The LLM-INSTRUCT system, developed by researchers at UZH, won the UZH Shared Task at ArgMining 2026 for its approach to paragraph-level argument mining in United Nations and UNESCO resolutions. The system achieved first place in F1 score by employing constraint-aware retrieval and selective debate among agents to classify paragraph types, predict tags, and identify relations within a strict JSON schema. This method, which reduces the decision space before generation, improved accuracy and submission robustness, with a final Task 1b Micro-F1 score of 40.08%. AI
IMPACT Demonstrates advanced techniques for structured prediction in NLP, potentially improving argument mining in legal and policy documents.
RANK_REASON The item describes a research paper detailing a system that won a shared task, which is a form of academic research output. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
