Masahiro Kaneko

Ryuto Koike, Masahiro Kaneko, Naoaki Okazaki

Large Language Models (LLMs) have achieved human-level fluency in text generation, making it difficult to distinguish between human-written and LLM-generated texts. This poses a growing risk of misuse of LLMs and demands the development of detectors to identify LLM-generated texts. However, existing detectors lack robustness against attacks: they degrade detection accuracy by simply paraphrasing LLM-generated texts. Furthermore, a malicious user might attempt to deliberately evade the detectors based on detection results, but this has not been assumed in previous studies. In this paper, we propose OUTFOX, a framework that improves the robustness of LLM-generated-text detectors by allowing both the detector and the attacker to consider each other's output. In this framework, the attacker uses the detector's prediction labels as examples for in-context learning and adversarially generates essays that are harder to detect, while the detector uses the adversarially generated essays as examples for in-context learning to learn to detect essays from a strong attacker. Experiments in the domain of student essays show that the proposed detector improves the detection performance on the attacker-generated texts by up to +41.3 points F1-score. Furthermore, the proposed detector shows a state-of-the-art detection performance: up to 96.9 points F1-score, beating existing detectors on non-attacked texts. Finally, the proposed attacker drastically degrades the performance of detectors by up to -57.0 points F1-score, massively outperforming the baseline paraphrasing method for evading detection.

Masahiro Kaneko, Aizhan Imankulova, Danushka Bollegala et al.

Masked Language Models (MLMs) pre-trained by predicting masked tokens on large corpora have been used successfully in natural language processing tasks for a variety of languages. Unfortunately, it was reported that MLMs also learn discriminative biases regarding attributes such as gender and race. Because most studies have focused on MLMs in English, the bias of MLMs in other languages has rarely been investigated. Manual annotation of evaluation data for languages other than English has been challenging due to the cost and difficulty in recruiting annotators. Moreover, the existing bias evaluation methods require the stereotypical sentence pairs consisting of the same context with attribute words (e.g. He/She is a nurse). We propose Multilingual Bias Evaluation (MBE) score, to evaluate bias in various languages using only English attribute word lists and parallel corpora between the target language and English without requiring manually annotated data. We evaluated MLMs in eight languages using the MBE and confirmed that gender-related biases are encoded in MLMs for all those languages. We manually created datasets for gender bias in Japanese and Russian to evaluate the validity of the MBE. The results show that the bias scores reported by the MBE significantly correlates with that computed from the above manually created datasets and the existing English datasets for gender bias.

Masahiro Kaneko, Danushka Bollegala, Naoaki Okazaki

We study the relationship between task-agnostic intrinsic and task-specific extrinsic social bias evaluation measures for Masked Language Models (MLMs), and find that there exists only a weak correlation between these two types of evaluation measures. Moreover, we find that MLMs debiased using different methods still re-learn social biases during fine-tuning on downstream tasks. We identify the social biases in both training instances as well as their assigned labels as reasons for the discrepancy between intrinsic and extrinsic bias evaluation measurements. Overall, our findings highlight the limitations of existing MLM bias evaluation measures and raise concerns on the deployment of MLMs in downstream applications using those measures.

Masahiro Kaneko, Sho Takase, Ayana Niwa et al.

Grammatical Error Correction (GEC) should not focus only on high accuracy of corrections but also on interpretability for language learning. However, existing neural-based GEC models mainly aim at improving accuracy, and their interpretability has not been explored. A promising approach for improving interpretability is an example-based method, which uses similar retrieved examples to generate corrections. In addition, examples are beneficial in language learning, helping learners understand the basis of grammatically incorrect/correct texts and improve their confidence in writing. Therefore, we hypothesize that incorporating an example-based method into GEC can improve interpretability as well as support language learners. In this study, we introduce an Example-Based GEC (EB-GEC) that presents examples to language learners as a basis for a correction result. The examples consist of pairs of correct and incorrect sentences similar to a given input and its predicted correction. Experiments demonstrate that the examples presented by EB-GEC help language learners decide to accept or refuse suggestions from the GEC output. Furthermore, the experiments also show that retrieved examples improve the accuracy of corrections.

Panatchakorn Anantaprayoon, Masahiro Kaneko, Naoaki Okazaki

Discriminatory gender biases have been found in Pre-trained Language Models (PLMs) for multiple languages. In Natural Language Inference (NLI), existing bias evaluation methods have focused on the prediction results of one specific label out of three labels, such as neutral. However, such evaluation methods can be inaccurate since unique biased inferences are associated with unique prediction labels. Addressing this limitation, we propose a bias evaluation method for PLMs, called NLI-CoAL, which considers all the three labels of NLI task. First, we create three evaluation data groups that represent different types of biases. Then, we define a bias measure based on the corresponding label output of each data group. In the experiments, we introduce a meta-evaluation technique for NLI bias measures and use it to confirm that our bias measure can distinguish biased, incorrect inferences from non-biased incorrect inferences better than the baseline, resulting in a more accurate bias evaluation. We create the datasets in English, Japanese, and Chinese, and successfully validate the compatibility of our bias measure across multiple languages. Lastly, we observe the bias tendencies in PLMs of different languages. To our knowledge, we are the first to construct evaluation datasets and measure PLMs' bias from NLI in Japanese and Chinese.

Masahiro Kaneko, Danushka Bollegala, Naoaki Okazaki

Numerous types of social biases have been identified in pre-trained language models (PLMs), and various intrinsic bias evaluation measures have been proposed for quantifying those social biases. Prior works have relied on human annotated examples to compare existing intrinsic bias evaluation measures. However, this approach is not easily adaptable to different languages nor amenable to large scale evaluations due to the costs and difficulties when recruiting human annotators. To overcome this limitation, we propose a method to compare intrinsic gender bias evaluation measures without relying on human-annotated examples. Specifically, we create multiple bias-controlled versions of PLMs using varying amounts of male vs. female gendered sentences, mined automatically from an unannotated corpus using gender-related word lists. Next, each bias-controlled PLM is evaluated using an intrinsic bias evaluation measure, and the rank correlation between the computed bias scores and the gender proportions used to fine-tune the PLMs is computed. Experiments on multiple corpora and PLMs repeatedly show that the correlations reported by our proposed method that does not require human annotated examples are comparable to those computed using human annotated examples in prior work.

Ryuto Koike, Masahiro Kaneko, Naoaki Okazaki

To combat the misuse of Large Language Models (LLMs), many recent studies have presented LLM-generated-text detectors with promising performance. When users instruct LLMs to generate texts, the instruction can include different constraints depending on the user's need. However, most recent studies do not cover such diverse instruction patterns when creating datasets for LLM detection. In this paper, we reveal that even task-oriented constraints -- constraints that would naturally be included in an instruction and are not related to detection-evasion -- cause existing powerful detectors to have a large variance in detection performance. We focus on student essay writing as a realistic domain and manually create task-oriented constraints based on several factors for essay quality. Our experiments show that the standard deviation (SD) of current detector performance on texts generated by an instruction with such a constraint is significantly larger (up to an SD of 14.4 F1-score) than that by generating texts multiple times or paraphrasing the instruction. We also observe an overall trend where the constraints can make LLM detection more challenging than without them. Finally, our analysis indicates that the high instruction-following ability of LLMs fosters the large impact of such constraints on detection performance.

Masahiro Kaneko, Danushka Bollegala, Naoaki Okazaki

Pre-trained language models trained on large-scale data have learned serious levels of social biases. Consequently, various methods have been proposed to debias pre-trained models. Debiasing methods need to mitigate only discriminatory bias information from the pre-trained models, while retaining information that is useful for the downstream tasks. In previous research, whether useful information is retained has been confirmed by the performance of downstream tasks in debiased pre-trained models. On the other hand, it is not clear whether these benchmarks consist of data pertaining to social biases and are appropriate for investigating the impact of debiasing. For example in gender-related social biases, data containing female words (e.g. ``she, female, woman''), male words (e.g. ``he, male, man''), and stereotypical words (e.g. ``nurse, doctor, professor'') are considered to be the most affected by debiasing. If there is not much data containing these words in a benchmark dataset for a target task, there is the possibility of erroneously evaluating the effects of debiasing. In this study, we compare the impact of debiasing on performance across multiple downstream tasks using a wide-range of benchmark datasets that containing female, male, and stereotypical words. Experiments show that the effects of debiasing are consistently \emph{underestimated} across all tasks. Moreover, the effects of debiasing could be reliably evaluated by separately considering instances containing female, male, and stereotypical words than all of the instances in a benchmark dataset.

Masahiro Kaneko, Naoaki Okazaki

In Grammatical Error Correction (GEC), it is crucial to ensure the user's comprehension of a reason for correction. Existing studies present tokens, examples, and hints as to the basis for correction but do not directly explain the reasons for corrections. Although methods that use Large Language Models (LLMs) to provide direct explanations in natural language have been proposed for various tasks, no such method exists for GEC. Generating explanations for GEC corrections involves aligning input and output tokens, identifying correction points, and presenting corresponding explanations consistently. However, it is not straightforward to specify a complex format to generate explanations, because explicit control of generation is difficult with prompts. This study introduces a method called controlled generation with Prompt Insertion (PI) so that LLMs can explain the reasons for corrections in natural language. In PI, LLMs first correct the input text, and then we automatically extract the correction points based on the rules. The extracted correction points are sequentially inserted into the LLM's explanation output as prompts, guiding the LLMs to generate explanations for the correction points. We also create an Explainable GEC (XGEC) dataset of correction reasons by annotating NUCLE, CoNLL2013, and CoNLL2014. Although generations from GPT-3 and ChatGPT using original prompts miss some correction points, the generation control using PI can explicitly guide to describe explanations for all correction points, contributing to improved performance in generating correction reasons.

Masahiro Kaneko, Danushka Bollegala, Naoaki Okazaki

Different methods have been proposed to develop meta-embeddings from a given set of source embeddings. However, the source embeddings can contain unfair gender-related biases, and how these influence the meta-embeddings has not been studied yet. We study the gender bias in meta-embeddings created under three different settings: (1) meta-embedding multiple sources without performing any debiasing (Multi-Source No-Debiasing), (2) meta-embedding multiple sources debiased by a single method (Multi-Source Single-Debiasing), and (3) meta-embedding a single source debiased by different methods (Single-Source Multi-Debiasing). Our experimental results show that meta-embedding amplifies the gender biases compared to input source embeddings. We find that debiasing not only the sources but also their meta-embedding is needed to mitigate those biases. Moreover, we propose a novel debiasing method based on meta-embedding learning where we use multiple debiasing methods on a single source embedding and then create a single unbiased meta-embedding.

Rem Hida, Masahiro Kaneko, Naoaki Okazaki

Warning: This paper contains examples of stereotypes and biases. Large Language Models (LLMs) exhibit considerable social biases, and various studies have tried to evaluate and mitigate these biases accurately. Previous studies use downstream tasks as prompts to examine the degree of social biases for evaluation and mitigation. While LLMs' output highly depends on prompts, previous studies evaluating and mitigating bias have often relied on a limited variety of prompts. In this paper, we investigate the sensitivity of LLMs when changing prompt variations (task instruction and prompt, few-shot examples, debias-prompt) by analyzing task performance and social bias of LLMs. Our experimental results reveal that LLMs are highly sensitive to prompts to the extent that the ranking of LLMs fluctuates when comparing models for task performance and social bias. Additionally, we show that LLMs have tradeoffs between performance and social bias caused by the prompts. Less bias from prompt setting may result in reduced performance. Moreover, the ambiguity of instances is one of the reasons for this sensitivity to prompts in advanced LLMs, leading to various outputs. We recommend using diverse prompts, as in this study, to compare the effects of prompts on social bias in LLMs.

Mengsay Loem, Masahiro Kaneko, Naoaki Okazaki

Large Language Models (LLMs) can justify or critique their predictions through discussions with other models or humans, thereby enriching their intrinsic understanding of instances. While proactive discussions in the inference phase have been shown to boost performance, such interactions have not been extensively explored during the training phase. We hypothesize that incorporating interactive discussions into the training process can enhance the models' understanding and improve their reasoning and verbal expression abilities during inference. This work introduces the SAIE framework, which facilitates supportive and adversarial discussions between learner and partner models. The learner model receives responses from the partner, and its parameters are then updated based on this discussion. This dynamic adjustment process continues throughout the training phase, responding to the evolving outputs of the learner model. Our empirical evaluation across various tasks, including math problems, commonsense reasoning, and multi-domain knowledge, demonstrates that models fine-tuned with the SAIE framework outperform those trained with conventional fine-tuning approaches. Furthermore, our method enhances the models' reasoning capabilities, improving both individual and multi-agent inference performance.

Yi Zhou, Masahiro Kaneko, Danushka Bollegala

Sense embedding learning methods learn different embeddings for the different senses of an ambiguous word. One sense of an ambiguous word might be socially biased while its other senses remain unbiased. In comparison to the numerous prior work evaluating the social biases in pretrained word embeddings, the biases in sense embeddings have been relatively understudied. We create a benchmark dataset for evaluating the social biases in sense embeddings and propose novel sense-specific bias evaluation measures. We conduct an extensive evaluation of multiple static and contextualised sense embeddings for various types of social biases using the proposed measures. Our experimental results show that even in cases where no biases are found at word-level, there still exist worrying levels of social biases at sense-level, which are often ignored by the word-level bias evaluation measures.

Daisuke Oba, Masahiro Kaneko, Danushka Bollegala

Despite their impressive performance in a wide range of NLP tasks, Large Language Models (LLMs) have been reported to encode worrying-levels of gender biases. Prior work has proposed debiasing methods that require human labelled examples, data augmentation and fine-tuning of LLMs, which are computationally costly. Moreover, one might not even have access to the model parameters for performing debiasing such as in the case of closed LLMs such as GPT-4. To address this challenge, we propose bias suppression that prevents biased generations of LLMs by simply providing textual preambles constructed from manually designed templates and real-world statistics, without accessing to model parameters. We show that, using CrowsPairs dataset, our textual preambles covering counterfactual statements can suppress gender biases in English LLMs such as LLaMA2. Moreover, we find that gender-neutral descriptions of gender-biased objects can also suppress their gender biases. Moreover, we show that bias suppression has acceptable adverse effect on downstream task performance with HellaSwag and COPA.

Mengsay Loem, Sho Takase, Masahiro Kaneko et al.

Impressive performance of Transformer has been attributed to self-attention, where dependencies between entire input in a sequence are considered at every position. In this work, we reform the neural $n$-gram model, which focuses on only several surrounding representations of each position, with the multi-head mechanism as in Vaswani et al.(2017). Through experiments on sequence-to-sequence tasks, we show that replacing self-attention in Transformer with multi-head neural $n$-gram can achieve comparable or better performance than Transformer. From various analyses on our proposed method, we find that multi-head neural $n$-gram is complementary to self-attention, and their combinations can further improve performance of vanilla Transformer.

Masahiro Kaneko, Ayana Niwa, Timothy Baldwin

Fake news undermines societal trust and decision-making across politics, economics, health, and international relations, and in extreme cases threatens human lives and societal safety. Because fake news reflects region-specific political, social, and cultural contexts and is expressed in language, evaluating the risks of large language models (LLMs) requires a multi-lingual and regional perspective. Malicious users can bypass safeguards through jailbreak attacks, inducing LLMs to generate fake news. However, no benchmark currently exists to systematically assess attack resilience across languages and regions. Here, we propose JailNewsBench, the first benchmark for evaluating LLM robustness against jailbreak-induced fake news generation. JailNewsBench spans 34 regions and 22 languages, covering 8 evaluation sub-metrics through LLM-as-a-Judge and 5 jailbreak attacks, with approximately 300k instances. Our evaluation of 9 LLMs reveals that the maximum attack success rate (ASR) reached 86.3% and the maximum harmfulness score was 3.5 out of 5. Notably, for English and U.S.-related topics, the defensive performance of typical multi-lingual LLMs was significantly lower than for other regions, highlighting substantial imbalances in safety across languages and regions. In addition, our analysis shows that coverage of fake news in existing safety datasets is limited and less well defended than major categories such as toxicity and social bias. Our dataset and code are available at https://github.com/kanekomasahiro/jail_news_bench.

Harry Stuart, Masahiro Kaneko, Timothy Baldwin

Effective hiring is integral to the success of an organisation, but it is very challenging to find the most suitable candidates because expert evaluation (e.g.\ interviews conducted by a technical manager) are expensive to deploy at scale. Therefore, automated resume scoring and other applicant-screening methods are increasingly used to coarsely filter candidates, making decisions on limited information. We propose that large language models (LLMs) can play the role of subject matter experts to cost-effectively elicit information from each candidate that is nuanced and role-specific, thereby improving the quality of early-stage hiring decisions. We present a system that leverages an LLM interviewer to update belief over an applicant's rubric-oriented latent traits in a calibrated way. We evaluate our system on simulated interviews and show that belief converges towards the simulated applicants' artificially-constructed latent ability levels. We release code, a modest dataset of public-domain/anonymised resumes, belief calibration tests, and simulated interviews, at \href{https://github.com/mbzuai-nlp/beyond-the-resume}{https://github.com/mbzuai-nlp/beyond-the-resume}. Our demo is available at \href{https://btr.hstu.net}{https://btr.hstu.net}.

Koshiro Saito, Ryuto Koike, Masahiro Kaneko et al.

Detecting machine-generated text is essential for transparency and accountability when deploying large language models (LLMs). Among detection approaches, watermarking is a statistically reliable method by design -- it embeds detectable signals into LLM outputs by biasing their token distributions. However, it has been reported that watermarked LLMs often perform worse on downstream tasks. We propose PUPPET, a framework that fine-tunes an LLM via reinforcement learning to generate text that is both more detectable and better performing on downstream tasks. We use two reward functions: a detector that outputs a machine-class likelihood and an evaluator that measures a task-specific metric. Experiments on long-form QA, summarization, and essay writing show that LLMs trained with PUPPET achieve high detectability competitive with watermarking methods while outperforming them on downstream tasks. The analysis shows that this optimization can be performed efficiently with only a few thousand samples in 1--2 GPU hours. Moreover, these gains are consistent across out-of-domain tasks, different LLM families, and model sizes, and are even robust to paraphrasing attacks.

Masanari Oi, Mahiro Ukai, Masahiro Kaneko et al.

Direct preference optimization (DPO) has emerged as a promising approach for aligning large language models (LLMs) with human preferences. However, the widespread reliance on the response-level Bradley-Terry (BT) model may limit its full potential, as the reference and learnable models are assumed to be autoregressive only after deriving the objective function. Motivated by this limitation, we revisit the theoretical foundations of DPO and propose a novel formulation that explicitly introduces the autoregressive assumption prior to applying the BT model. By reformulating and extending DPO, we derive a novel variant, termed Autoregressive DPO (ADPO), that explicitly integrates autoregressive modeling into the preference optimization framework. Without violating the theoretical foundations, the derived loss takes an elegant form: it shifts the summation operation in the DPO objective outside the log-sigmoid function. Furthermore, through theoretical analysis of ADPO, we show that there exist two length measures to be considered when designing DPO-based algorithms: the token length $μ$ and the feedback length $μ$'. To the best of our knowledge, we are the first to explicitly distinguish these two measures and analyze their implications for preference optimization in LLMs.

Daisuke Oba, Danushka Bollegala, Masahiro Kaneko et al.

Masked Diffusion Language Models generate sequences via iterative sampling that progressively unmasks tokens. However, they still recompute the attention and feed-forward blocks for every token position at every step -- even when many unmasked tokens are essentially fixed, resulting in substantial waste in compute. We propose SureLock: when the posterior at an unmasked position has stabilized across steps (our sure condition), we lock that position -- thereafter skipping its query projection and feed-forward sublayers -- while caching its attention keys and values so other positions can continue to attend to it. This reduces the dominant per-iteration computational cost from $O(N^2d)$ to $O(MNd)$ where $N$ is the sequence length, $M$ is the number of unlocked token positions, and $d$ is the model dimension. In practice, $M$ decreases as the iteration progresses, yielding substantial savings. On LLaDA-8B, SureLock reduces algorithmic FLOPs by 30--50% relative to the same sampler without locking, while maintaining comparable generation quality. We also provide a theoretical analysis to justify the design rationale of SureLock: monitoring only the local KL at the lock step suffices to bound the deviation in final token probabilities. Our project page is available at https://daioba.github.io/surelock .

Yuxia Wang, Artem Shelmanov, Jonibek Mansurov et al.

We present the GenAI Content Detection Task~1 -- a shared task on binary machine generated text detection, conducted as a part of the GenAI workshop at COLING 2025. The task consists of two subtasks: Monolingual (English) and Multilingual. The shared task attracted many participants: 36 teams made official submissions to the Monolingual subtask during the test phase and 26 teams -- to the Multilingual. We provide a comprehensive overview of the data, a summary of the results -- including system rankings and performance scores -- detailed descriptions of the participating systems, and an in-depth analysis of submissions. https://github.com/mbzuai-nlp/COLING-2025-Workshop-on-MGT-Detection-Task1

Masahiro Kaneko, Danushka Bollegala, Timothy Baldwin

Recent studies have demonstrated that large language models (LLMs) have ethical-related problems such as social biases, lack of moral reasoning, and generation of offensive content. The existing evaluation metrics and methods to address these ethical challenges use datasets intentionally created by instructing humans to create instances including ethical problems. Therefore, the data does not reflect prompts that users actually provide when utilizing LLM services in everyday contexts. This may not lead to the development of safe LLMs that can address ethical challenges arising in real-world applications. In this paper, we create Eagle datasets extracted from real interactions between ChatGPT and users that exhibit social biases, toxicity, and immoral problems. Our experiments show that Eagle captures complementary aspects, not covered by existing datasets proposed for evaluation and mitigation of such ethical challenges. Our code is publicly available at https://huggingface.co/datasets/MasahiroKaneko/eagle.

Ryuji Hashimoto, Masahiro Kaneko, Ryosuke Takata et al.

Opinion dynamics (OD) studies how individual opinions evolve and generate collective patterns such as consensus and polarization. While recent work explores OD using populations of LLM-based agents focusing on opinion exchange, it typically does not incorporate individuals' lived experiences, such as economic outcomes of past decisions, which play a critical role in shaping opinions. We propose a novel OD simulation framework that grounds LLM-based agents in an economic environment, allowing them to act and receive environmental feedback. Our simulations exhibit coherent OD at both individual and population levels: individual opinions follow structured trajectories shaped by economic experiences, with adverse conditions inducing opinion rigidity, while at the population level, collective opinions co-move with economic conditions, with inequality amplifying polarization and price instability driving larger distributional shifts. These results highlight the importance of grounding LLM-based agents in environments to capture collective OD.

Taihei Shiotani, Masahiro Kaneko, Ayana Niwa et al.

Social biases reflected in language are inherently shaped by cultural norms, which vary significantly across regions and lead to diverse manifestations of stereotypes. Existing evaluations of social bias in large language models (LLMs) for non-English contexts, however, often rely on translations of English benchmarks. Such benchmarks fail to reflect local cultural norms, including those found in Japanese. For instance, Western benchmarks may overlook Japan-specific stereotypes related to hierarchical relationships, regional dialects, or traditional gender roles. To address this limitation, we introduce Japanese cUlture adversarial BiAs benchmarK Under handcrafted creation (JUBAKU), a benchmark tailored to Japanese cultural contexts. JUBAKU uses adversarial construction to expose latent biases across ten distinct cultural categories. Unlike existing benchmarks, JUBAKU features dialogue scenarios hand-crafted by native Japanese annotators, specifically designed to trigger and reveal latent social biases in Japanese LLMs. We evaluated nine Japanese LLMs on JUBAKU and three others adapted from English benchmarks. All models clearly exhibited biases on JUBAKU, performing below the random baseline of 50% with an average accuracy of 23% (ranging from 13% to 33%), despite higher accuracy on the other benchmarks. Human annotators achieved 91% accuracy in identifying unbiased responses, confirming JUBAKU's reliability and its adversarial nature to LLMs.

Taihei Shiotani, Masahiro Kaneko, Naoaki Okazaki

In enhancing the fairness of Large Language Models (LLMs), evaluating social biases rooted in the cultural contexts of specific linguistic regions is essential. However, most existing Japanese benchmarks heavily rely on translating English data, which does not necessarily provide an evaluation suitable for Japanese culture. Furthermore, they only evaluate bias in the conclusion, failing to capture biases lurking in the reasoning. In this study, based on attribution theory in social psychology, we constructed a new dataset, ``JUBAKU-v2,'' which evaluates the bias in attributing behaviors to in-groups and out-groups within reasoning while fixing the conclusion. This dataset consists of 216 examples reflecting cultural biases specific to Japan. Experimental results verified that it can detect performance differences across models more sensitively than existing benchmarks.

Yang Liu, Masahiro Kaneko, Chenhui Chu

Today's large language models (LLMs) are capable of supporting multilingual scenarios, allowing users to interact with LLMs in their native languages. When LLMs respond to subjective questions posed by users, they are expected to align with the views of specific demographic groups or historical periods, shaped by the language in which the user interacts with the model. Existing studies mainly focus on researching the opinions represented by LLMs among demographic groups in the United States or a few countries, lacking worldwide country samples and studies on human opinions in different historical periods, as well as lacking discussion on using language to steer LLMs. Moreover, they also overlook the potential influence of prompt language on the alignment of LLMs' opinions. In this study, our goal is to fill these gaps. To this end, we create an evaluation framework based on the World Values Survey (WVS) to systematically assess the alignment of LLMs with human opinions across different countries, languages, and historical periods around the world. We find that LLMs appropriately or over-align the opinions with only a few countries while under-aligning the opinions with most countries. Furthermore, changing the language of the prompt to match the language used in the questionnaire can effectively steer LLMs to align with the opinions of the corresponding country more effectively than existing steering methods. At the same time, LLMs are more aligned with the opinions of the contemporary population. To our knowledge, our study is the first comprehensive investigation of the topic of opinion alignment in LLMs across global, language, and temporal dimensions. Our code and data are publicly available at https://github.com/ku-nlp/global-opinion-alignment and https://github.com/nlply/global-opinion-alignment.

Masahiro Kaneko, Masato Mita, Shun Kiyono et al.

This paper investigates how to effectively incorporate a pre-trained masked language model (MLM), such as BERT, into an encoder-decoder (EncDec) model for grammatical error correction (GEC). The answer to this question is not as straightforward as one might expect because the previous common methods for incorporating a MLM into an EncDec model have potential drawbacks when applied to GEC. For example, the distribution of the inputs to a GEC model can be considerably different (erroneous, clumsy, etc.) from that of the corpora used for pre-training MLMs; however, this issue is not addressed in the previous methods. Our experiments show that our proposed method, where we first fine-tune a MLM with a given GEC corpus and then use the output of the fine-tuned MLM as additional features in the GEC model, maximizes the benefit of the MLM. The best-performing model achieves state-of-the-art performances on the BEA-2019 and CoNLL-2014 benchmarks. Our code is publicly available at: https://github.com/kanekomasahiro/bert-gec.

Masahiro Kaneko, Danushka Bollegala, Naoaki Okazaki et al.

There exist both scalable tasks, like reading comprehension and fact-checking, where model performance improves with model size, and unscalable tasks, like arithmetic reasoning and symbolic reasoning, where model performance does not necessarily improve with model size. Large language models (LLMs) equipped with Chain-of-Thought (CoT) prompting are able to make accurate incremental predictions even on unscalable tasks. Unfortunately, despite their exceptional reasoning abilities, LLMs tend to internalize and reproduce discriminatory societal biases. Whether CoT can provide discriminatory or egalitarian rationalizations for the implicit information in unscalable tasks remains an open question. In this study, we examine the impact of LLMs' step-by-step predictions on gender bias in unscalable tasks. For this purpose, we construct a benchmark for an unscalable task where the LLM is given a list of words comprising feminine, masculine, and gendered occupational words, and is required to count the number of feminine and masculine words. In our CoT prompts, we require the LLM to explicitly indicate whether each word in the word list is a feminine or masculine before making the final predictions. With counting and handling the meaning of words, this benchmark has characteristics of both arithmetic reasoning and symbolic reasoning. Experimental results in English show that without step-by-step prediction, most LLMs make socially biased predictions, despite the task being as simple as counting words. Interestingly, CoT prompting reduces this unconscious social bias in LLMs and encourages fair predictions.

Masahiro Kaneko, Youmi Ma, Yuki Wata et al.

Large Language Models (LLMs) are trained on large-scale web data, which makes it difficult to grasp the contribution of each text. This poses the risk of leaking inappropriate data such as benchmarks, personal information, and copyrighted texts in the training data. Membership Inference Attacks (MIA), which determine whether a given text is included in the model's training data, have been attracting attention. Previous studies of MIAs revealed that likelihood-based classification is effective for detecting leaks in LLMs. However, the existing methods cannot be applied to some proprietary models like ChatGPT or Claude 3 because the likelihood is unavailable to the user. In this study, we propose a Sampling-based Pseudo-Likelihood (\textbf{SPL}) method for MIA (\textbf{SaMIA}) that calculates SPL using only the text generated by an LLM to detect leaks. The SaMIA treats the target text as the reference text and multiple outputs from the LLM as text samples, calculates the degree of $n$-gram match as SPL, and determines the membership of the text in the training data. Even without likelihoods, SaMIA performed on par with existing likelihood-based methods.

Masahiro Kaneko, Timothy Baldwin

Large Language Models (LLMs) are trained on massive web-crawled corpora. This poses risks of leakage, including personal information, copyrighted texts, and benchmark datasets. Such leakage leads to undermining human trust in AI due to potential unauthorized generation of content or overestimation of performance. We establish the following three criteria concerning the leakage issues: (1) leakage rate: the proportion of leaked data in training data, (2) output rate: the ease of generating leaked data, and (3) detection rate: the detection performance of leaked versus non-leaked data. Despite the leakage rate being the origin of data leakage issues, it is not understood how it affects the output rate and detection rate. In this paper, we conduct an experimental survey to elucidate the relationship between the leakage rate and both the output rate and detection rate for personal information, copyrighted texts, and benchmark data. Additionally, we propose a self-detection approach that uses few-shot learning in which LLMs detect whether instances are present or absent in their training data, in contrast to previous methods that do not employ explicit learning. To explore the ease of generating leaked information, we create a dataset of prompts designed to elicit personal information, copyrighted text, and benchmarks from LLMs. Our experiments reveal that LLMs produce leaked information in most cases despite less such data in their training set. This indicates even small amounts of leaked data can greatly affect outputs. Our self-detection method showed superior performance compared to existing detection methods.

Masanari Oi, Masahiro Kaneko, Ryuto Koike et al.

Large Language Models (LLMs) are widely used to evaluate natural language generation tasks as automated metrics. However, the likelihood, a measure of LLM's plausibility for a sentence, can vary due to superficial differences in sentences, such as word order and sentence structure. It is therefore possible that there might be a likelihood bias if LLMs are used for evaluation: they might overrate sentences with higher likelihoods while underrating those with lower likelihoods. In this paper, we investigate the presence and impact of likelihood bias in LLM-based evaluators. We also propose a method to mitigate the likelihood bias. Our method utilizes highly biased instances as few-shot examples for in-context learning. Our experiments in evaluating the data-to-text and grammatical error correction tasks reveal that several LLMs we test display a likelihood bias. Furthermore, our proposed method successfully mitigates this bias, also improving evaluation performance (in terms of correlation of models with human scores) significantly.

Panatchakorn Anantaprayoon, Masahiro Kaneko, Naoaki Okazaki

Self-Correction based on feedback improves the output quality of Large Language Models (LLMs). Moreover, as Self-Correction functions like the slow and conscious System-2 thinking from cognitive psychology's perspective, it can potentially reduce LLMs' social biases. LLMs are sensitive to contextual ambiguities and inconsistencies; therefore, explicitly communicating their intentions during interactions when applying Self-Correction for debiasing is crucial. In this study, we demonstrate that clarifying intentions is essential for effectively reducing biases in LLMs through Self-Correction. We divide the components needed for Self-Correction into three parts: instruction, response, and feedback, and clarify intentions at each component. We incorporate an explicit debiasing prompt to convey the intention of bias mitigation from the instruction for response generation. In the response, we use Chain-of-Thought (CoT) to clarify the reasoning process. In the feedback, we define evaluation aspects necessary for debiasing and propose clear feedback through multi-aspect critiques and scoring. Through experiments, we demonstrate that self-correcting CoT responses obtained from a debiasing prompt based on multi-aspect feedback can reduce biased responses more robustly and consistently than the baselines. We also find the variation in debiasing efficacy when using models with different bias levels or separating models for response and feedback generation.

Masahiro Kaneko, Zeerak Talat, Timothy Baldwin

Iterative jailbreak methods that repeatedly rewrite and input prompts into large language models (LLMs) to induce harmful outputs -- using the model's previous responses to guide each new iteration -- have been found to be a highly effective attack strategy. Despite being an effective attack strategy against LLMs and their safety mechanisms, existing defenses do not proactively disrupt this dynamic trial-and-error cycle. In this study, we propose a novel framework that dynamically updates its defense strategy through online learning in response to each new prompt from iterative jailbreak methods. Leveraging the distinctions between harmful jailbreak-generated prompts and typical harmless prompts, we introduce a reinforcement learning-based approach that optimizes prompts to ensure appropriate responses for harmless tasks while explicitly rejecting harmful prompts. Additionally, to curb overfitting to the narrow band of partial input rewrites explored during an attack, we introduce Past-Direction Gradient Damping (PDGD). Experiments conducted on three LLMs show that our approach significantly outperforms five existing defense methods against five iterative jailbreak methods. Moreover, our results indicate that our prompt optimization strategy simultaneously enhances response quality for harmless tasks.

Ryuto Koike, Masahiro Kaneko, Ayana Niwa et al.

Detecting texts generated by Large Language Models (LLMs) could cause grave mistakes due to incorrect decisions, such as undermining student's academic dignity. LLM text detection thus needs to ensure the interpretability of the decision, which can help users judge how reliably correct its prediction is. When humans verify whether a text is human-written or LLM-generated, they intuitively investigate with which of them it shares more similar spans. However, existing interpretable detectors are not aligned with the human decision-making process and fail to offer evidence that users easily understand. To bridge this gap, we introduce ExaGPT, an interpretable detection approach grounded in the human decision-making process for verifying the origin of a text. ExaGPT identifies a text by checking whether it shares more similar spans with human-written vs. with LLM-generated texts from a datastore. This approach can provide similar span examples that contribute to the decision for each span in the text as evidence. Our human evaluation demonstrates that providing similar span examples contributes more effectively to judging the correctness of the decision than existing interpretable methods. Moreover, extensive experiments in four domains and three generators show that ExaGPT massively outperforms prior powerful detectors by up to +40.9 points of accuracy at a false positive rate of 1%.

Ayana Niwa, Masahiro Kaneko, Kentaro Inui

Large language models (LLMs) can exhibit advanced reasoning yet still generate incorrect answers. We hypothesize that such errors frequently stem from spurious beliefs, propositions the model internally considers true but are incorrect. To address this, we propose a method to rectify the belief space by suppressing these spurious beliefs while simultaneously enhancing true ones, thereby enabling more reliable inferences. Our approach first identifies the beliefs that lead to incorrect or correct answers by prompting the model to generate textual explanations, using our Forward-Backward Beam Search (FBBS). We then apply unlearning to suppress the identified spurious beliefs and enhance the true ones, effectively rectifying the model's belief space. Empirical results on multiple QA datasets and LLMs show that our method corrects previously misanswered questions without harming overall model performance. Furthermore, our approach yields improved generalization on unseen data, suggesting that rectifying a model's belief space is a promising direction for mitigating errors and enhancing overall reliability.

Yuxia Wang, Rui Xing, Jonibek Mansurov et al.

Prior studies have shown that distinguishing text generated by large language models (LLMs) from human-written one is highly challenging, and often no better than random guessing. To verify the generalizability of this finding across languages and domains, we perform an extensive case study to identify the upper bound of human detection accuracy. Across 16 datasets covering 9 languages and 9 domains, 19 annotators achieved an average detection accuracy of 87.6\%, thus challenging previous conclusions. We find that major gaps between human and machine text lie in concreteness, cultural nuances, and diversity. Prompting by explicitly explaining the distinctions in the prompts can partially bridge the gaps in over 50\% of the cases. However, we also find that humans do not always prefer human-written text, particularly when they cannot clearly identify its source.

Masanari Ohi, Masahiro Kaneko, Naoaki Okazaki et al.

Vision-language models (VLMs) have shown impressive abilities across a range of multi-modal tasks. However, existing metrics for evaluating the quality of text generated by VLMs typically focus on an overall evaluation for a specific task, such as image captioning. While the overall evaluation is essential for any task, the criteria prioritized can differ depending on the task, making it challenging for current metrics to adapt to multi-task scenarios. To address this limitation, we propose HarmonicEval, a reference-free comprehensive evaluation metric that aggregates criterion-wise scores to produce the overall score in a bottom-up manner. Furthermore, we construct the Multi-task Multi-criteria Human Evaluation (MMHE) dataset, which comprises 18,000 expert human judgments across four multi-modal tasks. Our experiments demonstrate that HarmonicEval achieves higher correlations with human judgments than conventional metrics while providing numerical scores for each criterion.

Ryuto Koike, Liam Dugan, Masahiro Kaneko et al.

Although membership inference attacks (MIAs) and machine-generated text detection target different goals, identifying training samples and synthetic texts, their methods often exploit similar signals based on a language model's probability distribution. Despite this shared methodological foundation, the two tasks have been independently studied, which may lead to conclusions that overlook stronger methods and valuable insights developed in the other task. In this work, we theoretically and empirically investigate the transferability, i.e., how well a method originally developed for one task performs on the other, between MIAs and machine text detection. For our theoretical contribution, we prove that the metric that achieves the asymptotically highest performance on both tasks is the same. We unify a large proportion of the existing literature in the context of this optimal metric and hypothesize that the accuracy with which a given method approximates this metric is directly correlated with its transferability. Our large-scale empirical experiments, including 7 state-of-the-art MIA methods and 5 state-of-the-art machine text detectors across 13 domains and 10 generators, demonstrate very strong rank correlation (rho > 0.6) in cross-task performance. We notably find that Binoculars, originally designed for machine text detection, achieves state-of-the-art performance on MIA benchmarks as well, demonstrating the practical impact of the transferability. Our findings highlight the need for greater cross-task awareness and collaboration between the two research communities. To facilitate cross-task developments and fair evaluations, we introduce MINT, a unified evaluation suite for MIAs and machine-generated text detection, with implementation of 15 recent methods from both tasks.

Masahiro Kaneko, Timothy Baldwin

Adversarial attacks by malicious users that threaten the safety of large language models (LLMs) can be viewed as attempts to infer a target property $T$ that is unknown when an instruction is issued, and becomes knowable only after the model's reply is observed. Examples of target properties $T$ include the binary flag that triggers an LLM's harmful response or rejection, and the degree to which information deleted by unlearning can be restored, both elicited via adversarial instructions. The LLM reveals an \emph{observable signal} $Z$ that potentially leaks hints for attacking through a response containing answer tokens, thinking process tokens, or logits. Yet the scale of information leaked remains anecdotal, leaving auditors without principled guidance and defenders blind to the transparency--risk trade-off. We fill this gap with an information-theoretic framework that computes how much information can be safely disclosed, and enables auditors to gauge how close their methods come to the fundamental limit. Treating the mutual information $I(Z;T)$ between the observation $Z$ and the target property $T$ as the leaked bits per query, we show that achieving error $\varepsilon$ requires at least $\log(1/\varepsilon)/I(Z;T)$ queries, scaling linearly with the inverse leak rate and only logarithmically with the desired accuracy. Thus, even a modest increase in disclosure collapses the attack cost from quadratic to logarithmic in terms of the desired accuracy. Experiments on seven LLMs across system-prompt leakage, jailbreak, and relearning attacks corroborate the theory: exposing answer tokens alone requires about a thousand queries; adding logits cuts this to about a hundred; and revealing the full thinking process trims it to a few dozen. Our results provide the first principled yardstick for balancing transparency and security when deploying LLMs.

Masahiro Kaneko, Alham Fikri Aji, Timothy Baldwin

Multilingual large language models (MLLMs) are able to leverage in-context learning (ICL) to achieve high performance by leveraging cross-lingual knowledge transfer without parameter updates. However, their effectiveness is highly sensitive to example selection, particularly in multilingual settings. Based on the findings of existing work, three key factors influence multilingual ICL: (1) semantic similarity, (2) linguistic alignment, and (3) language-specific performance. However, existing approaches address these factors independently, without explicitly disentangling their combined impact, leaving optimal example selection underexplored. To address this gap, we propose balanced multi-factor ICL (\textbf{BMF-ICL}), a method that quantifies and optimally balances these factors for improved example selection. Experiments on mCSQA and TYDI across four MLLMs demonstrate that BMF-ICL outperforms existing methods. Further analysis highlights the importance of incorporating all three factors and the importance of selecting examples from multiple languages.

Masahiro Kaneko, Danushka Bollegala, Timothy Baldwin

The output tendencies of Pre-trained Language Models (PLM) vary markedly before and after Fine-Tuning (FT) due to the updates to the model parameters. These divergences in output tendencies result in a gap in the social biases of PLMs. For example, there exits a low correlation between intrinsic bias scores of a PLM and its extrinsic bias scores under FT-based debiasing methods. Additionally, applying FT-based debiasing methods to a PLM leads to a decline in performance in downstream tasks. On the other hand, PLMs trained on large datasets can learn without parameter updates via In-Context Learning (ICL) using prompts. ICL induces smaller changes to PLMs compared to FT-based debiasing methods. Therefore, we hypothesize that the gap observed in pre-trained and FT models does not hold true for debiasing methods that use ICL. In this study, we demonstrate that ICL-based debiasing methods show a higher correlation between intrinsic and extrinsic bias scores compared to FT-based methods. Moreover, the performance degradation due to debiasing is also lower in the ICL case compared to that in the FT case.

Mengsay Loem, Masahiro Kaneko, Sho Takase et al.

Large-scale pre-trained language models such as GPT-3 have shown remarkable performance across various natural language processing tasks. However, applying prompt-based methods with GPT-3 for Grammatical Error Correction (GEC) tasks and their controllability remains underexplored. Controllability in GEC is crucial for real-world applications, particularly in educational settings, where the ability to tailor feedback according to learner levels and specific error types can significantly enhance the learning process. This paper investigates the performance and controllability of prompt-based methods with GPT-3 for GEC tasks using zero-shot and few-shot setting. We explore the impact of task instructions and examples on GPT-3's output, focusing on controlling aspects such as minimal edits, fluency edits, and learner levels. Our findings demonstrate that GPT-3 could effectively perform GEC tasks, outperforming existing supervised and unsupervised approaches. We also showed that GPT-3 could achieve controllability when appropriate task instructions and examples are given.

Masahiro Kaneko, Naoaki Okazaki

Large Language Models (LLMs) have demonstrated remarkable performance in various tasks and gained significant attention. LLMs are also used for local sequence transduction tasks, including grammatical error correction (GEC) and formality style transfer, where most tokens in a source text are kept unchanged. However, the models that generate all target tokens in such tasks have a tendency to simply copy the input text as is, without making needed changes, because the difference between input and output texts is minimal in the training data. This is also inefficient because the computational cost grows quadratically with the target sequence length with Transformer. This paper proposes predicting edit spans for the source text for local sequence transduction tasks. Representing an edit span with a position of the source text and corrected tokens, we can reduce the length of the target sequence and the computational cost for inference. We apply instruction tuning for LLMs on the supervision data of edit spans. Experiments show that the proposed method achieves comparable performance to the baseline in four tasks, paraphrasing, formality style transfer, GEC, and text simplification, despite reducing the length of the target text by as small as 21%. Furthermore, we report that the task-specific fine-tuning with the proposed method achieved state-of-the-art performance in the four tasks.

Masahiro Kaneko, Graham Neubig, Naoaki Okazaki

Humans work together to solve common problems by having discussions, explaining, and agreeing or disagreeing with each other. Similarly, if a system can have discussions with humans when solving tasks, it can improve the system's performance and reliability. In previous research on explainability, it has only been possible for the system to make predictions and for humans to ask questions about them rather than having a mutual exchange of opinions. This research aims to create a dataset and computational framework for systems that discuss and refine their predictions through dialogue. Through experiments, we show that the proposed system can have beneficial discussions with humans improving the accuracy by up to 25 points in the natural language inference task.

Yujin Takahashi, Masahiro Kaneko, Masato Mita et al.

This study investigates how supervised quality estimation (QE) models of grammatical error correction (GEC) are affected by the learners' proficiency with the data. QE models for GEC evaluations in prior work have obtained a high correlation with manual evaluations. However, when functioning in a real-world context, the data used for the reported results have limitations because prior works were biased toward data by learners with relatively high proficiency levels. To address this issue, we created a QE dataset that includes multiple proficiency levels and explored the necessity of performing proficiency-wise evaluation for QE of GEC. Our experiments demonstrated that differences in evaluation dataset proficiency affect the performance of QE models, and proficiency-wise evaluation helps create more robust models.

Mengsay Loem, Sho Takase, Masahiro Kaneko et al.

Neural models trained with large amount of parallel data have achieved impressive performance in abstractive summarization tasks. However, large-scale parallel corpora are expensive and challenging to construct. In this work, we introduce a low-cost and effective strategy, ExtraPhrase, to augment training data for abstractive summarization tasks. ExtraPhrase constructs pseudo training data in two steps: extractive summarization and paraphrasing. We extract major parts of an input text in the extractive summarization step, and obtain its diverse expressions with the paraphrasing step. Through experiments, we show that ExtraPhrase improves the performance of abstractive summarization tasks by more than 0.50 points in ROUGE scores compared to the setting without data augmentation. ExtraPhrase also outperforms existing methods such as back-translation and self-training. We also show that ExtraPhrase is significantly effective when the amount of genuine training data is remarkably small, i.e., a low-resource setting. Moreover, ExtraPhrase is more cost-efficient than the existing approaches.

Seiichiro Kondo, Kengo Hotate, Masahiro Kaneko et al.

Neural machine translation (NMT) has recently gained widespread attention because of its high translation accuracy. However, it shows poor performance in the translation of long sentences, which is a major issue in low-resource languages. It is assumed that this issue is caused by insufficient number of long sentences in the training data. Therefore, this study proposes a simple data augmentation method to handle long sentences. In this method, we use only the given parallel corpora as the training data and generate long sentences by concatenating two sentences. Based on the experimental results, we confirm improvements in long sentence translation by the proposed data augmentation method, despite its simplicity. Moreover, the translation quality is further improved by the proposed method, when combined with back-translation.

Aomi Koyama, Kengo Hotate, Masahiro Kaneko et al.

Grammatical error correction (GEC) suffers from a lack of sufficient parallel data. Therefore, GEC studies have developed various methods to generate pseudo data, which comprise pairs of grammatical and artificially produced ungrammatical sentences. Currently, a mainstream approach to generate pseudo data is back-translation (BT). Most previous GEC studies using BT have employed the same architecture for both GEC and BT models. However, GEC models have different correction tendencies depending on their architectures. Thus, in this study, we compare the correction tendencies of the GEC models trained on pseudo data generated by different BT models, namely, Transformer, CNN, and LSTM. The results confirm that the correction tendencies for each error type are different for every BT model. Additionally, we examine the correction tendencies when using a combination of pseudo data generated by different BT models. As a result, we find that the combination of different BT models improves or interpolates the F_0.5 scores of each error type compared with that of single BT models with different seeds.

Masahiro Kaneko, Danushka Bollegala

Masked Language Models (MLMs) have shown superior performances in numerous downstream NLP tasks when used as text encoders. Unfortunately, MLMs also demonstrate significantly worrying levels of social biases. We show that the previously proposed evaluation metrics for quantifying the social biases in MLMs are problematic due to following reasons: (1) prediction accuracy of the masked tokens itself tend to be low in some MLMs, which raises questions regarding the reliability of the evaluation metrics that use the (pseudo) likelihood of the predicted tokens, and (2) the correlation between the prediction accuracy of the mask and the performance in downstream NLP tasks is not taken into consideration, and (3) high frequency words in the training data are masked more often, introducing noise due to this selection bias in the test cases. To overcome the above-mentioned disfluencies, we propose All Unmasked Likelihood (AUL), a bias evaluation measure that predicts all tokens in a test case given the MLM embedding of the unmasked input. We find that AUL accurately detects different types of biases in MLMs. We also propose AUL with attention weights (AULA) to evaluate tokens based on their importance in a sentence. However, unlike AUL and AULA, previously proposed bias evaluation measures for MLMs systematically overestimate the measured biases, and are heavily influenced by the unmasked tokens in the context.

Raj Dabre, Aizhan Imankulova, Masahiro Kaneko et al.

Parallel corpora are indispensable for training neural machine translation (NMT) models, and parallel corpora for most language pairs do not exist or are scarce. In such cases, pivot language NMT can be helpful where a pivot language is used such that there exist parallel corpora between the source and pivot and pivot and target languages. Naturally, the quality of pivot language translation is more inferior to what could be achieved with a direct parallel corpus of a reasonable size for that pair. In a real-time simultaneous translation setting, the quality of pivot language translation deteriorates even further given that the model has to output translations the moment a few source words become available. To solve this issue, we propose multi-pivot translation and apply it to a simultaneous translation setting involving pivot languages. Our approach involves simultaneously translating a source language into multiple pivots, which are then simultaneously translated together into the target language by leveraging multi-source NMT. Our experiments in a low-resource setting using the N-way parallel UN corpus for Arabic to English NMT via French and Spanish as pivots reveals that in a simultaneous pivot NMT setting, using two pivot languages can lead to an improvement of up to 5.8 BLEU.