Zilu Tang

Zilu Tang, Qiao Zhao, Gabriel Franco et al.

Entity tracking (ET), the ability to keep track of states, is a fundamental skill that underlies complex reasoning. An increasing amount of work investigates how transformer language models (LMs) solve entity binding $\textit{without}$ state changes. However, there is limited understanding of how non-toy LMs address ET problems of realistic difficulties expressed in natural language. To this end, we investigate the mechanisms underlying ET in more complex scenarios featuring multiple state-changing operations. We find that LMs do not incrementally track world states across tokens or query-relevant states across layers, but simply aggregate relevant information in parallel at the last token when the query becomes evident. We further investigate mechanisms of individual operations ($\texttt{PUT}$, $\texttt{REMOVE}$, $\texttt{MOVE}$) to characterize this non-incremental ET mechanism. Surprisingly, LMs implement the $\texttt{REMOVE}$ operation with a fragile global suppression tag; this global removal mechanism predicts various failure modes that we confirm behaviorally. We provide a mechanistic solution of nullifying this tag to partially address this issue. Overall, our findings reveal that LMs solve a fundamentally sequential task using a non-sequential strategy. More broadly, our work illustrates how behavioral and mechanistic analyses can fruitfully interact. Behavioral results inform mechanistic hypotheses, and insights from mechanistic analyses help build stronger behavioral evaluations by predicting failure modes missing from existing evaluations.

Zilu Tang, Mayank Agarwal, Alex Shypula et al.

This work explores the use of self-generated natural language explanations as an intermediate step for code-to-code translation with language models. Across three types of explanations and 19 programming languages constructed from the MultiPL-E dataset, we find the explanations to be particularly effective in the zero-shot case, improving performance by 12% on average. Improvements with natural language explanations are particularly pronounced on difficult programs. We release our dataset, code, and canonical solutions in all 19 languages.

Zilu Tang, Muhammed Yusuf Kocyigit, Derry Wijaya

Data augmentation techniques have been proven useful in many applications in NLP fields. Most augmentations are task-specific, and cannot be used as a general-purpose tool. In our work, we present AugCSE, a unified framework to utilize diverse sets of data augmentations to achieve a better, general purpose, sentence embedding model. Building upon the latest sentence embedding models, our approach uses a simple antagonistic discriminator that differentiates the augmentation types. With the finetuning objective borrowed from domain adaptation, we show that diverse augmentations, which often lead to conflicting contrastive signals, can be tamed to produce a better and more robust sentence representation. Our methods achieve state-of-the-art results on downstream transfer tasks and perform competitively on semantic textual similarity tasks, using only unsupervised data.

Zilu Tang, Derry Wijaya

Incorporating tagging into neural machine translation (NMT) systems has shown promising results in helping translate rare words such as named entities (NE). However, translating NE in low-resource setting remains a challenge. In this work, we investigate the effect of using tags and NE hypernyms from knowledge graphs (KGs) in parallel corpus in different levels of resource conditions. We find the tag-and-copy mechanism (tag the NEs in the source sentence and copy them to the target sentence) improves translation in high-resource settings only. Introducing copying also results in polarizing effects in translating different parts-of-speech (POS). Interestingly, we find that copy accuracy for hypernyms is consistently higher than that of entities. As a way of avoiding "hard" copying and utilizing hypernym in bootstrapping rare entities, we introduced a "soft" tagging mechanism and found consistent improvement in high and low-resource settings.

David Anugraha, Zilu Tang, Lester James V. Miranda et al. · cambridge

Reward models are essential for aligning language model outputs with human preferences, yet existing approaches often lack both controllability and interpretability. These models are typically optimized for narrow objectives, limiting their generalizability to broader downstream tasks. Moreover, their scalar outputs are difficult to interpret without contextual reasoning. To address these limitations, we introduce $\shortmethodname$, a novel reward modeling framework that is rubric-agnostic, generalizable across evaluation dimensions, and provides interpretable, reasoned score assignments. $\shortmethodname$ enables more transparent and flexible evaluation of language models, supporting robust alignment with diverse human values and use cases. Our models, data, and code are available as open source at https://github.com/rubricreward/r3.

David Anugraha, Shou-Yi Hung, Zilu Tang et al.

Evaluation using Large Language Model (LLM) judges has been widely adopted in English and shown to be effective for automatic evaluation. However, their performance does not generalize well to non-English settings, and it remains unclear what constitutes effective multilingual training for such judges. In this paper, we introduce mR3, a massively multilingual, rubric-agnostic reward reasoning model trained on 72 languages, achieving the broadest language coverage in reward modeling to date. We present a comprehensive study of data and curriculum selection for training to identify effective strategies and data sources for building high-quality reward models, including the integration of target-language reasoning datasets. Our approach attains state-of-the-art performance on multilingual reward model benchmarks, surpassing much larger models (i.e., GPT-OSS-120B) while being up to 9x smaller, and its effectiveness is further confirmed through extensive ablation studies. Our models, data, and code are available as open source at https://github.com/rubricreward/mr3.

Zilu Tang, Rajen Chatterjee, Sarthak Garg · apple-ml

Machine Translation (MT) is undergoing a paradigm shift, with systems based on fine-tuned large language models (LLM) becoming increasingly competitive with traditional encoder-decoder models trained specifically for translation tasks. However, LLM-based systems are at a higher risk of generating hallucinations, which can severely undermine user's trust and safety. Most prior research on hallucination mitigation focuses on traditional MT models, with solutions that involve post-hoc mitigation - detecting hallucinated translations and re-translating them. While effective, this approach introduces additional complexity in deploying extra tools in production and also increases latency. To address these limitations, we propose a method that intrinsically learns to mitigate hallucinations during the model training phase. Specifically, we introduce a data creation framework to generate hallucination focused preference datasets. Fine-tuning LLMs on these preference datasets reduces the hallucination rate by an average of 96% across five language pairs, while preserving overall translation quality. In a zero-shot setting our approach reduces hallucinations by 89% on an average across three unseen target languages.

Ryandito Diandaru, Lucky Susanto, Zilu Tang et al.

Large Language Models (LLMs) demonstrate strong machine translation capabilities on languages they are trained on. However, the impact of factors beyond training data size on translation performance remains a topic of debate, especially concerning languages not directly encountered during training. Our study delves into Llama2's translation capabilities. By modeling a linear relationship between linguistic feature distances and machine translation scores, we ask ourselves if there are potentially better central languages for LLMs other than English. Our experiments show that the 7B Llama2 model yields above 10 BLEU when translating into all languages it has seen, which rarely happens for languages it has not seen. Most translation improvements into unseen languages come from scaling up the model size rather than instruction tuning or increasing shot count. Furthermore, our correlation analysis reveals that syntactic similarity is not the only linguistic factor that strongly correlates with machine translation scores. Interestingly, we discovered that under specific circumstances, some languages (e.g. Swedish, Catalan), despite having significantly less training data, exhibit comparable correlation levels to English. These insights challenge the prevailing landscape of LLMs, suggesting that models centered around languages other than English could provide a more efficient foundation for multilingual applications.

Tyler A. Chang, Catherine Arnett, Abdelrahman Eldesokey et al. · uw

To date, there exist almost no culturally-specific evaluation benchmarks for large language models (LLMs) that cover a large number of languages and cultures. In this paper, we present Global PIQA, a participatory commonsense reasoning benchmark for over 100 languages, constructed by hand by 335 researchers from 65 countries around the world. The 116 language varieties in Global PIQA cover five continents, 14 language families, and 23 writing systems. In the non-parallel split of Global PIQA, over 50% of examples reference local foods, customs, traditions, or other culturally-specific elements. We find that state-of-the-art LLMs perform well on Global PIQA in aggregate, but they exhibit weaker performance in lower-resource languages (up to a 37% accuracy gap, despite random chance at 50%). Open models generally perform worse than proprietary models. Global PIQA highlights that in many languages and cultures, everyday knowledge remains an area for improvement, alongside more widely-discussed capabilities such as complex reasoning and expert knowledge. Beyond its uses for LLM evaluation, we hope that Global PIQA provides a glimpse into the wide diversity of cultures in which human language is embedded.

Zilu Tang, Afra Feyza Akyürek, Ekin Akyürek et al. · mit

A prominent issue in aligning language models (LMs) to personalized preferences is underspecification -- the lack of information from users about their preferences. A popular trend of injecting such specification is adding a prefix (e.g. prior relevant conversations) to the current user's conversation to steer preference distribution. Most methods passively model personal preferences with prior example preferences pairs. We ask whether models benefit from actively inferring preference descriptions, and address this question by creating a synthetic personalized alignment dataset based on famous people with known public preferences. We then test how effective finetuned 1-8B size models are at inferring and aligning to personal preferences. Results show that higher-quality active prefixes lead to better generalization, more contextually faithful models, and less systematic biases across different protected attributes. All our results suggest active alignment can lead to a more controllable and efficient path for personalized alignment.

Lucky Susanto, Musa Wijanarko, Prasetia Pratama et al.

Polarization is defined as divisive opinions held by two or more groups on substantive issues. As the world's third-largest democracy, Indonesia faces growing concerns about the interplay between political polarization and online toxicity, which is often directed at vulnerable minority groups. Despite the importance of this issue, previous NLP research has not fully explored the relationship between toxicity and polarization. To bridge this gap, we present a novel multi-label Indonesian dataset that incorporates toxicity, polarization, and annotator demographic information. Benchmarking this dataset using BERT-base models and large language models (LLMs) shows that polarization information enhances toxicity classification, and vice versa. Furthermore, providing demographic information significantly improves the performance of polarization classification.

Masataro Asai, Zilu Tang

We propose an unsupervised neural model for learning a discrete embedding of words. Unlike existing discrete embeddings, our binary embedding supports vector arithmetic operations similar to continuous embeddings. Our embedding represents each word as a set of propositional statements describing a transition rule in classical/STRIPS planning formalism. This makes the embedding directly compatible with symbolic, state of the art classical planning solvers.