/ACORN

Haruto Yoshida, Keito Kudo, Yoichi Aoki et al.

Large vision-language models (LVLMs) demonstrate strong performance on diagram understanding benchmarks, yet they still struggle with understanding relationships between elements, particularly those represented by nodes and directed edges (e.g., arrows and lines). To investigate the underlying causes of this limitation, we probe the internal representation of LVLMs using a carefully constructed synthetic diagram dataset based on directed graphs. Our probing experiments reveal that edge information is not linearly separable in the vision encoder and becomes linearly encoded only in the text tokens in the language model. In contrast, node information and global structural features are already linearly encoded in individual hidden states of the vision encoder. These findings suggest that the stage at which linearly separable representations are formed varies depending on the type of visual information. In particular, the delayed emergence of edge representations may help explain why LVLMs struggle with relational understanding, such as interpreting edge directions, which require more abstract, compositionally integrated processes.

Niket Tandon, Aman Madaan, Peter Clark et al.

How can an end-user provide feedback if a deployed structured prediction model generates inconsistent output, ignoring the structural complexity of human language? This is an emerging topic with recent progress in synthetic or constrained settings, and the next big leap would require testing and tuning models in real-world settings. We present a new dataset, Interscript, containing user feedback on a deployed model that generates complex everyday tasks. Interscript contains 8,466 data points -- the input is a possibly erroneous script and a user feedback, and the output is a modified script. We posit two use-cases of \ours that might significantly advance the state-of-the-art in interactive learning. The dataset is available at: https://github.com/allenai/interscript.

Masato Hagiwara, Joshua Tanner, Keisuke Sakaguchi

We present GrammarTagger, an open-source grammar profiler which, given an input text, identifies grammatical features useful for language education. The model architecture enables it to learn from a small amount of texts annotated with spans and their labels, which 1) enables easier and more intuitive annotation, 2) supports overlapping spans, and 3) is less prone to error propagation, compared to complex hand-crafted rules defined on constituency/dependency parses. We show that we can bootstrap a grammar profiler model with $F_1 \approx 0.6$ from only a couple hundred sentences both in English and Chinese, which can be further boosted via learning a multilingual model. With GrammarTagger, we also build Octanove Learn, a search engine of language learning materials indexed by their reading difficulty and grammatical features. The code and pretrained models are publicly available at \url{https://github.com/octanove/grammartagger}.

Jungmin Choi, Keisuke Sakaguchi, Hiroaki Yamada

Large language models (LLMs) have shown strong performance on legal benchmarks, including multiple-choice components of bar exams. However, their capacity for generating open-ended legal reasoning in realistic scenarios remains insufficiently explored. Notably, to our best knowledge, there are no prior studies or datasets addressing this issue in the Japanese context. This study presents the first dataset designed to evaluate the open-ended legal reasoning performance of LLMs within the Japanese jurisdiction. The dataset is based on the writing component of the Japanese bar examination, which requires examinees to identify multiple legal issues from long narratives and to construct structured legal arguments in free text format. Our key contribution is the manual evaluation of LLMs' generated responses by legal experts, which reveals limitations and challenges in legal reasoning. Moreover, we conducted a manual analysis of hallucinations to characterize when and how the models introduce content not supported by precedent or law. Our real exam questions, model-generated responses, and expert evaluations reveal the milestones of current LLMs in the Japanese legal domain. Our dataset and relevant resources will be available online.

Go Kamoda, Benjamin Heinzerling, Tatsuro Inaba et al.

According to the stages-of-inference hypothesis, early layers of language models map their subword-tokenized input, which does not necessarily correspond to a linguistically meaningful segmentation, to more meaningful representations that form the model's "inner vocabulary". Prior analysis of this detokenization stage has predominantly relied on probing and interventions such as path patching, which involve selecting particular inputs, choosing a subset of components that will be patched, and then observing changes in model behavior. Here, we show that several important aspects of the detokenization stage can be understood purely by analyzing model weights, without performing any model inference steps. Specifically, we introduce an analytical decomposition of first-layer attention in GPT-2. Our decomposition yields interpretable terms that quantify the relative contributions of position-related, token-related, and mixed effects. By focusing on terms in this decomposition, we discover weight-based explanations of attention bias toward close tokens and attention for detokenization.

Kosei Kobayashi, Kosuke Matsuzaki, Masaya Taniguchi et al.

The long-range dependencies among the tokens, which originate from hierarchical structures, are a defining hallmark of human language. However, whether similar dependencies exist within the sequential vocalization of non-human animals remains a topic of investigation. Transformer architectures, known for their ability to model long-range dependencies among tokens, provide a powerful tool for investigating this phenomenon. In this study, we employed the Transformer architecture to analyze the songs of Bengalese finch (Lonchura striata domestica), which are characterized by their highly variable and complex syllable sequences. To this end, we developed FinchGPT, a Transformer-based model trained on a textualized corpus of birdsongs, which outperformed other architecture models in this domain. Attention weight analysis revealed that FinchGPT effectively captures long-range dependencies within syllables sequences. Furthermore, reverse engineering approaches demonstrated the impact of computational and biological manipulations on its performance: restricting FinchGPT's attention span and disrupting birdsong syntax through the ablation of specific brain nuclei markedly influenced the model's outputs. Our study highlights the transformative potential of large language models (LLMs) in deciphering the complexities of animal vocalizations, offering a novel framework for exploring the structural properties of non-human communication systems while shedding light on the computational distinctions between biological brains and artificial neural networks.

Jaehyeok Lee, Keisuke Sakaguchi, JinYeong Bak

Self-training approach for large language models (LLMs) improves reasoning abilities by training the models on their self-generated rationales. Previous approaches have labeled rationales that produce correct answers for a given question as appropriate for training. However, a single measure risks misjudging rationale quality, leading the models to learn flawed reasoning patterns. To address this issue, we propose CREST (Consistency-driven Rationale Evaluation for Self-Training), a self-training framework that further evaluates each rationale through follow-up questions and leverages this evaluation to guide its training. Specifically, we introduce two methods: (1) filtering out rationales that frequently result in incorrect answers on follow-up questions and (2) preference learning based on mixed preferences from rationale evaluation results of both original and follow-up questions. Experiments on three question-answering datasets using open LLMs show that CREST not only improves the logical robustness and correctness of rationales but also improves reasoning abilities compared to previous self-training approaches.

Kosuke Matsuzaki, Masaya Taniguchi, Kentaro Inui et al.

We introduce a Japanese Morphology dataset, J-UniMorph, developed based on the UniMorph feature schema. This dataset addresses the unique and rich verb forms characteristic of the language's agglutinative nature. J-UniMorph distinguishes itself from the existing Japanese subset of UniMorph, which is automatically extracted from Wiktionary. On average, the Wiktionary Edition features around 12 inflected forms for each word and is primarily dominated by denominal verbs (i.e., [noun] +suru (do-PRS)). Morphologically, this form is equivalent to the verb suru (do). In contrast, J-UniMorph explores a much broader and more frequently used range of verb forms, offering 118 inflected forms for each word on average. It includes honorifics, a range of politeness levels, and other linguistic nuances, emphasizing the distinctive characteristics of the Japanese language. This paper presents detailed statistics and characteristics of J-UniMorph, comparing it with the Wiktionary Edition. We release J-UniMorph and its interactive visualizer publicly available, aiming to support cross-linguistic research and various applications.

Tarek Naous, Anagha Savit, Carlos Rafael Catalan et al.

As Large Language Models (LLMs) gain stronger multilingual capabilities, their ability to handle culturally diverse entities becomes crucial. Prior work has shown that LLMs often favor Western-associated entities in Arabic, raising concerns about cultural fairness. Due to the lack of multilingual benchmarks, it remains unclear if such biases also manifest in different non-Western languages. In this paper, we introduce Camellia, a benchmark for measuring entity-centric cultural biases in nine Asian languages spanning six distinct Asian cultures. Camellia includes 19,530 entities manually annotated for association with the specific Asian or Western culture, as well as 2,173 naturally occurring masked contexts for entities derived from social media posts. Using Camellia, we evaluate cultural biases in four recent multilingual LLM families across various tasks such as cultural context adaptation, sentiment association, and entity extractive QA. Our analyses show a struggle by LLMs at cultural adaptation in all Asian languages, with performance differing across models developed in regions with varying access to culturally-relevant data. We further observe that different LLM families hold their distinct biases, differing in how they associate cultures with particular sentiments. Lastly, we find that LLMs struggle with context understanding in Asian languages, creating performance gaps between cultures in entity extraction.

Mutsumi Sasaki, Go Kamoda, Ryosuke Takahashi et al.

Temporal reasoning and knowledge are essential capabilities for language models (LMs). While much prior work has analyzed and improved temporal reasoning in LMs, most studies have focused solely on the Gregorian calendar. However, many non-Gregorian systems, such as the Japanese, Hijri, and Hebrew calendars, are in active use and reflect culturally grounded conceptions of time. If and how well current LMs can accurately handle such non-Gregorian calendars has not been evaluated so far. Here, we present a systematic evaluation of how well language models handle one such non-Gregorian system: the Japanese wareki. We create datasets that require temporal knowledge and reasoning in using wareki dates. Evaluating open and closed LMs, we find that some models can perform calendar conversions, but GPT-4o, Deepseek V3, and even Japanese-centric models struggle with Japanese calendar arithmetic and knowledge involving wareki dates. Error analysis suggests corpus frequency of Japanese calendar expressions and a Gregorian bias in the model's knowledge as possible explanations. Our results show the importance of developing LMs that are better equipped for culture-specific tasks such as calendar understanding.

Steven Coyne, Diana Galvan-Sosa, Ryan Spring et al.

Recent advances in natural language processing (NLP) have contributed to the development of automated writing evaluation (AWE) systems that can correct grammatical errors. However, while these systems are effective at improving text, they are not optimally designed for language learning. They favor direct revisions, often with a click-to-fix functionality that can be applied without considering the reason for the correction. Meanwhile, depending on the error type, learners may benefit most from simple explanations and strategically indirect hints, especially on generalizable grammatical rules. To support the generation of such feedback, we introduce an annotation framework that models each error's error type and generalizability. For error type classification, we introduce a typology focused on inferring learners' knowledge gaps by connecting their errors to specific grammatical patterns. Following this framework, we collect a dataset of annotated learner errors and corresponding human-written feedback comments, each labeled as a direct correction or hint. With this data, we evaluate keyword-guided, keyword-free, and template-guided methods of generating feedback using large language models (LLMs). Human teachers examined each system's outputs, assessing them on grounds including relevance, factuality, and comprehensibility. We report on the development of the dataset and the comparative performance of the systems investigated.

Keito Kudo, Yoichi Aoki, Tatsuki Kuribayashi et al.

This study investigates the incremental, internal problem-solving process of language models (LMs) with arithmetic multi-hop reasoning as a case study. We specifically investigate when LMs internally resolve sub/whole problems through first reading the problem statements, generating reasoning chains, and achieving the final answer to mechanistically interpret LMs' multi-hop problem-solving process. Our experiments reveal a systematic incremental reasoning strategy underlying LMs. They have not derived an answer at the moment they first read the problem; instead, they obtain (sub)answers while generating the reasoning chain. Therefore, the generated reasoning chains can be regarded as faithful reflections of the model's internal computation.

Yoichi Aoki, Keito Kudo, Tatsuki Kuribayashi et al.

Multi-step reasoning instruction, such as chain-of-thought prompting, is widely adopted to explore better language models (LMs) performance. We report on the systematic strategy that LMs employ in such a multi-step reasoning process. Our controlled experiments reveal that LMs rely more heavily on heuristics, such as lexical overlap, in the earlier stages of reasoning, where more reasoning steps remain to reach a goal. Conversely, their reliance on heuristics decreases as LMs progress closer to the final answer through multiple reasoning steps. This suggests that LMs can backtrack only a limited number of future steps and dynamically combine heuristic strategies with rationale ones in tasks involving multi-step reasoning.

Ryosuke Takahashi, Go Kamoda, Benjamin Heinzerling et al.

Language models (LMs) encode world knowledge in their internal parameters through training. However, LMs may learn personal and confidential information from the training data, leading to privacy concerns such as data leakage. Therefore, research on knowledge deletion from LMs is essential. This study focuses on the knowledge stored in LMs and analyzes the relationship between the side effects of knowledge deletion and the entities related to the knowledge. Our findings reveal that deleting knowledge related to popular entities can have catastrophic side effects. Furthermore, this research is the first to analyze knowledge deletion in models trained on synthetic knowledge graphs, indicating a new direction for controlled experiments.

Faeze Brahman, Chandra Bhagavatula, Valentina Pyatkin et al.

Procedural planning, which entails decomposing a high-level goal into a sequence of temporally ordered steps, is an important yet intricate task for machines. It involves integrating common-sense knowledge to reason about complex and often contextualized situations, e.g. ``scheduling a doctor's appointment without a phone''. While current approaches show encouraging results using large language models (LLMs), they are hindered by drawbacks such as costly API calls and reproducibility issues. In this paper, we advocate planning using smaller language models. We present PlaSma, a novel two-pronged approach to endow small language models with procedural knowledge and (constrained) language planning capabilities. More concretely, we develop symbolic procedural knowledge distillation to enhance the commonsense knowledge in small language models and an inference-time algorithm to facilitate more structured and accurate reasoning. In addition, we introduce a new related task, Replanning, that requires a revision of a plan to cope with a constrained situation. In both the planning and replanning settings, we show that orders-of-magnitude smaller models (770M-11B parameters) can compete and often surpass their larger teacher models' capabilities. Finally, we showcase successful application of PlaSma in an embodied environment, VirtualHome.

Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras et al.

Natural language processing researchers have identified limitations of evaluation methodology for generation tasks, with new questions raised about the validity of automatic metrics and of crowdworker judgments. Meanwhile, efforts to improve generation models tend to depend on simple n-gram overlap metrics (e.g., BLEU, ROUGE). We argue that new advances on models and metrics should each more directly benefit and inform the other. We therefore propose a generalization of leaderboards, bidimensional leaderboards (Billboards), that simultaneously tracks progress in language generation models and metrics for their evaluation. Unlike conventional unidimensional leaderboards that sort submitted systems by predetermined metrics, a Billboard accepts both generators and evaluation metrics as competing entries. A Billboard automatically creates an ensemble metric that selects and linearly combines a few metrics based on a global analysis across generators. Further, metrics are ranked based on their correlation with human judgments. We release four Billboards for machine translation, summarization, and image captioning. We demonstrate that a linear ensemble of a few diverse metrics sometimes substantially outperforms existing metrics in isolation. Our mixed-effects model analysis shows that most automatic metrics, especially the reference-based ones, overrate machine over human generation, demonstrating the importance of updating metrics as generation models become stronger (and perhaps more similar to humans) in the future. Our project website is available at https://nlp.cs.washington.edu/billboard/.

Jungo Kasai, Keisuke Sakaguchi, Lavinia Dunagan et al.

We establish THumB, a rubric-based human evaluation protocol for image captioning models. Our scoring rubrics and their definitions are carefully developed based on machine- and human-generated captions on the MSCOCO dataset. Each caption is evaluated along two main dimensions in a tradeoff (precision and recall) as well as other aspects that measure the text quality (fluency, conciseness, and inclusive language). Our evaluations demonstrate several critical problems of the current evaluation practice. Human-generated captions show substantially higher quality than machine-generated ones, especially in coverage of salient information (i.e., recall), while most automatic metrics say the opposite. Our rubric-based results reveal that CLIPScore, a recent metric that uses image features, better correlates with human judgments than conventional text-only metrics because it is more sensitive to recall. We hope that this work will promote a more transparent evaluation protocol for image captioning and its automatic metrics.

Liwei Jiang, Jena D. Hwang, Chandra Bhagavatula et al.

As AI systems become increasingly powerful and pervasive, there are growing concerns about machines' morality or a lack thereof. Yet, teaching morality to machines is a formidable task, as morality remains among the most intensely debated questions in humanity, let alone for AI. Existing AI systems deployed to millions of users, however, are already making decisions loaded with moral implications, which poses a seemingly impossible challenge: teaching machines moral sense, while humanity continues to grapple with it. To explore this challenge, we introduce Delphi, an experimental framework based on deep neural networks trained directly to reason about descriptive ethical judgments, e.g., "helping a friend" is generally good, while "helping a friend spread fake news" is not. Empirical results shed novel insights on the promises and limits of machine ethics; Delphi demonstrates strong generalization capabilities in the face of novel ethical situations, while off-the-shelf neural network models exhibit markedly poor judgment including unjust biases, confirming the need for explicitly teaching machines moral sense. Yet, Delphi is not perfect, exhibiting susceptibility to pervasive biases and inconsistencies. Despite that, we demonstrate positive use cases of imperfect Delphi, including using it as a component model within other imperfect AI systems. Importantly, we interpret the operationalization of Delphi in light of prominent ethical theories, which leads us to important future research questions.

Aman Madaan, Niket Tandon, Dheeraj Rajagopal et al.

A class of explainable NLP models for reasoning tasks support their decisions by generating free-form or structured explanations, but what happens when these supporting structures contain errors? Our goal is to allow users to interactively correct explanation structures through natural language feedback. We introduce MERCURIE - an interactive system that refines its explanations for a given reasoning task by getting human feedback in natural language. Our approach generates graphs that have 40% fewer inconsistencies as compared with the off-the-shelf system. Further, simply appending the corrected explanation structures to the output leads to a gain of 1.2 points on accuracy on defeasible reasoning across all three domains. We release a dataset of over 450k graphs for defeasible reasoning generated by our system at https://tinyurl.com/mercurie .

Keisuke Sakaguchi, Chandra Bhagavatula, Ronan Le Bras et al.

Scripts - standardized event sequences describing typical everyday activities - have been shown to help understand narratives by providing expectations, resolving ambiguity, and filling in unstated information. However, to date they have proved hard to author or extract from text. In this work, we demonstrate for the first time that pre-trained neural language models (LMs) can be be finetuned to generate high-quality scripts, at varying levels of granularity, for a wide range of everyday scenarios (e.g., bake a cake). To do this, we collected a large (6.4k), crowdsourced partially ordered scripts (named proScript), which is substantially larger than prior datasets, and developed models that generate scripts with combining language generation and structure prediction. We define two complementary tasks: (i) edge prediction: given a scenario and unordered events, organize the events into a valid (possibly partial-order) script, and (ii) script generation: given only a scenario, generate events and organize them into a (possibly partial-order) script. Our experiments show that our models perform well (e.g., F1=75.7 in task (i)), illustrating a new approach to overcoming previous barriers to script collection. We also show that there is still significant room for improvement toward human level performance. Together, our tasks, dataset, and models offer a new research direction for learning script knowledge.

Niket Tandon, Keisuke Sakaguchi, Bhavana Dalvi Mishra et al.

We present the first dataset for tracking state changes in procedural text from arbitrary domains by using an unrestricted (open) vocabulary. For example, in a text describing fog removal using potatoes, a car window may transition between being foggy, sticky,opaque, and clear. Previous formulations of this task provide the text and entities involved,and ask how those entities change for just a small, pre-defined set of attributes (e.g., location), limiting their fidelity. Our solution is a new task formulation where given just a procedural text as input, the task is to generate a set of state change tuples(entity, at-tribute, before-state, after-state)for each step,where the entity, attribute, and state values must be predicted from an open vocabulary. Using crowdsourcing, we create OPENPI1, a high-quality (91.5% coverage as judged by humans and completely vetted), and large-scale dataset comprising 29,928 state changes over 4,050 sentences from 810 procedural real-world paragraphs from WikiHow.com. A current state-of-the-art generation model on this task achieves 16.1% F1 based on BLEU metric, leaving enough room for novel model architectures.

Jena D. Hwang, Chandra Bhagavatula, Ronan Le Bras et al.

Recent years have brought about a renewed interest in commonsense representation and reasoning in the field of natural language understanding. The development of new commonsense knowledge graphs (CSKG) has been central to these advances as their diverse facts can be used and referenced by machine learning models for tackling new and challenging tasks. At the same time, there remain questions about the quality and coverage of these resources due to the massive scale required to comprehensively encompass general commonsense knowledge. In this work, we posit that manually constructed CSKGs will never achieve the coverage necessary to be applicable in all situations encountered by NLP agents. Therefore, we propose a new evaluation framework for testing the utility of KGs based on how effectively implicit knowledge representations can be learned from them. With this new goal, we propose ATOMIC 2020, a new CSKG of general-purpose commonsense knowledge containing knowledge that is not readily available in pretrained language models. We evaluate its properties in comparison with other leading CSKGs, performing the first large-scale pairwise study of commonsense knowledge resources. Next, we show that ATOMIC 2020 is better suited for training knowledge models that can generate accurate, representative knowledge for new, unseen entities and events. Finally, through human evaluation, we show that the few-shot performance of GPT-3 (175B parameters), while impressive, remains ~12 absolute points lower than a BART-based knowledge model trained on ATOMIC 2020 despite using over 430x fewer parameters.

Aaron Steven White, Elias Stengel-Eskin, Siddharth Vashishtha et al.

We present the Universal Decompositional Semantics (UDS) dataset (v1.0), which is bundled with the Decomp toolkit (v0.1). UDS1.0 unifies five high-quality, decompositional semantics-aligned annotation sets within a single semantic graph specification---with graph structures defined by the predicative patterns produced by the PredPatt tool and real-valued node and edge attributes constructed using sophisticated normalization procedures. The Decomp toolkit provides a suite of Python 3 tools for querying UDS graphs using SPARQL. Both UDS1.0 and Decomp0.1 are publicly available at http://decomp.io.

Niket Tandon, Bhavana Dalvi Mishra, Keisuke Sakaguchi et al.

We introduce WIQA, the first large-scale dataset of "What if..." questions over procedural text. WIQA contains three parts: a collection of paragraphs each describing a process, e.g., beach erosion; a set of crowdsourced influence graphs for each paragraph, describing how one change affects another; and a large (40k) collection of "What if...?" multiple-choice questions derived from the graphs. For example, given a paragraph about beach erosion, would stormy weather result in more or less erosion (or have no effect)? The task is to answer the questions, given their associated paragraph. WIQA contains three kinds of questions: perturbations to steps mentioned in the paragraph; external (out-of-paragraph) perturbations requiring commonsense knowledge; and irrelevant (no effect) perturbations. We find that state-of-the-art models achieve 73.8% accuracy, well below the human performance of 96.3%. We analyze the challenges, in particular tracking chains of influences, and present the dataset as an open challenge to the community.

Tongfei Chen, Zhengping Jiang, Adam Poliak et al.

We introduce Uncertain Natural Language Inference (UNLI), a refinement of Natural Language Inference (NLI) that shifts away from categorical labels, targeting instead the direct prediction of subjective probability assessments. We demonstrate the feasibility of collecting annotations for UNLI by relabeling a portion of the SNLI dataset under a probabilistic scale, where items even with the same categorical label differ in how likely people judge them to be true given a premise. We describe a direct scalar regression modeling approach, and find that existing categorically labeled NLI data can be used in pre-training. Our best models approach human performance, demonstrating models may be capable of more subtle inferences than the categorical bin assignment employed in current NLI tasks.

Chandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya et al.

Abductive reasoning is inference to the most plausible explanation. For example, if Jenny finds her house in a mess when she returns from work, and remembers that she left a window open, she can hypothesize that a thief broke into her house and caused the mess, as the most plausible explanation. While abduction has long been considered to be at the core of how people interpret and read between the lines in natural language (Hobbs et al., 1988), there has been relatively little research in support of abductive natural language inference and generation. We present the first study that investigates the viability of language-based abductive reasoning. We introduce a challenge dataset, ART, that consists of over 20k commonsense narrative contexts and 200k explanations. Based on this dataset, we conceptualize two new tasks -- (i) Abductive NLI: a multiple-choice question answering task for choosing the more likely explanation, and (ii) Abductive NLG: a conditional generation task for explaining given observations in natural language. On Abductive NLI, the best model achieves 68.9% accuracy, well below human performance of 91.4%. On Abductive NLG, the current best language generators struggle even more, as they lack reasoning capabilities that are trivial for humans. Our analysis leads to new insights into the types of reasoning that deep pre-trained language models fail to perform--despite their strong performance on the related but more narrowly defined task of entailment NLI--pointing to interesting avenues for future research.

Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula et al.

The Winograd Schema Challenge (WSC) (Levesque, Davis, and Morgenstern 2011), a benchmark for commonsense reasoning, is a set of 273 expert-crafted pronoun resolution problems originally designed to be unsolvable for statistical models that rely on selectional preferences or word associations. However, recent advances in neural language models have already reached around 90% accuracy on variants of WSC. This raises an important question whether these models have truly acquired robust commonsense capabilities or whether they rely on spurious biases in the datasets that lead to an overestimation of the true capabilities of machine commonsense. To investigate this question, we introduce WinoGrande, a large-scale dataset of 44k problems, inspired by the original WSC design, but adjusted to improve both the scale and the hardness of the dataset. The key steps of the dataset construction consist of (1) a carefully designed crowdsourcing procedure, followed by (2) systematic bias reduction using a novel AfLite algorithm that generalizes human-detectable word associations to machine-detectable embedding associations. The best state-of-the-art methods on WinoGrande achieve 59.4-79.1%, which are 15-35% below human performance of 94.0%, depending on the amount of the training data allowed. Furthermore, we establish new state-of-the-art results on five related benchmarks - WSC (90.1%), DPR (93.1%), COPA (90.6%), KnowRef (85.6%), and Winogender (97.1%). These results have dual implications: on one hand, they demonstrate the effectiveness of WinoGrande when used as a resource for transfer learning. On the other hand, they raise a concern that we are likely to be overestimating the true capabilities of machine commonsense across all these benchmarks. We emphasize the importance of algorithmic bias reduction in existing and future benchmarks to mitigate such overestimation.

Keisuke Sakaguchi, Benjamin Van Durme

We describe a novel method for efficiently eliciting scalar annotations for dataset construction and system quality estimation by human judgments. We contrast direct assessment (annotators assign scores to items directly), online pairwise ranking aggregation (scores derive from annotator comparison of items), and a hybrid approach (EASL: Efficient Annotation of Scalar Labels) proposed here. Our proposal leads to increased correlation with ground truth, at far greater annotator efficiency, suggesting this strategy as an improved mechanism for dataset creation and manual system evaluation.

Keisuke Sakaguchi, Matt Post, Benjamin Van Durme

We propose a neural encoder-decoder model with reinforcement learning (NRL) for grammatical error correction (GEC). Unlike conventional maximum likelihood estimation (MLE), the model directly optimizes towards an objective that considers a sentence-level, task-specific evaluation metric, avoiding the exposure bias issue in MLE. We demonstrate that NRL outperforms MLE both in human and automated evaluation metrics, achieving the state-of-the-art on a fluency-oriented GEC corpus.

Courtney Napoles, Keisuke Sakaguchi, Joel Tetreault

We present a new parallel corpus, JHU FLuency-Extended GUG corpus (JFLEG) for developing and evaluating grammatical error correction (GEC). Unlike other corpora, it represents a broad range of language proficiency levels and uses holistic fluency edits to not only correct grammatical errors but also make the original text more native sounding. We describe the types of corrections made and benchmark four leading GEC systems on this corpus, identifying specific areas in which they do well and how they can improve. JFLEG fulfills the need for a new gold standard to properly assess the current state of GEC.

Courtney Napoles, Keisuke Sakaguchi, Joel Tetreault

Current methods for automatically evaluating grammatical error correction (GEC) systems rely on gold-standard references. However, these methods suffer from penalizing grammatical edits that are correct but not in the gold standard. We show that reference-less grammaticality metrics correlate very strongly with human judgments and are competitive with the leading reference-based evaluation metrics. By interpolating both methods, we achieve state-of-the-art correlation with human judgments. Finally, we show that GEC metrics are much more reliable when they are calculated at the sentence level instead of the corpus level. We have set up a CodaLab site for benchmarking GEC output using a common dataset and different evaluation metrics.