Literature Survey: Re-ranking Methods

This section contains detailed surveys and analyses of individual papers related to Stage 2 re-ranking methods, including cross-encoders, late interaction models, and LLM-based rerankers.

Papers:

• MUVERA: Multi-Vector Retrieval via Fixed Dimensional Encodings
  • Overview
  • Problem Statement
  • Core Innovation: Fixed Dimensional Encodings (FDEs)
  • FDE Construction Algorithm
  • Theoretical Guarantees
  • Asymmetric Encoding Strategy
  • Two-Stage Retrieval Pipeline
  • Empirical Results
  • Implementation Details
  • Connections to Existing Work
  • Integration into RAG Pipelines
  • Practical Considerations
  • Open Questions and Future Directions
  • Code and Resources
  • Key Takeaways
  • Citation
  • Related Papers in This Repository
  • Connections to Advanced Retrieval Concepts
• zELO: ELO-inspired Training Method for Rerankers and Embedding Models
  • Paper Information
  • Abstract
  • Motivation: The Laffer Curve Problem
  • The zELO Method
  • Sparse Matrix Subsampling
  • Training Details
  • Results
  • Key Contributions
  • Practical Applications
  • Model Availability
  • References

Overview

The papers in this section provide in-depth technical analysis of key contributions to the re-ranking literature. Each survey includes:

• Problem formulation and mathematical foundations

• Algorithmic innovations with theoretical guarantees

• Empirical results on standard benchmarks

• Practical considerations for deployment

• Connections to other methods in the retrieval-reranking pipeline

Featured Papers

MUVERA: Multi-Vector Retrieval via Fixed Dimensional Encodings (NeurIPS 2024)

A principled approach to reduce multi-vector similarity search to single-vector MIPS, achieving 10% improved recall with 90% lower latency compared to prior state-of-the-art. Enables ColBERT-quality retrieval at production scale.

Zelo: Addressing the Laffer Curve in Hard Negative Mining

Introduces a theoretical framework identifying the Laffer curve relationship between hard negative miner intelligence and model performance. Reveals that increasingly sophisticated miners eventually degrade performance by selecting false negatives that are more relevant than human-annotated positives.

Contributing

To add a new paper survey to this section:

1. Create a new .rst file following the structure of existing surveys

2. Include: problem statement, core innovation, theoretical analysis, empirical results

3. Add the file to the toctree above

4. Ensure proper citations and links to related papers