Self-Consistency (SC) is an effective decoding strategy that improves the reasoning performance of Large Language Models (LLMs) by generating multiple chain-of-thought reasoning paths and selecting the final answer via majority voting. However, it suffers from substantial inference costs because it requires a large number of samples. To mitigate this issue, Difficulty-Adaptive Self-Consistency (DSC) was proposed to reduce unnecessary token usage for easy problems by adjusting the number of samples according to problem difficulty. However, DSC requires additional model calls and pre-sampling to estimate difficulty, and this process is repeated when applying to each dataset, leading to significant computational overhead. In this work, we propose Activation-Informed Difficulty-Aware Self-Consistency (ACTSC) to address these limitations. ACTSC leverages internal difficulty signals reflected in the feed-forward network neuron activations to construct a lightweight difficulty estimation probe, without any additional token generation or model calls. The probe dynamically adjusts the number of samples for SC and can be applied to new datasets without requiring pre-sampling for difficulty estimation. To validate its effectiveness, we conduct experiments on five benchmarks. Experimental results show that ACTSC effectively reduces inference costs while maintaining accuracy relative to existing methods.
@article{yoon2026breaking,title={Breaking the Pre-Sampling Barrier: Activation-Informed Difficulty-Aware Self-Consistency},author={Yoon, Taewoong and Jeong, Geunyeong and Park, Geon and Yeom, Sihyeong and Kim, Harksoo},journal={arXiv preprint arXiv:2602.09438},year={2026},url={https://arxiv.org/abs/2602.09438},note={Under review at IJCAI 2026}}
SemEval
K-NLPers at SemEval-2026 Task 7: Multiple LLM Agent Debate System for Everyday Knowledge Across Diverse Languages and Cultures
Jiwoo Song*, Sihyeong Yeom*, and Harksoo Kim
In Proceedings of the 20th International Workshop on Semantic Evaluation (SemEval-2026), Jul 2026
This paper presents K-NLPers’ system for SemEval-2026 Task 7. We propose a Multiple LLM Agent Debate System designed to capture and evaluate everyday knowledge across diverse languages and cultures, showcasing advanced multi-agent coordination and robust cross-cultural reasoning capability.
@inproceedings{semeval2026,title={K-NLPers at SemEval-2026 Task 7: Multiple LLM Agent Debate System for Everyday Knowledge Across Diverse Languages and Cultures},author={Song, Jiwoo and Yeom, Sihyeong and Kim, Harksoo},booktitle={Proceedings of the 20th International Workshop on Semantic Evaluation (SemEval-2026)},year={2026},month=jul,address={San Diego, California (USA)},note={To be published (Presented on July 02-07)},}
2025
HCLT
Improving the Efficiency of Iterative Retrieval-Augmented Generation with an External Classifier
Sihyeong Yeom, Geunyeong Jeong, Seonghee Lee, and Harksoo Kim
In Annual Conference on Human and Language Technology, Oct 2025
검색 증강 생성은 질문과 관련된 정보를 검색하여 맥락으로 사용함으로써 대형 언어 모델이 학습할 때 보지 못한 질문에도 답변할 수 있게 만든다. 하지만, 일반적인 검색 증강 생성은 단 한 번의 검색과 생성을 수행하기 때문에 복잡한 질의를 효과적으로 다루는 데 한계가 존재한다. 최근 연구에서는 복잡한 질의를 하위 질의로 분해하고, 반복적으로 검색과 생성을 수행하는 학습 기반 반복적 검색 증강 생성(Iterative RAG)이 뛰어난 성능을 보여주었지만, 언어 모델의 반복 호출로 인한 높은 계산 비용이 요구된다. 이에 본 연구는 외부 분류기를 활용하여 성능을 크게 저하시키지 않으면서도, 토큰 소비량을 효과적으로 줄일 수 있는 방법을 제안한다.
@inproceedings{hclt_2025,title={Improving the Efficiency of Iterative Retrieval-Augmented Generation with an External Classifier},author={Yeom, Sihyeong and Jeong, Geunyeong and Lee, Seonghee and Kim, Harksoo},booktitle={Annual Conference on Human and Language Technology},pages={652--657},year={2025},month=oct,}
2024
HCLT
Analysis of knowledge forgetting problem in Locate-Then-Edit knowledge editing method
Sihyeong Yeom, SeongHee Lee, SeongSik Park, and Hark-Soo Kim
In Annual Conference on Human and Language Technology, 2024
지식 편집은 거대 언어 모델에서 잘못되었거나 오래된 지식을 수정하고 새로운 지식을 주입하기 위한 기술이다. 지식 편집의 목적은 기존 모델의 성능을 유지하면서 특정 지식만을 효율적으로 변경하는 것이다. 그 중 Locate-Then-Edit 방법은 인과 매개 분석을 수행하여 모델에서 지식이 저장된 특정 위치를 탐색하고 일부 매개변수만을 편집함으로써 효율적인 지식 편집의 가능성을 보여준다. 하지만, 이와 같은 편집 방법은 동일한 subject에 대한 연속적인 지식 편집을 수행했을 때, 이전에 편집된 지식이 망각된다는 문제가 발생한다. 본 논문에서는 이러한 동일 subject에 대한 동시 및 연속 지식 편집에서 발생하는 문제점을 구체적으로 탐구한다.
@inproceedings{hclt_2024,title={Analysis of knowledge forgetting problem in Locate-Then-Edit knowledge editing method},author={Yeom, Sihyeong and Lee, SeongHee and Park, SeongSik and Kim, Hark-Soo},booktitle={Annual Conference on Human and Language Technology},pages={590--594},year={2024},}