LLM-INSTRUCT at UZH Shared Task 2026: Constraint-Aware Retrieval and Selective Debate for Paragraph-Level Argument Mining LLM-INSTRUCT在UZH 2026共享任务中的表现：面向段落级论证挖掘的约束感知检索与选择性辩论

TL;DR

• LLM-INSTRUCT won the UZH Shared Task 2026 on paragraph-level argument mining in UN/UNESCO resolutions using only open-weight models up to 8B parameters.
• The system employs a constrained structured prediction framework that combines metadata-aware dense retrieval with constrained decoding and per-dimension caps.
• A selective three-agent debate branch is utilized to resolve uncertain cases, improving robustness without applying heavy computation to all inputs.
• Performance optimization via configuration search raised Task 1b Micro-F1 from 35.83% to 40.08%, securing the top rank in F1 scores.
• The core technical insight is that reducing the decision space prior to generation significantly enhances both accuracy and submission reliability.

Why It Matters

This work demonstrates that high-performance complex information extraction tasks can be achieved with resource-efficient, open-weight models under 8B parameters, challenging the necessity for massive proprietary models in specialized NLP tasks. It provides a practical blueprint for handling strict schema constraints and multi-label classification in legal or diplomatic texts, which are critical for compliance and policy analysis. The approach offers a scalable alternative for industries requiring precise, structured outputs from large document corpora without prohibitive computational costs.

Technical Details

• Task Definition: The system addresses paragraph-type classification, prediction of a subset of 141 official tags, and directed relation prediction within a strict JSON schema setting.
• Constraint-Aware Retrieval: Metadata-aware dense retrieval is used first to narrow down the candidate tag space, effectively reducing the complexity of the subsequent generation step.
• Constrained Decoding: The model applies constrained decoding with per-dimension caps to ensure outputs adhere strictly to the required schema, minimizing invalid generations.
• Selective Debate Mechanism: Uncertain predictions are escalated to a three-agent debate branch, allowing for higher-confidence resolution only where needed, balancing cost and accuracy.
• Performance Metrics: The system ranked 1st overall and 1st in F1 on the official leaderboard, with internal development showing a significant F1 improvement through configuration search while maintaining stable Task 2 scores.

Industry Insight

Practitioners should prioritize pre-generation filtering and constraint enforcement over brute-force model scaling when dealing with highly structured output requirements. Implementing selective reasoning mechanisms, such as debate branches for uncertain cases, can optimize inference costs while maintaining high precision in critical applications. This architecture serves as a strong reference for building efficient, open-source solutions for legal, regulatory, and diplomatic text analysis tasks.

TL;DR

• LLM-INSTRUCT 在 ArgMining 2026 联合国决议段落级论证挖掘共享任务中夺冠，仅使用 8B 以下开源模型。
• 核心架构采用约束结构化预测框架，结合元数据感知的密集检索以缩小候选标签空间。
• 引入受限解码（带维度上限）与基于不确定性的三智能体辩论机制，并辅以输出模式验证。
• 通过配置搜索优化，Task 1b Micro-F1 从 35.83% 提升至 40.08%，整体排名 F1 第一。
• 关键结论表明：在生成前缩减决策空间能显著提升准确率及提交的鲁棒性。

为什么值得看

本文展示了在资源受限（<8B 参数）条件下，如何通过精细的工程策略（检索增强+约束解码+多智能体协作）超越单纯依赖大模型规模的传统路径。其“先缩小决策空间再生成”的方法论为复杂结构化信息抽取任务提供了极具参考价值的高效解决方案。

技术解析

• 任务定义与约束：针对联合国和 UNESCO 决议的段落级论证挖掘，要求执行段落类型分类、141个官方标签子集预测及有向关系预测，且必须在严格的 JSON 模式下运行，限制使用 8B 及以下开源权重模型。
• 检索增强缩小搜索空间：系统首先利用元数据感知的密集检索技术，从庞大的标签库中筛选出相关候选集，从而大幅降低后续生成的复杂度，这是提升性能的关键前置步骤。
• 受限解码与多智能体辩论：采用带有每维度上限的受限解码确保格式合规；对于置信度低的案例，系统将其升级至一个由三个智能体组成的辩论分支进行深度推理，其余情况直接输出，以此平衡计算成本与准确性。
• 结果优化：最终系统在官方排行榜上获得总分第一及 F1 分数第一，LLM-as-a-Judge 排名第五。开发阶段的配置搜索进一步将 Task 1b Micro-F1 提升了约 4.25 个百分点。

行业启示

• 小模型+强工程优于盲目堆料：在特定垂直领域或严格约束任务中，通过检索增强、约束解码和逻辑编排等工程手段，小型开源模型完全可以击败更大规模的通用模型，应重视推理流程的设计而非仅关注模型参数量。
• 结构化输出的可靠性依赖于前置过滤：在处理高维标签或多重关系预测时，先通过检索等手段缩小候选范围，比让模型直接在全量空间中生成更能保证结果的准确性和格式的正确性。
• 混合推理策略的价值：结合确定性方法（受限解码）与探索性方法（多智能体辩论），并根据不确定性动态路由，是一种兼顾效率与精度的有效范式，适用于对鲁棒性要求高的生产环境。