HUFS Prof. Junhyun Lee's Team Has 4 AI Papers Accepted Across ICLR, ICML, and NeurIPS 2026
A research team led by Professor Junhyun Lee at Hankuk University of Foreign Studies has had four papers accepted across ICLR, ICML, and NeurIPS 2026, introduci
On October 11, 2026, Hankuk University of Foreign Studies (HUFS) announced that four artificial intelligence research papers corresponding-authored by Professor Junhyun Lee from the Department of Computer Science and Engineering were officially accepted into the world's top three AI conferences: ICLR 2026 (one paper), ICML 2026 (one paper), and NeurIPS 2026 (two papers).
Image source: 한국외국어대학교 / 메트로신문
The accepted studies focus on core engineering challenges to enhance the efficiency and reliability of next-generation AI systems, ranging from multi-agent collaboration techniques that reduce communication bottlenecks between large language models (LLMs) to active reasoning evaluation and domain-preserving model compression for fields like healthcare and finance.
Direct LLM Representation Exchange Cuts Communication Latency by 11x and Enables Selective AI Debate
A major bottleneck in multi-agent workflows is the network latency and redundant compute costs caused by frequent inter-agent data transfers. The research team introduced two structural solutions to streamline collaboration:
- Direct Representation Exchange Between LLMs (NeurIPS 2026): In collaboration with LG AI Research, the University of Illinois Chicago (UIC), Capital One AI Foundations, and Chung-Ang University, the team developed a method for LLMs to directly exchange internal representations rather than relying on natural language text serialization. Under specific experimental benchmark conditions, this reduced inter-agent communication latency to approximately one-eleventh of previous baselines.
- Selective Multi-Agent Debate (ICML 2026): Rather than enforcing compute-heavy multi-agent debate across every incoming prompt, the proposed adaptive mechanism initiates AI debate only when model confidence is low, significantly cutting unnecessary compute overhead and API call costs.
Active Problem Generation for Reasoning Evaluation and Domain-Preserving Model Compression
The accepted papers also address rigorous verification of model reasoning and practical compression techniques for domain-specific applications:
- Active Generation Reasoning Evaluation (ICLR 2026): Moving beyond static multiple-choice benchmark datasets, this evaluation framework prompts AI agents to actively formulate and verify questions, providing a more rigorous measure of multi-step reasoning abilities.
- Domain-Preserving Knowledge Compression (NeurIPS 2026): Addressing the loss of specialized domain knowledge during standard model compression, this technique prioritizes and protects mission-critical representations in high-stakes domains such as healthcare, law, and finance while pruning model parameters.
Benchmark Conditions and Implementation Availability
The reported performance gains represent milestone academic results, with production-level validation expected as the papers are presented:
- Benchmark Constraints: The 11x communication latency reduction and debate compute savings were measured under specific benchmark conditions configured by the researchers; practical savings in production environments will vary depending on network topology and third-party API infrastructure.
- Code and Proceeding Availability: Official arXiv preprints and open-source code repositories were not linked in the initial announcement; full algorithm implementations and benchmark suites are expected alongside the official conference proceedings.