Papers by Hwanjun Song
Learning to Verify Summary Facts with Fine-Grained LLM Feedback (2025.coling-main)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have significantly enhanced the text summarization performance, but hallucination issues still occur in summaries. |
| Approach: | They propose a large-scale dataset containing fine-grained factual feedback on summaries that can be fine tuned by using Large Language Models (LLMs) they employ 10 distinct LLMs for diverse summary generation and Llama-3-70B-Instruct for feedback. |
| Outcome: | The proposed model outperforms models trained on smaller human-annotated datasets while maintaining high performance. |
FineSurE: Fine-grained Summarization Evaluation using LLMs (2024.acl-long)
Copied to clipboard
| Challenge: | Existing methods for text summarization evaluation do not correlate well with human judgments . evaluators that use Likert scale scores are limited in their ability to perform deeper analysis. |
| Approach: | They propose a fine-grained evaluator specifically tailored for the summarization task using large language models. |
| Outcome: | The proposed method improves on open-source and proprietary LLMs and shows better completeness and conciseness than existing methods. |
MAGID: An Automated Pipeline for Generating Synthetic Multi-modal Datasets (2024.naacl-long)
Copied to clipboard
Hossein Aboutalebi, Hwanjun Song, Yusheng Xie, Arshit Gupta, Lijia Sun, Hang Su, Igor Shalyminov, Nikolaos Pappas, Siffi Singh, Saab Mansour
| Challenge: | Existing approaches to augment textual dialogues with retrieved images pose privacy, diversity, and quality constraints. |
| Approach: | They propose a framework to augment text-only dialogues with diverse and high-quality images by using a diffusion model and a feedback loop. |
| Outcome: | The proposed framework is comparable to or better than baselines, with significant improvements in human evaluation, especially against retrieval baselines where the image database is small. |
UniSumEval: Towards Unified, Fine-grained, Multi-dimensional Summarization Evaluation for LLMs (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing benchmarks for summarization quality evaluation lack diverse input scenarios, focus on narrowly defined dimensions, and struggle with subjective and coarse-grained annotation schemes. |
| Approach: | They propose to use AI to help human annotations and identifie potentially hallucinogenic input texts. |
| Outcome: | The proposed benchmarks improve on existing benchmarks in terms of input diversity, granularity of human annotations, and evaluation dimensions. |
Distilling LLM Reasoning into Dense Encoders: Bridging the Accuracy-Efficiency Gap in Recommendation (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing distillation approaches target Small Language Models (SLMs) or Conventional Recommendation Models, but face a critical trade-off between computational cost and semantic reasoning capacity. |
| Approach: | They propose a framework that establishes a text encoder as the optimal student architecture for scalable recommendation. |
| Outcome: | Experiments on four datasets show that the proposed framework outperforms state-of-the-art models and achieves significantly reduced latency. |
TofuEval: Evaluating Hallucinations of LLMs on Topic-Focused Dialogue Summarization (2024.naacl-long)
Copied to clipboard
Liyan Tang, Igor Shalyminov, Amy Wong, Jon Burnsky, Jake Vincent, Yu’an Yang, Siffi Singh, Song Feng, Hwanjun Song, Hang Su, Lijia Sun, Yi Zhang, Saab Mansour, Kathleen McKeown
| Challenge: | Existing LLMs hallucinate significant amounts of factual errors in the dialogue domain, regardless of the model’s size. |
| Approach: | They propose to evaluate topic-focused dialogue summarization by using large language models (LLMs) they use human annotations to evaluate factual consistency and explain factually inconsistent sentences. |
| Outcome: | The proposed evaluation benchmark on topic-focused dialogue summarization shows that existing LLMs hallucinate significant amounts of factual errors regardless of the model’s size. |
Towards a Holistic and Automated Evaluation Framework for Multi-Level Comprehension of LLMs in Book-Length Contexts (2025.emnlp-main)
Copied to clipboard
Yuho Lee, Jiaqi Deng, Nicole Hee-Yeon Kim, Hyangsuk Min, Taewon Yun, Minjeong Ban, Kim Yul, Hwanjun Song
| Challenge: | HAMLET is a framework for evaluating the long-context comprehension of large language models. |
| Approach: | They propose a framework for evaluating the long-context comprehension of large language models . HAMLET structures key information of source texts into a three-level hierarchy . |
| Outcome: | HAMLET achieves 90% agreement with expert judgments while reducing evaluation cost by up to 25. |
Enhancing Abstractiveness of Summarization Models through Calibrated Distillation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to generate abstractive summarizations are slow and abstractive, but we propose a novel approach to enhance the level of abstractiveness without sacrificing the informativeness of generated summaries. |
| Approach: | They propose a novel approach to enhance the level of abstractiveness without sacrificing the informativeness of generated summaries by exposing diverse pseudo summary with two supervision to the student model. |
| Outcome: | The proposed method outperforms previous methods in abstractive summarization distillation, producing highly abstractive and informative summaries. |
Distilling Long-CoT Reasoning through Collaborative Step-wise Multi-Teacher Decoding (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing curation-based approaches to inference are inefficient and fail to adapt dynamically, leading to redundant sampling and missed opportunities for complementary reasoning. |
| Approach: | They propose a collaborative multi-teacher decoding framework that performs step-wise reasoning synthesis guided by predictive perplexity–based scoring and beam search. |
| Outcome: | The proposed framework generates higher-quality reasoning data and achieves student-level results, demonstrating that fine-grained collaboration yields structured, efficient, and robust reasoning distillation. |
Word2Passage: Word-level Importance Re-weighting for Query Expansion (2025.findings-acl)
Copied to clipboard
| Challenge: | Retrieval-augmented generation (RAG) enhances the quality of LLM generation by providing relevant chunks, but retrieving accurately from external knowledge remains challenging due to missing contextually important words in query expansion. |
| Approach: | They propose a method that generates word, sentence, and passage references for query expansion and assigns distinct importance scores to words based on their origin and characteristics. |
| Outcome: | The proposed method outperforms existing methods across datasets and LLM configurations, effectively enhancing retrieval accuracy and generation quality. |
Can Your Model Tell a Negation from an Implicature? Unravelling Challenges With Intent Encoders (2024.acl-long)
Copied to clipboard
| Challenge: | Conversational systems often rely on embedding models for intent classification and intent clustering tasks. |
| Approach: | They propose a toolkit that gives a more holistic view of intent embedding models by considering three tasks– (1) intent classification, (2) intent clustering, and (3) a novel triplet task. |
| Outcome: | The proposed model improves on the linguistic dimensions while affecting performance on downstream task metrics. |
QuDAR: Query-Wise Dual-Perspective Adaptive Retrieval (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to grounding large language models rely on static weights and a static retrieval component. |
| Approach: | They propose a dual-perspective adaptive retrieval framework that adapts along two perspectives: retriever type (sparse vs. dense) and query format (original v. expanded). |
| Outcome: | The proposed framework adapts along two perspectives: retriever type (sparse vs. dense) and query format (original v. expanded). |
Semi-Supervised Dialogue Abstractive Summarization via High-Quality Pseudolabel Selection (2024.naacl-long)
Copied to clipboard
| Challenge: | Semi-supervised dialogue summarization (SSDS) leverages model-generated summaries to reduce reliance on human-labeled data. |
| Approach: | They propose a scoring approach that encapsulates three primary dimensions of summarization model quality. |
| Outcome: | The proposed method reduces reliance on human-labeled data and improves the performance of summarization models. |
Fast and Robust Early-Exiting Framework for Autoregressive Language Models with Synchronized Parallel Decoding (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to speed up inference latency have shown performance degradation caused by a state copying mechanism or numerous exit paths. |
| Approach: | They propose a framework that allocates adaptive computation paths for each token based on the complexity of generating the subsequent token. |
| Outcome: | The proposed framework outperforms existing frameworks on extensive generation tasks. |
Faithful, Unfaithful or Ambiguous? Multi-Agent Debate with Initial Stance for Summary Evaluation (2025.naacl-long)
Copied to clipboard
Mahnaz Koupaee, Jake W. Vincent, Saab Mansour, Igor Shalyminov, Han He, Hwanjun Song, Raphael Shu, Jianfeng He, Yi Nian, Amy Wing-mei Wong, Kyu J. Han, Hang Su
| Challenge: | Existing approaches to evaluate faithfulness of summaries are often fooled by the fluency of the text and struggle with identifying errors. |
| Approach: | They propose an approach to summary faithfulness evaluation where multiple LLM-based agents are assigned initial stances and forced to come up with a reason to justify belief. |
| Outcome: | The proposed approach can identify ambiguities and have even stronger performance on non-ambiguous summaries. |
Towards Multi-dimensional Evaluation of LLM Summarization across Domains and Languages (2025.acl-long)
Copied to clipboard
Hyangsuk Min, Yuho Lee, Minjeong Ban, Jiaqi Deng, Nicole Hee-Yeon Kim, Taewon Yun, Hang Su, Jason Cai, Hwanjun Song
| Challenge: | Existing evaluation frameworks for text summarization lack domain-specific assessment criteria and are predominantly English-centric. |
| Approach: | They propose a multi-dimensional, multi-domain evaluation of summarization in English and Chinese that incorporates specialized assessment criteria for each domain and leverages a debate system to enhance annotation quality. |
| Outcome: | The proposed evaluation framework provides a multi-dimensional, multi-domain evaluation of summarization in English and Chinese. |
Rethinking LLM-Based Recommendations: A Personalized Query-Driven Parallel Integration (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Query-to-Recommendation framework integrates large langucage models into recommendation systems . but it faces training-induced bias and bottlenecks from serialized architecture . |
| Approach: | They propose a parallel recommendation framework that decouples LLMs from candidate pre-selection and direct retrieval over the entire item pool. |
| Outcome: | The proposed framework decouples LLMs from candidate pre-selection and enables direct retrieval over the entire item pool. |
Learning to Summarize from LLM-generated Feedback (2025.naacl-long)
Copied to clipboard
| Challenge: | Developing effective text summarizers remains a challenge due to issues like unfaithful statements, key information omissions, and verbosity. |
| Approach: | They propose a large-scale dataset containing multi-dimensional feedback on LLM-generated summaries of varying quality across diverse domains to align them with human preferences for faithfulness, completeness, and conciseness. |
| Outcome: | The proposed model outperforms the 10x larger Llama3-70b-instruct in generating human-preferred summaries. |
Alignment Tuning for Large Language Models: A Data-Centric Lens on Alignment Data Pipelines (2026.findings-acl)
Copied to clipboard
| Challenge: | a large number of alignment tuning literature is organized around optimization objectives, but the construction of alignment data is often treated implicitly. |
| Approach: | They propose to decompose alignment data construction into three interacting stages . they identify recurring design trade-offs and failure modes observed across prior alignment methods . |
| Outcome: | The proposed framework identifies recurring design trade-offs and failure modes observed across prior alignment methods and outlines open challenges for alignment data pipelines including prompt-level alignment, agentic settings, and alignment under evolving objectives. |