Back to research

Research Project

Single-Pass Document Scanning for Question Answering

Training state space models for long-context question answering that scan documents once and answer efficiently.

Co-author 2025 arXiv Preprint

Overview

This project studies retrieval-augmented question answering with self-trained Mamba-2 state space models. The goal is to answer questions over very long documents while reducing repeated document passes and inference cost.

What This Page Can Include

  • Research motivation and core problem.
  • Model or system diagram.
  • Dataset, training, and evaluation details.
  • Selected results and qualitative examples.
  • Links to paper, code, demos, or related talks.

Current Summary

We trained state space models for long-context question answering, reaching performance comparable to strong frontier baselines on extremely long documents while improving computational efficiency.