We aspire to put people first and integrate information and communications technology into our lives to solve the challenges our customers and society face.
Global
Joined June 2010
- Tweets4.8K
- Following2.4K
- Followers2.4K
- Likes3.4K
材料計算の準備に、どれだけ時間を使っていますか？
富士通の材料開発向けAIエージェントを活用した社内事例では、人手で約1か月半を要すると見積もられる準備作業を、約3.5時間で完了しました。
材料計算では、ツールごとの操作手順に加え、構造の作成や計算条件の設定に専門知識が求められます。本AIエージェントは、材料計算のノウハウをスキルとして活用し、指示に応じて必要なツールを選択・実行します。
全固体電池のLi金属／Li₆PS₅Cl界面を対象に、構造作成からニューラルネットワークポテンシャル（NNP）の構築までを実行。30,672原子の界面系で、1ナノ秒のMDシミュレーションを安定に実行できることも確認しました。
準備の負担を減らし、研究者が材料の分析や検討により多くの時間を充てられるように。
適用事例と技術の詳細を、Tech Blogでご紹介しています。blog.fltech.dev/entry/2026/1…
技術にご興味をお持ちの方は、以下の連絡先までお問い合わせください。[email protected]
#FujitsuTechBlog #MaterialsInformatics #AIエージェント #材料開発
Fujitsu Research at EMNLP 2026 (@emnlpmeeting)
How can we make AI agents safer, more efficient, and better suited to real-world use?
At the recently concluded EMNLP 2026 and its affiliated workshops, Fujitsu Research and our collaborators explored this question across seven papers - one oral presentation, four Findings papers, and two workshop papers.
Here’s a glimpse of our contributions:
🔒 Safer AI agents
• AgenTRIM: Tool Risk Mitigation for Agentic AI
Adapting tool permissions at each step to reduce attack success while improving task performance.
• RIFT-Bench: Dynamic Red-teaming For Agentic AI Systems
Automated, adaptive red-teaming demonstrated across 45 agentic systems.
• Training-Free Policy Violation Detection via Activation-Space Whitening in LLM
Using LLM internal representations to detect organizational policy violations with substantially less computation.
🧑 More faithful human simulations
• Measuring the Behavioral Fidelity of Long-Horizon Human Activity Simulations (Oral)
Drawing on 43 hours of real-world office activity to examine how closely simulated people reproduce human behavior - and why no single metric tells the full story.
⚡ More efficient multi-agent systems
• RCR-Router: Role-Conditioned Context Routing for Resource-Efficient Multi-Agent LLM Systems
Routing context according to each agent’s role, reducing token usage by 35% and runtime by 38% on average while maintaining or improving task performance.
🛠️ Scalable memory and enterprise deployment | REALM Workshop
• From Trajectories to Workflows: Online Procedural Memory Abstraction for Language Agents
Consolidating agent trajectories into reusable workflows, with competitive task success and 3.5× fewer memory tokens on average.
• EnterpriseLab: A Full-Stack Platform for developing and deploying agents in Enterprises
A unified platform for developing and deploying specialized enterprise agents from modular environment schemas.
Congratulations to all our researchers and collaborators! We look forward to building on these contributions and continuing the conversation on reliable, practical agentic AI.
Read the full papers via the links in the comments.
#EMNLP2026 #FujitsuResearch #AgenticAI #AISecurity #LLMAgents
7
3
158
7. EnterpriseLab: A Full-Stack Platform for developing and deploying agents in Enterprises: arxiv.org/abs/2603.21630
25
6. From Trajectories to Workflows: Online Procedural Memory Abstraction for Language Agents: aclanthology.org/2026.realm-…
10
5. RCR-Router:Role-Conditioned Context Routing for Resource-Efficient Multi-Agent LLM Systems: aclanthology.org/2026.findin…
8
4. Measuring the Behavioral Fidelity of Long-Horizon Human Activity Simulations: aclanthology.org/2026.emnlp-…
7
3. Training-Free Policy Violation Detection via Activation-Space Whitening in LLM: aclanthology.org/2026.findin…
7
2. RIFT-Bench: Dynamic Red-teaming For Agentic AI Systems: aclanthology.org/2026.findin…
7
Explore our papers:
1. AgenTRIM: Tool Risk Mitigation for Agentic AI: aclanthology.org/2026.findin…
18
How can businesses move beyond correlation to understand what truly drives outcomes?
In our latest Fujitsu Blog article, Cristiano Bellucci explores how Fujitsu Causal AI can help organizations turn complex data into actionable and explainable insights.The article also explores how Causal AI, together with Graph AI, can uncover cause-and-effect relationships and support more confident decision-making across healthcare, manufacturing, retail, energy, and other industries. These insights also build on research developed through Fujitsu’s collaboration with Business Advantage.
Discover how Fujitsu is working to move AI from identifying what is happening to understanding why: corporate-blog.global.fujits…
#FujitsuBlog #FujitsuCausalAI #CausalAI #GraphAI #AI
1
125
An AI-generated patch can look convincing. But how do we verify that it works?
At ASE 2026 in Munich, executable evidence, agent harnesses and trustworthy AI for software engineering were recurring themes. Mehdi Bahrami from Fujitsu Research of America presented Kozuchi Agent in the Industry Showcase, sharing a language-agnostic, open-weight approach to software repair.
Highlights from the report:
✅ 374 of 500 SWE-bench Verified instances resolved (74.80%) using a locally hosted Qwen3.5-27B model, with no fine-tuning and test-based selection across eight candidate runs (TTS@8).
✅ 41 of 128 Java instances resolved (32.03%) on Multi-SWE-bench using the same 27B model and eight-phase workflow.
✅ Related presentations explored how agent planning, coordination and execution feedback can support software repair.
What does it take to move from a generated diff to a test-backed patch?
Read the full tech blog: blog-en.fltech.dev/entry/202…
#FujitsuTechBlog #KozuchiAgent #AgenticAI #ASE2026 #ACM #SoftwareEngineering
1
2
182
OneComp v1.4のアップデート内容を、デモ付きで紹介する動画を公開しました！
OneCompは、富士通が研究開発を行っている生成AI圧縮のためのOSSです。
今回の動画では、量子化後のモデル精度回復機能の強化やNPU対応など、OneComp v1.4の新機能をご紹介します。
✅ OneComp v1.4のリリースハイライト
✅ MoEモデルの量子化精度を改善する「Router Fine-tuning」
✅ 独自量子化フォーマット「MDBF」へのGlobal PTQ対応
✅ OpenVINO IR形式へのエクスポートによるNPU推論への対応
✅ OpenVINOへのエクスポートと推論の実際のデモ
✅ Router Fine-tuning / MDBF Global PTQを使った量子化後の精度改善デモ
▶ 動画はこちら
youtu.be/JlWCyoW41fw
#OneComp #OneCompression #量子化 #LLM #MoE #MDBF #OpenVINO @akihiro_ysd
Leadership is changing with AI. What does that mean for the people leading organisations?
Fujitsu is co-hosting Leadership Models for the AI Era with the EIT Campus on 11 November from 11:00–12:00 CET.
The webinar will explore how AI is changing leadership, how people and AI can work together effectively, and why women and diverse perspectives matter in shaping responsible AI.
Katri Peälä from Fujitsu will join speakers from HR Nomad, the EIT Campus and Beam Institute of Technology for the discussion, moderated by Climate KIC.
Register here: bit.ly/4hHwhKn
112
複雑な帳票チェックを、AIでもっと効率的に。
調達業務では、社外から届く見積書のフォーマットが企業ごとに異なり、紙やPDFも混在するため、帳票チェックを人手に頼らざるを得ませんでした。
例えば富士通の調達部門では、誤発注を防ぐため、数十項目のチェック表を使って二重、三重に確認。1日にチェックできる件数にも限りがありました。
そこで研究チームが開発したのが、AIが人に代わって帳票をチェックする「帳票チェック支援AI」です。
✅ 既存のチェック表と帳票から、チェック結果を自動出力
✅ 汎用LLMと比べ、NG案件の見逃しを大幅に削減
✅ 社内実証では、約50％の工数削減が見込まれると評価
AIで帳票チェック業務をどう変えられるのか。詳しくはtech blogをご覧ください：
blog.fltech.dev/entry/2026/0…
#FujitsuTechBlog #AI #業務自動化
115
Fujitsu’s survey project to establish a next-generation digital infrastructure business framework in Singapore has been selected for METI’s Global South Future-Oriented Co-creation Project Subsidy, paving the way for trials of “FUJITSU-MONAKA” in Singapore toward digital infrastructure commercialization. global.fujitsu/en-global/pr/…
146
“The future of Agentic AI needs more than powerful models. It needs infrastructure we can trust.”
This was one of the key messages highlighted during Dr. Priyanka Sharma’s keynote, “Architecting Sovereign Infrastructure for Agentic AI,” at the IEEE Agentic AI Summit at IISc Bangalore.
As AI evolves from generating responses to reasoning, using tools and executing workflows autonomously, the infrastructure behind it becomes increasingly critical.
The keynote brought a systems perspective to five interconnected priorities: compute scalability, energy efficiency, security, governance and sovereignty.
Using FUJITSU-MONAKA as a practical case study, Dr. Sharma explored how hardware capabilities, confidential computing and an open-source software ecosystem can come together to create a trusted foundation for autonomous intelligence.
Because advancing Agentic AI is not only about what AI can do. It is also about ensuring that what it does is secure, scalable and accountable.
Learn more about FUJITSU-MONAKA: global.fujitsu/en-global/tec…
#Fujitsu #FujitsuMONAKA #AgenticAI #IISc
2
208
Advancing Disinformation Countermeasures at MisD 2026
How can we improve fact-checking during disasters when misinformation spreads rapidly on social media? At the 2nd Workshop on Misinformation Detection in the Era of LLMs (MisD 2026), co-located with ICWSM 2026 in Los Angeles, our researcher Okumura presented their work:“DTCD-AFC: Disaster-Type Classification Dataset designed for Automated Fact-Checking.”
The research introduces DTCD-AFC (Disaster-Type Classification Dataset designed for Automated Fact-Checking), a new dataset of 18,082 labeled text-image pairs designed to support automated fact-checking. By helping AI identify disaster types in social media posts and connect them with reliable evidence sources, DTCD-AFC can contribute to more accurate and efficient fact-checking during disasters.
In the blog, Okumura also shares highlights from other presentations and key takeaways from the conference.
Want to learn more about our research and key takeaways from the conference? Read the full tech blog: blog-en.fltech.dev/entry/202…
#MisD2026 #ICWSM2026 #FactChecking #Disinformation
100
Fujitsu is proud to sponsor Quantum Taiwan 2026! Join us at the Taipei International Convention Center (TICC) on November 10 - 11 for two days of discussions and insights into the evolving quantum ecosystem and global quantum supply chain, alongside leading researchers and industry experts from around the world. Don't miss the Industry Talk by Shintaro Sato, Fujitsu Fellow and Head of Quantum at Fujitsu, who will share Fujitsu's perspectives on advancing quantum computing and its potential.
📅 Nov 10 (Tue) – Nov 11 (Wed), 2026
📌 TICC, Taipei
🎟️ Exhibition, 1F (free): forms.gle/ePDVLqhGCsDTF1AA9
🎟️ Conference, 2F (paid): pse.is/7z58qw
🌐 Full agenda: 2026.quantumtaiwan.org/
📍 Visit Fujitsu at Booth A03 to Discover Fujitsu’s latest quantum computing initiatives and explore OpenQARP, our open-source software package designed to accelerate quantum computing research and application development.
OpenQARP offers:
✅ ~70% reduction in code volume
✅ 70+ composable building blocks
✅ 20+ ready-to-run algorithms
✅ One codebase that runs from your laptop to GPU systems and HPC simulators
✅ Built for applications in chemistry, materials discovery, drug discovery, manufacturing, finance, and supply chain
Whether you're exploring quantum algorithms or developing practical applications, come talk with our team and discover what you can build with OpenQARP.
See you at Booth A03!
#QuantumTaiwan2026 #FujitsuQuantum #OpenQARP #OpenSource #TICC
ALT Quantum Taiwan 2026
1
170
What if we could turn what’s in the air into valuable resources? The material system behind the 2025 Nobel Prize in Chemistry, known as MOFs (Metal-Organic Frameworks), suggests we can.
MOFs can selectively capture molecules such as CO₂ and water from the air, turning them into valuable resources.
In this tech blog, Hiroyoshi Ohtsu from Fujitsu's Advanced Materials Research Center explains:
🔹 What makes MOFs unique
🔹 Why they were recognized by the Nobel Prize
🔹 How MOFs may contribute to a more sustainable future
MOFs are also well suited to Materials Informatics (MI), which combine databases and calculations to narrow down candidate materials with desired functions, a task previously reliant on intuition. In addition, due to the difficulty of material exploration caused by combinatorial explosion, MOFs are also targets for material exploration using quantum computers.
Learn about MOFs and discover how they connect to Fujitsu's research in Materials Informatics and quantum computing:
English: blog-en.fltech.dev/entry/202…
Japanese: blog.fltech.dev/entry/2025/1…
#FujitsuTechBlog #MOFs #MaterialsInformatics
1
145