Dunamu said on Nov. 17 that three of its research papers were accepted at ICAIF 2026, the Association for Computing Machinery’s international conference on financial artificial intelligence. One paper was selected for the main conference and two were accepted for specialized workshops.
ICAIF 2026 was held at Bocconi University in Milan from Nov. 14 to Nov. 17 and drew more than 1,000 paper submissions—the highest on record. The main conference acceptance rate was about 22%. The event, organized by ACM, included participation from global financial and payments firms such as BlackRock, JPMorgan, Visa, Mastercard and Standard Chartered.
Dunamu’s main-conference paper on exchange wallet identification
The paper accepted to the main conference, “Classifying On-Chain Behavioral Characteristics of Ethereum-Based Centralized Exchange Wallets Using Proof of Reserves,” presented a machine-learning methodology for identifying exchange wallets. Researchers including Park Sung-chan of Dunamu’s quant team analyzed about 3.03 million Ethereum wallet addresses using proof-of-reserves data disclosed by 15 exchanges and derived 58 behavioral characteristics, including transaction patterns, token transfers and gas usage.
When tested on wallets from exchanges not used in training, the model recorded an average PR-AUC of 0.9594, a classification-performance metric in which a score closer to 1 indicates better performance. Dunamu said the findings confirmed the potential to identify previously unknown exchange wallets. Parts of the exchange wallet identification technology are already applied in Upbit Data Lab’s blockchain analytics service to analyze holdings by participant type, including exchanges, whales and foundations.
Workshop papers: asset classification and fund-tracing under incomplete records
One workshop paper, “Semantic Translation for an Institutional Digital Asset Classification System,” explored classifying digital assets by actual use through analysis of smart-contract code with a large language model. Researchers including Jeon Hyun-ho, head of Dunamu’s quant team, ran experiments on Ethereum-based DeFi protocols such as Uniswap, Aave and Curve. After translating code into natural language and classifying it by financial function, 16 of the 17 cases analyzed matched existing classification standards. A highlighted feature of the methodology is that humans can directly review and verify the basis for the AI’s classifications. Dunamu plans to apply the technique to enhance the Upbit Digital Asset Classification System (UBCS).
The other workshop paper, “Bounds on Estimating the Use of Deposited Funds Under Incomplete Transaction Records,” examined what information and conditions are needed to reliably analyze fund flows when transaction records are incomplete. Researchers including Kim Sung-yong, a project leader on Dunamu’s FDI team, and So Min-seop, head of the FDI team, identified key information supporting reliable source-of-funds analysis using Ethereum wallet transaction case studies and blockchain execution experiments. Dunamu said the research provides a technical basis for improving the analytical reliability of its on-chain fund-tracing solution, OTS, and will be used to advance AI-based fund tracing and financial-crime analysis technologies.
Kim Dae-hyun, Dunamu’s chief data officer, said the research confirmed the potential for AI to analyze blockchain data with greater precision and improve result reliability. He added that the company will continue research combining AI and blockchain technology to lead technological innovation in the digital-asset industry.

