Yupei Du

Raoyuan Zhao, Yihong Liu, Yupei Du et al.

Mathematical reasoning benchmarks are vital for evaluating large language models (LLMs), but many are static and repeatedly exposed through public evaluation and training pipelines, making it difficult to separate genuine reasoning from memorization. Meanwhile, manually constructing new math problems with reliable answers remains costly. We introduce ReverseMath, a scalable method for generating new math problems through answer inversion. Given a problem and its answer, ReverseMath masks a numerical value in the original problem, treats the original answer as a known condition, and rewrites the problem so that the masked value becomes the new answer. The generated problem reverses the original input-output relation, making its answer known by construction. We study ReverseMath for both evaluation and training. For evaluation, paired original/reversed problems reveal substantial behavioral shifts: models sometimes fail on reversed problems and even incorrectly output the original answer, suggesting memorization-like behavior. For training, ReverseMath provides automatically labeled reversed problems as data augmentation for reinforcement learning (RL). Experiments show that including ReverseMath-generated data improves mathematical reasoning performance across multiple benchmarks, demonstrating its value as both an analysis tool and a scalable source of verifiable training data.

Olusanmi Hundogan, Xixi Lu, Yupei Du et al.

Predictive process analytics focuses on predicting future states, such as the outcome of running process instances. These techniques often use machine learning models or deep learning models (such as LSTM) to make such predictions. However, these deep models are complex and difficult for users to understand. Counterfactuals answer ``what-if'' questions, which are used to understand the reasoning behind the predictions. For example, what if instead of emailing customers, customers are being called? Would this alternative lead to a different outcome? Current methods to generate counterfactual sequences either do not take the process behavior into account, leading to generating invalid or infeasible counterfactual process instances, or heavily rely on domain knowledge. In this work, we propose a general framework that uses evolutionary methods to generate counterfactual sequences. Our framework does not require domain knowledge. Instead, we propose to train a Markov model to compute the feasibility of generated counterfactual sequences and adapt three other measures (delta in outcome prediction, similarity, and sparsity) to ensure their overall viability. The evaluation shows that we generate viable counterfactual sequences, outperform baseline methods in viability, and yield similar results when compared to the state-of-the-art method that requires domain knowledge.

Yupei Du, Dong Nguyen

Fine-tuning pre-trained language models on downstream tasks with varying random seeds has been shown to be unstable, especially on small datasets. Many previous studies have investigated this instability and proposed methods to mitigate it. However, most studies only used the standard deviation of performance scores (SD) as their measure, which is a narrow characterization of instability. In this paper, we analyze SD and six other measures quantifying instability at different levels of granularity. Moreover, we propose a systematic framework to evaluate the validity of these measures. Finally, we analyze the consistency and difference between different measures by reassessing existing instability mitigation methods. We hope our results will inform the development of better measurements of fine-tuning instability.

Yash Sarrof, Yupei Du, Katharina Stein et al.

Transformers have shown inconsistent success in AI planning tasks, and theoretical understanding of when generalization should be expected has been limited. We take important steps towards addressing this gap by analyzing the ability of decoder-only models to verify whether a given plan correctly solves a given planning instance. To analyse the general setting where the number of objects -- and thus the effective input alphabet -- grows at test time, we introduce C*-RASP, an extension of C-RASP designed to establish length generalization guarantees for transformers under the simultaneous growth in sequence length and vocabulary size. Our results identify a large class of classical planning domains for which transformers can provably learn to verify long plans, and structural properties that significantly affects the learnability of length generalizable solutions. Empirical experiments corroborate our theory.

Yupei Du, Albert Gatt, Dong Nguyen

Despite the massive success of fine-tuning Pre-trained Language Models (PLMs), they remain susceptible to out-of-distribution input. Dataset cartography is a simple yet effective dual-model approach that improves the robustness of fine-tuned PLMs. It involves fine-tuning a model on the original training set (i.e. reference model), selecting a subset of important training instances based on the training dynamics, and fine-tuning again only on these selected examples (i.e. main model). However, this approach requires fine-tuning the same model twice, which is computationally expensive for large PLMs. In this paper, we show that (1) training dynamics are highly transferable across model sizes and pre-training methods, and that (2) fine-tuning main models using these selected training instances achieves higher training efficiency than empirical risk minimization (ERM). Building on these observations, we propose a novel fine-tuning approach: Fine-Tuning by transFerring Training dynamics (FTFT). Compared with dataset cartography, FTFT uses more efficient reference models and aggressive early stopping. FTFT achieves robustness improvements over ERM while lowering the training cost by up to $\sim 50\%$.

Yingjin Song, Yupei Du, Denis Paperno et al.

This paper introduces the TempVS benchmark, which focuses on temporal grounding and reasoning capabilities of Multimodal Large Language Models (MLLMs) in image sequences. TempVS consists of three main tests (i.e., event relation inference, sentence ordering and image ordering), each accompanied with a basic grounding test. TempVS requires MLLMs to rely on both visual and linguistic modalities to understand the temporal order of events. We evaluate 38 state-of-the-art MLLMs, demonstrating that models struggle to solve TempVS, with a substantial performance gap compared to human capabilities. We also provide fine-grained insights that suggest promising directions for future research. Our TempVS benchmark data and code are available at https://github.com/yjsong22/TempVS.

Goya van Boven, Yupei Du, Dong Nguyen

Gender-neutral pronouns are increasingly being introduced across Western languages. Recent evaluations have however demonstrated that English NLP systems are unable to correctly process gender-neutral pronouns, with the risk of erasing and misgendering non-binary individuals. This paper examines a Dutch coreference resolution system's performance on gender-neutral pronouns, specifically hen and die. In Dutch, these pronouns were only introduced in 2016, compared to the longstanding existence of singular they in English. We additionally compare two debiasing techniques for coreference resolution systems in non-binary contexts: Counterfactual Data Augmentation (CDA) and delexicalisation. Moreover, because pronoun performance can be hard to interpret from a general evaluation metric like LEA, we introduce an innovative evaluation metric, the pronoun score, which directly represents the portion of correctly processed pronouns. Our results reveal diminished performance on gender-neutral pronouns compared to gendered counterparts. Nevertheless, although delexicalisation fails to yield improvements, CDA substantially reduces the performance gap between gendered and gender-neutral pronouns. We further show that CDA remains effective in low-resource settings, in which a limited set of debiasing documents is used. This efficacy extends to previously unseen neopronouns, which are currently infrequently used but may gain popularity in the future, underscoring the viability of effective debiasing with minimal resources and low computational costs.

Yuekun Yao, Yupei Du, Dawei Zhu et al.

Implicit reasoning is the ability of a language model to solve multi-hop reasoning tasks in a single forward pass, without chain of thought. We investigate this capability using GPT2-style language models trained from scratch on controlled $k$-hop reasoning datasets ($k = 2, 3, 4$). We show that while such models can indeed learn implicit $k$-hop reasoning, the required training data grows exponentially in $k$, and the required number of transformer layers grows linearly in $k$. We offer a theoretical explanation for why this depth growth is necessary. We further find that the data requirement can be mitigated, but not eliminated, through curriculum learning.

Yupei Du, Philipp Mondorf, Silvia Casola et al.

Large language models readily memorize arbitrary training instances, such as label noise, yet they perform strikingly well on reasoning tasks. In this work, we investigate how language models memorize label noise, and why such memorization in many cases does not heavily affect generalizable reasoning capabilities. Using two controllable synthetic reasoning datasets with noisy labels, four-digit addition (FDA) and two-hop relational reasoning (THR), we discover a reliance of memorization on generalizable reasoning mechanisms: models continue to compute intermediate reasoning outputs even when retrieving memorized noisy labels, and intervening reasoning adversely affects memorization. We further show that memorization operates through distributed encoding, i.e., aggregating various inputs and intermediate results, rather than building a look-up mechanism from inputs to noisy labels. Moreover, our FDA case study reveals memorization occurs via outlier heuristics, where existing neuron activation patterns are slightly shifted to fit noisy labels. Together, our findings suggest that memorization of label noise in language models builds on, rather than overrides, the underlying reasoning mechanisms, shedding lights on the intriguing phenomenon of benign memorization.

Yupei Du, Yingjin Song, Hugh Mee Wong et al.

Data pruning, selecting small but impactful subsets, offers a promising way to efficiently scale NLP model training. However, existing methods often involve many different design choices, which have not been systematically studied. This limits future developments. In this work, we decompose data pruning into two key components: the data representation and the selection algorithm, and we systematically analyze their influence on the selection of instances. Our theoretical and empirical results highlight the crucial role of representations: better representations, e.g., training gradients, generally lead to a better selection of instances, regardless of the chosen selection algorithm. Furthermore, different selection algorithms excel in different settings, and none consistently outperforms the others. Moreover, the selection algorithms do not always align with their intended objectives: for example, algorithms designed for the same objective can select drastically different instances, highlighting the need for careful evaluation.

Yuqian Li, Yupei Du, Yufang Liu et al.

Language models excel in various tasks by making complex decisions, yet understanding the rationale behind these decisions remains a challenge. This paper investigates \emph{data-centric interpretability} in language models, focusing on the next-word prediction task. Using representer theorem, we identify two types of \emph{support samples}-those that either promote or deter specific predictions. Our findings reveal that being a support sample is an intrinsic property, predictable even before training begins. Additionally, while non-support samples are less influential in direct predictions, they play a critical role in preventing overfitting and shaping generalization and representation learning. Notably, the importance of non-support samples increases in deeper layers, suggesting their significant role in intermediate representation formation. These insights shed light on the interplay between data and model decisions, offering a new dimension to understanding language model behavior and interpretability.

Yupei Du, Jacek Golebiowski, Philipp Schmidt et al.

Techniques for knowledge graph (KGs) enrichment have been increasingly crucial for commercial applications that rely on evolving product catalogues. However, because of the huge search space of potential enrichment, predictions from KG completion (KGC) methods suffer from low precision, making them unreliable for real-world catalogues. Moreover, candidate facts for enrichment have varied relevance to users. While making correct predictions for incomplete triplets in KGs has been the main focus of KGC method, the relevance of when to apply such predictions has been neglected. Motivated by the product search use case, we address the angle of generating relevant completion for a catalogue using user search behaviour and the users property association with a product. In this paper, we present our intuition for identifying enrichable data points and use general-purpose KGs to show-case the performance benefits. In particular, we extract entity-predicate pairs from user queries, which are more likely to be correct and relevant, and use these pairs to guide the prediction of KGC methods. We assess our method on two popular encyclopedia KGs, DBPedia and YAGO 4. Our results from both automatic and human evaluations show that query guidance can significantly improve the correctness and relevance of prediction.

Yupei Du, Qixiang Fang, Dong Nguyen

Various measures have been proposed to quantify human-like social biases in word embeddings. However, bias scores based on these measures can suffer from measurement error. One indication of measurement quality is reliability, concerning the extent to which a measure produces consistent results. In this paper, we assess three types of reliability of word embedding gender bias measures, namely test-retest reliability, inter-rater consistency and internal consistency. Specifically, we investigate the consistency of bias scores across different choices of random seeds, scoring rules and words. Furthermore, we analyse the effects of various factors on these measures' reliability scores. Our findings inform better design of word embedding gender bias measures. Moreover, we urge researchers to be more critical about the application of such measures.

Guanyi Chen, Yinhe Zheng, Yupei Du

Personalised response generation enables generating human-like responses by means of assigning the generator a social identity. However, pragmatics theory suggests that human beings adjust the way of speaking based on not only who they are but also whom they are talking to. In other words, when modelling personalised dialogues, it might be favourable if we also take the listener's social identity into consideration. To validate this idea, we use gender as a typical example of a social variable to investigate how the listener's identity influences the language used in Chinese dialogues on social media. Also, we build personalised generators. The experiment results demonstrate that the listener's identity indeed matters in the language use of responses and that the response generator can capture such differences in language use. More interestingly, by additionally modelling the listener's identity, the personalised response generator performs better in its own identity.