Papers by Giwon Hong
Steering Knowledge Selection Behaviours in LLMs via SAE-Based Representation Engineering (2025.naacl-long)
Copied to clipboard
Yu Zhao, Alessio Devoto, Giwon Hong, Xiaotang Du, Aryo Pradipta Gema, Hongru Wang, Xuanli He, Kam-Fai Wong, Pasquale Minervini
| Challenge: | Large language models store factual knowledge in their parameters but their parametric knowledge can conflict with the information provided in the context. |
| Approach: | They propose a training-free representation engineering method that uses pre-trained sparse auto-encoders to control the knowledge selection behaviour of large language models. |
| Outcome: | The proposed method can control the use of both knowledge sources to resolve knowledge conflict in open-domain question-answering tasks surpassing existing representation engineering methods (+10%) and contrastive decoding methods (+5%). |
Exploiting Numerical-Contextual Knowledge to Improve Numerical Reasoning in Question Answering (2022.findings-naacl)
Copied to clipboard
| Challenge: | Existing numerical reasoning models overly rely on parametric knowledge at inference time . previous studies show that understanding numbers in text improves numerical reasoning accuracy . |
| Approach: | They propose a numerical reasoning model that leverages parametric knowledge to alleviate this over-reliance on parametric information. |
| Outcome: | The proposed model improves numerical reasoning accuracy and performance in DROP. |
Handling Anomalies of Synthetic Questions in Unsupervised Question Answering (2020.coling-main)
Copied to clipboard
| Challenge: | Existing approaches to improve unsupervised Question Answering (UQA) are expensive and require additional datasets. |
| Approach: | They propose an unsupervised QA approach that generates QA training data automatically. |
| Outcome: | The proposed method improves unsupervised QA significantly across a number of QA tasks. |
Mixtures of In-Context Learners (2025.acl-long)
Copied to clipboard
| Challenge: | In-context learning is sensitive to the choice of in-con context demonstrations and processing many demonstrations can be computationally demanding. |
| Approach: | They propose a method that uses subsets of demonstrations to train experts via ICL and learns a weighting function to merge their output distributions via gradient-based optimisation. |
| Outcome: | The proposed approach improves on 5 out of 7 classification datasets compared to strong baselines and reduces the inference time needed to achieve the same performance with fewer demonstrations. |
Have You Seen That Number? Investigating Extrapolation in Question Answering Models (2021.emnlp-main)
Copied to clipboard
| Challenge: | Numerical reasoning in machine reading comprehension (MRC) has shown drastic improvements over the past few years. |
| Approach: | They propose an E-digit number form that alleviates the lack of extrapolation in numerical MRC models. |
| Outcome: | The proposed model can't extrapolate to unseen numbers, the authors say . they also show that the model needs to treat numbers differently from regular words . |
Why So Gullible? Enhancing the Robustness of Retrieval-Augmented Models against Counterfactual Noise (2024.findings-naacl)
Copied to clipboard
| Challenge: | Existing retrieval-augmented language models assume query relevance and irrelevance as dichotomy . existing models are highly brittle to the presence of conflicting information in both the fine-tuning and in-context few-shot learning scenarios. |
| Approach: | They propose methods for handling knowledge conflicts by fine-tuning a discriminator or prompting it to elicit its discriminative capability. |
| Outcome: | The proposed approaches significantly enhance model robustness on open-domain QA. |
Analyzing LLM Instruction Optimization for Tabular Fact Verification (2026.findings-eacl)
Copied to clipboard
Xiaotang Du, Giwon Hong, Wai-Chung Kwan, Rohit Saxena, Ivan Titov, Pasquale Minervini, Emily Allaway
| Challenge: | evaluating instruction optimization for tabular fact verification is a key challenge for reliable NLP systems. |
| Approach: | They compare instruction optimization for tabular fact verification with a framework based on DSPy . they find that instruction optimization consistently improves verification accuracy . |
| Outcome: | The proposed method improves verification accuracy across four benchmarks and three model families. |
Regularization of Distinct Strategies for Unsupervised Question Generation (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Unsupervised question answering (UQA) is a task of answering questions from a context that contains the answer. |
| Approach: | They propose a method to generate higher-quality questions with a teacher-student architecture and a regularization module to avoid bias toward a particular question generation strategy. |
| Outcome: | The proposed method generates higher-quality questions across diverse datasets and tasks and can be used to create a model with few-shot learning. |
FinePrompt: Unveiling the Role of Finetuned Inductive Bias on Compositional Reasoning in GPT-4 (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models such as GPT-4 have demonstrated impressive capability to solve textual understanding problems at a level parallel to or surpassing state-of-the-art taskspecific models. |
| Approach: | They propose to transfer task-specific inductive biases from finetuned models to prompts to improve GPT-4's compositional reasoning capabilities. |
| Outcome: | The proposed prompt scheme shows competitive zero-shot and few-shot performances compared to existing prompts on complicated reasoning tasks. |
Ultra-High Dimensional Sparse Representations with Binarization for Efficient Text Retrieval (2021.emnlp-main)
Copied to clipboard
| Challenge: | Recent approaches to information retrieval (IR) and natural language processing (NLP) use contextual language models, which can improve both synonymy and polysemy problems associated with words. |
| Approach: | They propose an ultra-high dimensional representation scheme equipped with directly controllable sparsity and a bucketing method where embeddings from multiple layers of BERT are selected/merged to represent diverse linguistic aspects. |
| Outcome: | The proposed representation scheme outperforms sparse models with MS MARCO and TREC CAR, and shows that it is highly efficient for storage and search. |
Are We Done with MMLU? (2025.naacl-long)
Copied to clipboard
Aryo Pradipta Gema, Joshua Ong Jun Leang, Giwon Hong, Alessio Devoto, Alberto Carlo Maria Mancino, Rohit Saxena, Xuanli He, Yu Zhao, Xiaotang Du, Mohammad Reza Ghasemi Madani, Claire Barale, Robert McHardy, Joshua Harris, Jean Kaddour, Emile Van Krieken, Pasquale Minervini
| Challenge: | MMLU is widely adopted but its ground truth errors obscure the true capabilities of LLMs. |
| Approach: | They propose a framework for identifying dataset errors using a novel error annotation protocol and a subset of 5,700 manually re-annotated questions. |
| Outcome: | The proposed framework is based on 5,700 re-annotated questions from the MMLU benchmark. |
Theorem Prover as a Judge for Synthetic Data Generation (2025.acl-long)
Copied to clipboard
| Challenge: | Recent studies show that large language models are increasingly capable of tackling mathematical problems. |
| Approach: | They propose an approach that iteratively refines theorem prover formalisation to mitigate errors. |
| Outcome: | The proposed method increases execution rate on the Lean prover from 60% to 87%, while human annotation is replaced with theorem prover feedback. |
Graph-Induced Transformers for Efficient Multi-Hop Question Answering (2022.emnlp-main)
Copied to clipboard
| Challenge: | Recent MHQA tasks that require inter-paragraph/sentence linkages use graphs to model internal structural information within text. |
| Approach: | They propose a graph-induced transformer that applies graph-derived attention patterns directly into a PLM without external graph modules. |
| Outcome: | The proposed model can replace external graph modules while preserving model performance. |
GRADA: Graph-based Reranking against Adversarial Documents Attack (2025.emnlp-main)
Copied to clipboard
Jingjie Zheng, Aryo Pradipta Gema, Giwon Hong, Xuanli He, Pasquale Minervini, Youcheng Sun, Qiongkai Xu
| Challenge: | Retrieval Augmented Generation (RAG) frameworks are susceptible to adversarial attacks that manipulate the retrieval process by introducing documents that are adversarially similar to the query. |
| Approach: | They propose a framework that integrates external retrieval modules into RAG frameworks to improve the factual accuracy of large language models. |
| Outcome: | The proposed framework reduces adversarial attacks by 80% while maintaining minimal loss in accuracy. |
Compositional Steering of Large Language Models with Steering Tokens (2026.acl-long)
Copied to clipboard
| Challenge: | Existing work addresses compositional steering for a single behavior, but it's limited by computational costs and performance. |
| Approach: | They propose compositional steering tokens for multi-behavior steering by embedding individual behaviors into dedicated tokens via self-distillation. |
| Outcome: | The proposed steering tokens generalize well to unseen compositions compared to competing approaches. |