LLM-INSTRUCT wins ArgMining 2026 with constraint-aware debate
Open-weight models under 8B parameters beat larger systems by narrowing decision space...
Researchers from Vietnam introduced LLM-INSTRUCT, the winning system for the UZH Shared Task at ArgMining 2026, focused on paragraph-level argument mining in UN and UNESCO resolutions. The task requires paragraph-type classification, prediction of a subset of 141 official tags, and directed relation prediction under a strict JSON schema, all using only open-weight models up to 8B parameters. LLM-INSTRUCT frames this as constrained structured prediction: it first uses metadata-aware dense retrieval to narrow candidate tags, then applies constrained decoding with per-dimension caps to enforce limits, escalates only uncertain cases to a three-agent debate branch for consensus, and validates the final output schema.
On the official leaderboard, the system ranked 1st overall, with 1st in F1 and 5th in LLM-as-a-Judge. During development, configuration search further improved Task 1b Micro-F1 from 35.83% to 40.08% while keeping internal Task 2 score at 4.421. The main lesson is straightforward: reducing the decision space before generation improves both accuracy and submission robustness. This approach demonstrates that smaller, constrained models can outperform larger systems by focusing computational resources on narrowing choices first, rather than brute-force reasoning. Code and scripts are publicly available.
- Uses only open-weight models with up to 8B parameters, proving small models can win with smart architecture
- Metadata-aware dense retrieval reduces candidate tags from 141 to a smaller set before generation
- Three-agent selective debate handles uncertain cases, helping boost Task 1b Micro-F1 from 35.83% to 40.08%
Why It Matters
Shows that reducing decision space before generation can make small open models outperform larger ones in structured prediction tasks.