The VANGUARD team has developed IROH, a three-stage retrieval system that secured first place in the JOKER 2026 Track Task 1 English competition. This system combines hybrid retrieval methods, cross-encoder reranking, and an ensemble of Large Language Model judges. Key findings indicate that rationale-distilled judges are crucial for ranking quality, while structured hard negatives can harm generalization. Notably, a smaller, well-calibrated model like Qwen2.5-7B with generic rationales outperformed larger Gemma-4-31B configurations. AI
IMPACT This research demonstrates advanced LLM judge techniques for information retrieval tasks, potentially improving humor ranking systems.
RANK_REASON The cluster describes a research paper detailing a system that achieved first place in a specific competition track, including findings from ablation studies.
Read on arXiv cs.IR (Information Retrieval) →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
