Unsupervised Bootstrapping of G Function with LLM Based Rerankers

[Yikai-Liao]

Hi MFAR authors,

Thank you for sharing the "Multi-Field Adaptive Retrieval" (ICLR 2025) paper and code! Last day I visited your poster and had a talk.

The adaptive field weighting via the G function is a compelling approach for semi-structured retrieval.

Training G typically requires labeled query-document pairs, which may be scarce for new datasets. Could G be trained unsupervised using specialized rerankers, possibly iteratively?

Proposed Approach:

1. Initial Retrieval: Retrieve candidates using BM25 or a pre-trained dense retriever.
2. Reranking: Rerank candidates with a fine-tuned reranker (e.g., ~some 7B-scale LLM from MTEB leaderboards).
3. Pseudo-Labels: Use top-ranked documents as pseudo-positives and others as pseudo-negatives.
4. Training:
   • Initial: Train MFAR (including G) with contrastive loss on pseudo-labels.
   • Iterative: Use trained MFAR for better retrieval, rerank, and refine G.

Has you considered similar approach? What challenges might arise, such as pseudo-label diversity or iterative convergence, for semi-structured data?

Thanks for your thoughts!