Papers by Heeseung Kim
FlashAdventure: A Benchmark for GUI Agents Solving Full Story Arcs in Diverse Adventure Games (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing game benchmarks lack diversity and evaluate GUI agents on completing entire storylines. |
| Approach: | They propose a benchmark of 34 Flash-based adventure games to test full story arc completion and tackle observation-behavior gap. |
| Outcome: | The proposed benchmarks show GUI agents struggle with full story arc completion while others improve on observation-behavior gaps. |
WoW-Bench: Evaluating Fine-Grained Acoustic Perception in Audio-Language Models via Marine Mammal Vocalizations (2026.findings-acl)
Copied to clipboard
| Challenge: | Large audio-language models extend language understanding into the auditory domain, yet their ability to perform low-level listening, such as pitch and duration detection, remains underexplored. |
| Approach: | They propose a global benchmark to evaluate low-level auditory perception and cognition using marine mammal vocalizations to better assess models’ low- level listening. |
| Outcome: | The proposed models show performance far below human levels, indicating a need for stronger auditory grounding in LALMs. |
Rare Tokens Degenerate All Tokens: Improving Neural Text Generation via Adaptive Gradient Gating for Rare Token Embeddings (2022.acl-long)
Copied to clipboard
| Challenge: | Recent studies have determined that the learned token embeddings of large-scale neural language models are degenerated to be anisotropic with a narrow-cone shape. |
| Approach: | They propose a method to degenerate the learning gradient for rare token embeddings by gating the specific part of the gradient for all tokens during training stage. |
| Outcome: | The proposed method improves the performance of the models but lacks the training dynamics needed to solve the representation degeneration problem. |
Can LLMs Deceive CLIP? Benchmarking Adversarial Compositionality of Pre-trained Multimodal Representation via Text Updates (2025.acl-long)
Copied to clipboard
| Challenge: | Recent advances in multimodal systems have demonstrated remarkable capabilities in generating multimodal content from multimodal inputs. |
| Approach: | They propose a benchmark that leverages large language models to generate deceptive text samples to exploit compositional vulnerabilities across different modalities. |
| Outcome: | The proposed approach exploits compositional vulnerabilities across images, videos, and audios. |
EdiText: Controllable Coarse-to-Fine Text Editing with Diffusion Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for text editing have been proposed for various types of data with diverse attributes. |
| Approach: | They propose a novel text editing method that modifies reference text to desired attributes at various scales. |
| Outcome: | The proposed method is capable of making precise adjustments within the desired range while maintaining the accuracy of the reference text. |
Does Your Voice Assistant Remember? Analyzing Conversational Context Recall and Utilization in Voice Interaction Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in multi-turn voice interaction models have improved user-model communication, but whether open-source models share this ability remains unexplored. |
| Approach: | They propose to use ContextDialog to evaluate open-source interaction models' ability to recall past utterances to identify key limitations. |
| Outcome: | The proposed model retains and recalls past utterances better than closed-source models, but still struggles with questions about past . findings highlight key limitations in open-source model and suggest ways to improve memory retention and retrieval robustness. |
Still Between Us? Evaluating and Improving Voice Assistant Robustness to Third-Party Interruptions (2026.acl-long)
Copied to clipboard
| Challenge: | Recent Spoken Language Models lack the capability to discern Third-Party Interruptions (TPI) from the primary user’s ongoing flow, leaving them vulnerable to contextual failures. |
| Approach: | They propose a dataset with speaker-aware hard negatives to enforce acoustic cue prioritization for interruption handling and a framework to measure the interruption-handling strategy and precise speaker discrimination in deceptive contexts. |
| Outcome: | The proposed framework mitigates semantic shortcut learning while neglecting acoustic signals essential for discerning speaker changes. |