AI BriefingKaLM-Reranker-V1 Released
Key point
A new high-performance Reranker model with a novel architecture that separates query and passage computation to improve computational efficiency has been released.
Details
Existing Encoder-Decoder or Encoder-only based Rerankers encode the query and passage simultaneously, which limits computational efficiency and flexibility.
KaLM-Reranker-V1 addresses this by adopting an FBNL (Fast but Not Late Interaction) architecture that separates the computation of queries and passages. It is based on an Encoder-Decoder architecture, where the Encoder pre-encodes passages via Matryoshka embedding pooling, and the Decoder models the query's intent before capturing the relevance between the two pieces of information through Cross-attention.
The model is provided in three sizes: Nano (0.27B), Small (1B), and Large (4B).
Key results are as follows:
- BEIR benchmark: Achieved SOTA (State-of-the-art) performance on par with the Qwen3-Reranker series
- MIRACL: Demonstrated strong performance despite not being focused on multilingual data training
- LMEB: The 0.27B Nano model showed performance competitive with Embedding models in the 7-12B range
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Why LLM Research Agents Generalize Without Benchmark Overfitting: 'Compressibility' Is Key