4 papers from NTT accepted to EMNLP2026, a leading international conference on natural language processing

4 papers submitted by NTT laboratories has been accepted for publication in Findings of EMNLP 2026. EMNLP 2026 (The 2026 Conference on Empirical Methods in Natural Language Processing), a highly selective international conference in natural language processing, will be held in Budapest, Hungary, from October 24 to 29, 2026. Among the four papers, two were accepted as Findings, one was accepted to the Industry Track, and one additional paper co-authored with a university was also accepted.

Abbreviated names of the laboratories:
SI: Social Informatics Labs., NTT, Inc.
HI: Human Informatics Labs., NTT, Inc.
CD: Computer and Data Science Labs., NTT, Inc.
NS: Network Service Systems Labs., NTT, Inc.

■Dummy Backdoor as a Defense: Removing Unknown Backdoors via Shared Internal Mechanisms for Generative LLMs

Kazuki Iwahana (SI), Masaru Matsubayashi (SI), Takuma Koyama (SI), Toshiki Shibahara (SI), Kenichiro Ominato (SI), Akira Ito (Tohoku Univ.)

Backdoor attacks poison a model so that it produces malicious outputs only when a specific trigger known exclusively to the attacker is present. Defending against such attacks is challenging because the content and form of the attacker’s trigger are unknown. In this study, we focus on the fact that different backdoors share common internal states in LLMs. By embedding and subsequently removing a controllable backdoor (“dummy backdoor”), we demonstrated that our method effectively suppresses unknown backdoor attacks while preserving the original utility of LLMs . Our proposed method is expected to suppress harmful responses and unintended behaviors in LLMs provided by third parties, contributing to the development of safer AI services.

■Conservative Inference-Time Tuning with an Improvement Guarantee

Daiki Chijiwa (CD), Taku Hasegawa (HI)

Large language models (LLMs) can acquire specialized capabilities through fine-tuning for specific domains or tasks, but the emergence of new foundation models requires repeated fine-tuning, leading to substantial computational and data management costs. To address this challenge, we have developed a new technique that transfers the benefits of fine-tuning to new foundation models solely through inference-time computation, without any retraining, and provided theoretical guarantees for its performance. Unlike existing techniques, which lack such guarantees and can sometimes cause significant performance degradation, our technique is theoretically guaranteed to prevent such degradation under certain conditions. Further development of this principle is expected to pave the way for efficient and reliable AI technologies that enable specialized models to be updated in line with advances in foundation models, without the need for retraining.

■What Else Needs Fixing? Exploring Cost-Effective Test-Time Compute for Revision Propagation in Artifacts Generated Through Conversation

Daisuke Kikuta (NS)

In human–LLM collaboration, artifacts are created and revised iteratively through conversation. Since users often specify only the parts they want to revise, LLMs are expected to identify how these changes affect other parts and propagate necessary revisions throughout the artifact on behalf of users. In this paper, we propose a benchmark for evaluating this ability of LLMs to propagate revisions in structured data generated through conversation. Furthermore, we compare the accuracy and computational cost of test-time compute methods under various conditions. Based on these comparisons, we provide practical guidelines for selecting effective methods and determining the number of responses to generate to improve revision accuracy while keeping computational costs low.