Skip to main content

# LLM Reranking

2 related articles

Subscribe to updates

Engineering at Meta Nov 11, 2026
Facebook recommendations stop guessing from clicks alone: Your algorithm lets people rewrite Feed and Reels in a sentence
Meta introduced Your algorithm, which lets Facebook users state in natural language what they want and do not want to see, with those preferences applied to both Feed and Reels. Previously the recommendation systems learned mainly from implicit engagement signals, which scale well but adapt slowly and capture little nuance. The system uses a collection of large language models to infer interests, validate preference changes and perform a final rerank, while a supplementary fetch and cache invalidation make changes take effect faster; Meta notes that interpreting ambiguous input and responding more precisely remain open problems.
Meta introduced Your algorithm, which lets Facebook users state in natural language what they want and do not want to see, with those preferences applied to both Feed and Reels. Previously the recommendation systems learned mainly from implicit engagement signals, which scale well but adapt slowly and capture little nuance. The system uses a collection of large language models to infer interests, validate preference changes and perform a final rerank, while a supplementary fetch and cache invalidation make changes take effect faster; Meta notes that interpreting ambiguous input and responding more precisely remain open problems.
Your algorithm LLM Reranking Preference-based recommendation
arXiv Jun 1, 2026
Whole-Pool Setwise Reranking with Long-Context LLMs: DualEnd Builds a Full 100-Candidate Ranking in 50 Comparisons, 59.4% Fewer Than Windowed HeapSort
The work introduces Whole-Pool Setwise reranking, in which each comparison ranks the entire candidate pool, and proposes DualEnd, which jointly selects the candidates predicted to be most and least relevant and fills the ranking from both ends, constructing a complete ranking of 100 candidates in 50 LLM comparisons; with nine open-weight LLMs on TREC DL19 and DL20 it requires 59.4% fewer comparisons than top-oriented windowed Setwise with heapsort and 88.8% fewer than with bubblesort, keeps nDCG@100 within 0.008 of a single-end whole-pool top-oriented approach while roughly halving token consumption and ranking time, and across six BEIR datasets reduces mean token consumption and ranking time by 49.4% and 50.8% relative to that single-end approach.
The work introduces Whole-Pool Setwise reranking, in which each comparison ranks the entire candidate pool, and proposes DualEnd, which jointly selects the candidates predicted to be most and least relevant and fills the ranking from both ends, constructing a complete ranking of 100 candidates in 50 LLM comparisons; with nine open-weight LLMs on TREC DL19 and DL20 it requires 59.4% fewer comparisons than top-oriented windowed Setwise with heapsort and 88.8% fewer than with bubblesort, keeps nDCG@100 within 0.008 of a single-end whole-pool top-oriented approach while roughly halving token consumption and ranking time, and across six BEIR datasets reduces mean token consumption and ranking time by 49.4% and 50.8% relative to that single-end approach.
The University of Queensland
Hang Li（First author）
Whole-Pool Setwise Reranking DualEnd LLM Reranking

You’re all caught up