tech
December 3, 2025
What We Learned at SIGIR 2025
SIGIR (Special Interest Group on Information Retrieval) is a top-tier information retrieval conference bringing together researchers, developers, industry experts, and educators from across the globe to share the latest ground-breaking research. Jina AI was at this year’s conference in Padua in July, presenting our work on late chunking at the Robust IR Workshop.
TL;DR
- Jina AI presented their 'late chunking' method at SIGIR 2025's Robust IR workshop, improving text retrieval by applying chunking after embedding.
- SIGIR 2025 highlighted research in reranking, sparse retrieval, and LLM integration for information retrieval.
- Keynotes included Stephen Robertson on BM25 and Iryna Gurevych on AI in scientific research.
- CLIP-AdaM was presented for open-set 3D object retrieval using multi-view CLIP models.
- A framework for 'compound retrieval systems' was proposed to optimize the combination of multiple rerankers.
- RE-AdaptIR suggests using weight differences from fine-tuned models to improve embeddings for new domains.
- Evaluations of LLM-based relevance judgment methods indicate binary judgments and pairwise comparisons perform best.
- Research explored the interplay of LLMs as rankers, judges, and assistants in IR evaluation, noting potential biases.
- A distinction was made between relevance and usefulness in search results, with LLMs showing alignment with human judgments on usefulness.
- Limitations of LLM-based relevance assessment were discussed, including insufficient evidence, vulnerability to manipulation, bias, and overfitting risks.
Continue reading
the original article