LLM-INSTRUCT tops UZH argument-mining task using 8B open models

A team from the University of Zurich just posted a result worth pausing on, less for the leaderboard win itself and more for what it says about how to get useful structured work out of small models. Their system, LLM-INSTRUCT, took first place overall on the UZH Shared Task at ArgMining 2026, a paragraph-level argument mining challenge on UN and UNESCO resolutions, using only open-weight models capped at 8 billion parameters.

The task is not a toy. Systems have to classify each paragraph, pick from a subset of 141 official tags, and predict directed relations between paragraphs, all conforming to a strict JSON schema. That kind of structured output is where general-purpose LLMs tend to fall apart, especially at the small end. The authors' framing is that the trick is not scaling the model, it is shrinking the decision the model has to make before generation begins.

Their pipeline does that in stages. Metadata-aware dense retrieval narrows the candidate tag space. Constrained decoding then applies per-dimension caps. Only the uncertain cases get escalated to a three-agent debate branch, and a schema validation step catches malformed output at the end. On the official leaderboard the system ranked 1st overall, with 1st in F1 and 5th in LLM-as-a-Judge. During development, configuration search pushed Task 1b Micro-F1 from 35.83% to 40.08% while the internal Task 2 score held at 4.421.

The honest caveat is that a shared-task win is a controlled evaluation, not proof the same recipe holds on messier real-world policy corpora, and the paper does not spell out which specific 8B open-weight models were used or what the debate branch costs in latency. The gap between the top F1 and the mid-pack LLM-as-a-Judge ranking is also worth watching if you care about human-perceived quality over exact-match metrics.

What is generalisable is the scaffolding pattern. For any team building structured extraction over a fixed vocabulary, retrieval-plus-constraint plumbing around a small open model may buy more accuracy per dollar than reaching for a bigger one.