📝 FIRST: Faster Improved Listwise Reranking with Single Token Decoding
ft. Revanth Gangi Reddy, JaeHyeok Doo, Yifei Xu, Deevya Swain, and IBM Research
We are excited to introduce FIRST, in collaboration with IBM Research!
Paper: https://arxiv.org/pdf/2406.15657 (under review)
Our novel LLM reranking approach boosts efficiency by 50% while maintaining performance.
“FIRST leverages the output logits of the first generated identifier to obtain the entire ranking, streamlining the process compared to traditional listwise LLM reranking which outputs the ranking order as a generation sequence. We believe FIRST marks a significant step forward, and we hope it paves the way for future advancements in improving both the performance and practicality of LLM rerankers in real-world applications.”
¹
- Yifei Xu¹, student researcher
This research is based on work supported by U.S. DARPA KAIROS Program
No. FA8750-19-2-1004, and the Molecule Maker Lab Institute: an AI research institute program supported by NSF under award No. 2019897 and No. 2034562. Compute resources utlized on this project (Delta, ICCP) were provided through Lapis Labs.