NABLAS to Present Three Papers at JSAI 2026 on AI-Generated Video Forgery Detection and Japan-Specific Video QA

NABLAS Co., Ltd. (Headquarters: Hongo, Bunkyo-ku, Tokyo; Representative Director and Director of Research: Kotaro Nakayama; hereinafter 'the Company'), operating as a comprehensive AI research institute, announces that three papers by its R&D Division have been accepted for oral and poster presentations at the 40th Annual Conference of the Japanese Society for Artificial Intelligence (JSAI 2026), one of Japan's largest AI research conferences, to be held in June 2026.

NABLAS is actively advancing research and development (R&D) in cutting-edge AI technologies, including deepfake detection, multimodal large language models (MLLMs), and data analysis agents, to solve social issues and create new value. From these research activities, papers on three themes have been accepted: marketing support, an AI evaluation platform specialized for Japanese culture, and fake video detection to counter the evolving threat of generative AI. These will be presented in oral and poster sessions.

In addition to the Company's own papers, two papers resulting from joint research with Toyota Motor Corporation are scheduled for oral and poster presentations.

[Recommendation System for Food Name Marketing Support via SNS Analysis] With support from 'GENIAC (Generative AI Accelerator Challenge),' a project by the Ministry of Economy, Trade and Industry and NEDO aimed at strengthening Japan's generative AI development capabilities, the Company worked on developing visual language models and services utilizing them for approximately six months from October 2024 to the end of April 2025. Based on these R&D outcomes, further research and development towards the practical application of trend prediction from SNS data has been advanced.

Since food trends directly impact manufacturing, inventory, and logistics planning, there is a demand for methods to early extract and organize food name candidates worthy of attention in marketing analysis from SNS mentions. However, existing criteria for selection are not sufficiently established. This study proposes a framework that supports exploratory analysis in food trend analysis by integrating and ranking food names based on the assumption that they possess both 'current low frequency (rarity)' and 'atypical word composition (unexpectedness).'

General Session: AI Application: Marketing and Recommendation Monday, June 8, 2026, 13:40~15:10, Room J (Medium Conference Room 201B)

[Video Forgery Detection with Optical Flow Residuals and Spatial-Temporal Consistency] The rapid advancement of diffusion model-based video generation technology has made it possible to generate increasingly realistic synthetic content, including natural-looking fake videos. Traditional fake video detection methods identified fakes by capturing subtle inconsistencies between frames, but the evolution of generative AI has made capturing these subtle inconsistencies difficult. This study proposes a novel detection framework to capture these subtle inconsistencies (temporal consistency disruptions) and identify fake videos.

This detection framework utilizes spatial and temporal consistency by combining RGB appearance features with optical flow residuals. Furthermore, large-scale experiments on text-to-video and image-to-video tasks across 10 diverse generative models demonstrated high robustness and excellent generalization performance for the proposed method.

Poster Session Monday, June 8, 2026, 14:10~15:40, Hall Y (Exhibition Hall AB-1)

[Japanese Video-QA: Construction and Evaluation of a Video Question-Answering Benchmark Specialized for Japanese Culture] This paper proposes 'Japanese Video-QA,' a video question-answering benchmark specialized for Japanese culture. While Multimodal Large Language Models (MLLMs) have rapidly enhanced their ability to process images, audio, and video, there are few benchmarks that can quantitatively evaluate video understanding capabilities that are strongly dependent on the Japanese language and Japanese culture. Japanese Video-QA is a benchmark built to fill this gap, representing an initial attempt to visualize the difficulty of video understanding in the specific domain of Japanese culture.

The dataset consists of 800 question-answer pairs constructed from 428 YouTube videos related to Japan, using question generation by Gemini 2.5 Flash followed by manual verification and correction. The videos cover 6 domains (seasons/events, tourist attractions, traditional culture, food culture, nature/landscapes, pop culture) and 100 sub-domains.

Poster Session Monday, June 8, 2026, 16:10~17:40, Hall Y (Exhibition Hall AB-1)

[Multimodal LLM Design for Time-Series Data Interpretation and Its Application to Vehicle Driving Behavior Explanation] Traditionally, developing machine learning models that take time-series data as input required building specialized models for each application, with limitations in accuracy. To solve this challenge, this study proposes a novel multimodal LLM (Time-Series Language Model: TSLM) capable of processing time-series data.

General Session: AI Application: Multimodal Understanding and Interpretation Tuesday, June 9, 2026, 15:30~17:00, Hall E (Main Hall C)

[Construction of a Vehicle Driving Dataset for the Development of a Multimodal LLM for Time-Series Data Interpretation] In recent years, extensive research has been conducted on making large language models (LLMs) multimodal by utilizing diverse datasets including images, videos, and audio in addition to natural language. However, the development of multimodal LLMs capable of interpreting time-series data has not progressed sufficiently due to a lack of large-scale datasets pairing time-series data with explanatory text. This study focuses on vehicle driving data among time-series data, aiming to generate explanatory text for driving situations and construct a large-scale dataset. It proposes a pipeline that graphs vehicle driving data and processes it with a visual language model (VLM).

Poster Session Tuesday, June 9, 2026, 16:00~17:30, Hall Y (Exhibition Hall AB-1)

The Company is also challenging itself with applications beyond fake detection technology, including automation of data analysis by AI agents, automation of causal inference including automatic DAG construction, and improving productivity and anomaly detection reporting through the use of time-series data in factories and manufacturing sites, particularly in the manufacturing industry.