Team Naver Has 24 Papers Accepted at EMNLP 2026, Expanding AI Research Output

Team Naver on the 15th unveiled 24 papers accepted at EMNLP 2026. EMNLP (Empirical Methods in Natural Language Processing) is a major international conference in the natural language processing field, often mentioned alongside ACL and NAACL. This year's EMNLP will be held in Budapest, Hungary, from the 24th to the 29th.

The papers highlighted this time reflect research achievements from Team Naver's technology organizations, including Naver, Naver Cloud, and Naver Labs Europe. The accepted papers span a range of Natural Language Processing subtopics, including Korean AI, multimodal AI, inference analysis of large language models (LLM), and AI agents.

The research also includes cases in which problems identified during AI model and service development were expanded into research topics. Team Naver also presented results that developed into papers by broadening issues uncovered in actual development work into research tasks.

Naver Cloud conducted research centered on Korean inference data and multimodal AI evaluation. It proposed a method for synthesizing Korean reasoning training data using English data from specialized domains, and analyzed the causes of degraded language capability during the process of combining vision and language models. It also developed diagnostic metrics to identify language capability degradation in advance.

The accepted studies covered topics ranging from long-video and omnimodal AI evaluation standards to internal reasoning interpretation in LLMs, improved training efficiency, and agent generalization. A benchmark study was accepted, and its evaluation targets included the performance of long-video and omnimodal models. The study analyzed the process of finding sections in videos lasting several hours that answer natural-language queries, and proposed evaluation standards for context-appropriate speech tone and speed responses in omnimodal AI.

Naver conducted research on internal reasoning processes in LLMs and training efficiency analysis. Naver AI Lab analyzed the internal differentiation of LLMs across different reasoning processes, including problem understanding, planning, and calculation, and expanded research into interpreting how models arrive at their answers.

Naver and its affiliated organizations presented research aimed at understanding reasoning structures and improving training and utilization methods rather than simply scaling up model size. In this regard, they presented research on post-training efficiency, as well as studies on how to preserve existing capabilities in mixture-of-experts (MoE) models. The focus of the research was on understanding reasoning processes and improving training efficiency rather than expanding model scale.

Naver Labs Europe studied retrieval-augmented LLM agents and model alignment technologies. It also researched methods for retrieving an AI agent's past task experiences and using them to solve new tasks, with the goal of improving generalization performance on unfamiliar tasks.

The scope of the research included work on adapting LLM-generated sentences for TTS suitability. It also addressed LLM alignment aimed at generating text that sounds natural when converted to speech without post-processing.

Team Naver continued to deliver research results at EMNLP following 6 papers accepted at 'ICML 2026' in July this year and 23 papers accepted at 'ECCV 2026' last September. The company plans to reflect these research results in next-generation AI model and service development.

Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=408254

References

• https://www.epnc.co.kr/news/articleView.html?idxno=408254