LLM-INSTRUCT at UZH Shared Task 2026: Constraint-Aware Retrieval and Selective Debate for Paragraph-Level Argument Mining

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

LLM-INSTRUCT, the winning system for the UZH Shared Task at ArgMining 2026, tackles paragraph-level argument mining within UN and UNESCO resolutions. This complex task requires precise paragraph-type classification, the prediction of a subset from 141 official tags, and accurate directed relation prediction. All operations must strictly conform to a specified JSON schema and be performed using only open-weight models limited to 8B parameters. The LLM-INSTRUCT system approaches this as a constrained structured prediction problem. Its multi-stage methodology first narrows the candidate tag space using metadata-aware dense retrieval, then applies constrained decoding with per-dimension caps, and finally escalates only the most uncertain cases to a specialized three-agent debate branch for refined decision-making.

Key takeaway

For NLP Engineers developing argument mining systems under strict constraints, LLM-INSTRUCT demonstrates a robust approach. You should consider integrating metadata-aware dense retrieval to pre-filter candidate tags and implement constrained decoding to ensure output adheres to complex schemas. Furthermore, designing a multi-agent debate mechanism for ambiguous cases can significantly improve prediction accuracy and reliability, especially when working with open-weight models up to 8B parameters.

Key insights

The LLM-INSTRUCT system combines retrieval, constrained decoding, and multi-agent debate for accurate, constrained argument mining.

Principles

• Frame complex tasks as constrained structured prediction.
• Narrow candidate space early with metadata.
• Use multi-agent debate for uncertain cases.

Method

LLM-INSTRUCT performs metadata-aware dense retrieval to narrow tags, applies constrained decoding with per-dimension caps, and then uses a three-agent debate branch for uncertain predictions.

In practice

• Implement dense retrieval for tag filtering.
• Apply constrained decoding for schema adherence.
• Design multi-agent systems for ambiguity resolution.

Topics

• Argument Mining
• Constrained Structured Prediction
• LLM-INSTRUCT
• Dense Retrieval
• Multi-Agent Systems
• Open-Weight Models

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• From Argument Components to Graphs: A Multi-Agent Debate with Confidence Gating for Argument Relations — Multi-agent debate with confidence gating improves training-free argument relation identification and offers interpretability.
• The State of Information Retrieval in 2026 — Information retrieval has evolved into a reasoning problem, integrating LLMs and agentic search for complex, context-aware query processing.
• MADRAG: Multi-Agent Debate with Retrieval-Augmented Generation for Training-Free Analytic Essay Scoring — Multi-agent debate with retrieval augmentation offers a training-free, reliable approach to analytic essay scoring, outperforming prompt-based LLMs.
• CAF-Gen: A Multi-Agent System for Enriching Argumentation Structures — Multi-agent LLM systems with iterative Creator-Reviewer pipelines reliably enrich shallow argument structures into complex formal models.
• Argument Collapse: LLMs Flatten Long-Form Public Debate — LLMs flatten public debate by converging on fewer, more generalized arguments and fixed structures.
• ARMOR-MAD: Adaptive Routing for Heterogeneous Multi-Agent Debate in Large Language Model Reasoning — Adaptive control and genuine model heterogeneity are crucial for efficient and accurate multi-agent LLM reasoning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.