Papers by Yupei Du
Burn After Reading: Do Multimodal Large Language Models Truly Capture Order of Events in Image Sequences? (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing benchmarks focus on single image settings, but some focus on multi-image settings. |
| Approach: | They introduce the TempVS benchmark which focuses on temporal grounding and reasoning capabilities of Multimodal Large Language Models in image sequences. |
| Outcome: | The proposed model performs poorly compared to human models in vision and language tasks. |
Measuring the Instability of Fine-Tuning (2023.acl-long)
Copied to clipboard
| Challenge: | Many previous studies have investigated fine-tuning pre-trained language models on downstream tasks with varying random seeds, but they only used the standard deviation of performance scores (SD) as their measure, which is a narrow characterization of instability. |
| Approach: | They propose a systematic evaluation framework for the standard deviation of performance scores (SD) and six other measures quantifying instability of different granularity levels. |
| Outcome: | The proposed framework will be used to evaluate the validity of these measures and to improve them. |
Reason to Rote: Rethinking Memorization in Reasoning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models readily memorize arbitrary training instances, such as label noise . however, such memorization does not affect generalizable reasoning abilities . |
| Approach: | They investigate how large language models memorize label noise and why it affects generalizability. |
| Outcome: | The proposed model performs well on reasoning tasks even when memorized labels are missing . the proposed model is able to generalize to correctly answer "87+19=106" |
Disentangling the Roles of Representation and Selection in Data Pruning (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for data pruning involve many different design choices, which have not been systematically studied. |
| Approach: | They decompose data pruning into two key components: data representation and selection algorithm . theoretical and empirical results highlight crucial role of representations . |
| Outcome: | The proposed method can be used to train models with less data. |
On Support Samples of Next Word Prediction (2025.acl-long)
Copied to clipboard
| Challenge: | Language models excel in various tasks by making complex decisions, but understanding the rationale behind these decisions remains a challenge. |
| Approach: | They investigate data-centric interpretability in language models by focusing on the next-word prediction task. |
| Outcome: | The proposed model supports or deteres specific predictions, while non-support samples play a critical role in generalization and representation learning. |
Language models can learn implicit multi-hop reasoning, but only if they have lots of training data (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies explore the ability of language models to solve multi-hop reasoning tasks without chain of thought. |
| Approach: | They propose to use GPT2-style language models to train k-hop reasoning models . they show that the required training data grows exponentially in k . |
| Outcome: | The proposed models can learn implicit reasoning without chain-of-thoughts, the authors show . their training data grows exponentially in k, and the required number of transformer layers grows linearly in the model. |
Understanding Gender Bias in Knowledge Base Embeddings (2022.acl-long)
Copied to clipboard
| Challenge: | Knowledge base (KB) embeddings have been shown to contain gender biases . authors develop two new bias measures to quantify them and trace their origins in KB . |
| Approach: | They propose two ways to quantify gender biases in knowledge base (KB) embeddings . they use the influence function to inspect the contribution of each triple in KB to the overall group bias . |
| Outcome: | The proposed measures are compared with real-world census data to examine gender biases. |
Assessing the Reliability of Word Embedding Gender Bias Measures (2021.emnlp-main)
Copied to clipboard
| Challenge: | Various measures have been proposed to quantify human-like social biases in word embeddings, but they can suffer from measurement error. |
| Approach: | They propose to assess the reliability of word embedding gender bias measures by examining their reliability across different choices of random seeds, scoring rules and words. |
| Outcome: | The proposed measures can suffer from measurement error, and the results inform better design of word embedding gender bias measures. |
Exploring Human Gender Stereotypes with Word Association Test (D19-1)
Copied to clipboard
| Challenge: | Existing word embeddings have been used to study gender stereotypes in texts . however, evaluating their validities is still an open problem . et al.: this study investigates gender bias using the lens of language, especially, the words . |
| Approach: | They use word association test to derive bias scores for large amount of words . they find that these bias scores correlate well with bias in the real world . |
| Outcome: | The proposed method correlates well with bias in the real world, and with census data, it provides a different perspective on gender stereotypes in words. |
FTFT: Efficient and Robust Fine-Tuning by Transferring Training Dynamics (2025.coling-main)
Copied to clipboard
| Challenge: | Despite the success of fine-tuning Pre-trained Language Models, they remain susceptible to out-of-distribution input. |
| Approach: | They propose a novel approach that fine-tunes Pre-trained Language Models by transFerring Training dynamics (FTFT) FTFT uses more efficient reference models and aggressive early stopping . |
| Outcome: | The proposed approach improves the robustness of fine-tuned PLMs while reducing training costs. |