Papers by Hoyun Song
Does Rationale Quality Matter? Enhancing Mental Disorder Detection via Selective Reasoning Distillation (2025.findings-acl)
Copied to clipboard
| Challenge: | incorporating clinical symptom information into a model enhances domain expertise, improving its detection and interpretation performance. large language models are effective for generating explanatory rationales, but inconsistencies in relevance and domain alignment of LLM-generated rationale are challenging. |
| Approach: | They propose a framework that fine-tunes smaller language models with rationales exhibiting high domain relevance . they propose 'quality-focused' approach that selects rationale based on their alignment with clinical reasoning . |
| Outcome: | The proposed framework improves mental health detection and interpretation performance by ensuring high-quality rationales with domain relevance. |
Beyond Static Benchmarks: Synthesizing Harmful Content via Persona-based Simulation for Robust Evaluation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing static benchmarks for harmful content detection face limitations in scalability and diversity. |
| Approach: | They propose a framework for synthesizing harmful content using persona-guided large language model agents. |
| Outcome: | The proposed framework achieves a high success rate in harmful generation tests across multiple detection systems. |
Lossless Acceleration of Large Language Models with Hierarchical Drafting based on Temporal Locality in Speculative Decoding (2025.findings-naacl)
Copied to clipboard
Sukmin Cho, Sangjin Choi, Taeho Hwang, Jeongyeon Seo, Soyeong Jeong, Huije Lee, Hoyun Song, Jong C. Park, Youngjin Kwon
| Challenge: | Existing methods for drafting and verifying tokens require significant fine-tuning or have inconsistent performance across tasks. |
| Approach: | They propose a lossless drafting approach that organizes various token sources into multiple databases in a hierarchical framework based on temporal locality. |
| Outcome: | The proposed method outperforms existing database drafting methods on Spec-Bench using 7B and 13B parameters. |
TReX: Tokenizer Regression for Optimal Data Mixture (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing approaches to train and inference tokenizers rely on heuristics or large-scale searches to determine optimal data mixtures. |
| Approach: | They propose a regression-based framework that efficiently predicts the optimal data mixture for tokenizer training. |
| Outcome: | The proposed model outperforms mixtures based on LLaMA3 and uniform distributions by up to 12% in both in- and out-of-distribution compression efficiency. |
ELF22: A Context-based Counter Trolling Dataset to Combat Internet Trolls (2022.lrec-1)
Copied to clipboard
| Challenge: | a new dataset aims to automate the method to counter trolls . trolleds cause psychological damage to individuals and increase social costs . |
| Approach: | They propose to use a dataset to generate counter responses by varying counter responses according to a given strategy. |
| Outcome: | The proposed method improves strategy-controlled sentence generation. |
Towards Effective Counter-Responses: Aligning Human Preferences with Strategies to Combat Online Trolling (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to counter trolling in online communities are not yet available to address the diversity of trolling behaviors. |
| Approach: | They propose a method for generating counter-responses to trolls by aligning these strategies with human preferences across different trolled contexts. |
| Outcome: | The proposed approach reduces negative effects of trolling and improves the online community environment. |
EXIT: Context-Aware Extractive Compression for Enhancing Retrieval-Augmented Generation (2025.findings-acl)
Copied to clipboard
| Challenge: | Current retrieval-augmented generation systems struggle when retrieval models fail to rank the most relevant documents . existing extractive methods reduce latency but rely on independent, non-adaptive sentence selection . |
| Approach: | They introduce an extractive context compression framework that enhances retrieval-augmented generation in question answering. |
| Outcome: | EXIT surpasses existing compression methods and uncompressed baselines in QA accuracy . the framework reduces inference time and token count while preserving contextual dependencies . |
Ask LLMs Directly, “What shapes your bias?”: Measuring Social Bias in Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to evaluate social bias in large language models have limitations . et al., 1995: stereotypes shape social perceptions without objective basis . |
| Approach: | They propose a method to intuitively quantify social perceptions and suggest metrics to evaluate biases within LLMs. |
| Outcome: | The proposed metrics capture the multi-dimensional aspects of social bias, the paper shows . they show that the proposed metrics can be used to evaluate bias in large language models . |
Spotting Out-of-Character Behavior: Atomic-Level Evaluation of Persona Fidelity in Open-Ended Generation (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing evaluation methods struggle to capture subtle inconsistencies in large language models. |
| Approach: | They propose an atomic-level evaluation framework that quantifies persona fidelity at a finer granularity. |
| Outcome: | The proposed framework detects inconsistencies that prior evaluation methods overlook . it captures subtle deviations that real users would encounter . |
RoleConflictBench: A Benchmark of Role Conflict Scenarios for Evaluating LLMs’ Contextual Sensitivity (2026.findings-acl)
Copied to clipboard
| Challenge: | a new benchmark measures the contextual sensitivity of large language models in role conflict scenarios . role conflicts are social dilemmas where multiple roles cannot be fulfilled simultaneously . authors: models are forced to arbitrate between dynamic contextual cues and learned preferences . |
| Approach: | They propose a benchmark to measure the contextual sensitivity of large language models in role conflict scenarios. |
| Outcome: | The proposed benchmark measures the contextual sensitivity of large language models in role conflict scenarios. |
Semantic Hardness Is Not Visual Hardness: Sign-Aware Hard Negative Mining for Sign Language Retrieval (2026.acl-long)
Copied to clipboard
Junmyeong Lee, Chan Hur, ChangSu Choi, Sukmin Cho, Fitsum Gaim, Eui Jun Hwang, Hoyun Song, KyungTae Lim
| Challenge: | Existing methods for sign language retrieval fail to capture visual ambiguity . semantically distinct yet visually confusable signs are rarely treated as hard negatives . |
| Approach: | They propose a method that constructs hard negatives based on visual confusability rather than linguistic similarity. |
| Outcome: | The proposed method significantly improves fine-grained retrieval performance while preserving coarse-grain accuracy. |
A Simple and Flexible Modeling for Mental Disorder Detection by Learning from Clinical Questionnaires (2023.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to detecting mental disorders lack domain-based interpretation . lack of quality data or complexity of models can cause problems . |
| Approach: | They propose a model that captures semantic meanings directly from social media and compares them to symptom-related descriptions. |
| Outcome: | The proposed model outperforms baselines on mental disorder detection tasks. |
Social Dynamics as Critical Vulnerabilities that Undermine Objective Decision-Making in LLM Collectives (2026.acl-long)
Copied to clipboard
| Challenge: | Large language model (LLM) agents are increasingly acting as human delegates in multi-agent environments, where a representative agent integrates diverse peer perspectives to make a final decision. |
| Approach: | They define four key phenomena—social conformity, perceived expertise, dominant speaker effect, and rhetorical persuasion—and manipulate the number of adversaries, relative intelligence, argument length, and argumentative styles. |
| Outcome: | The results show that the reliability of the representative agent is undermined by the social context of its network. |