learn.ifs.name

Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection

Asai et al., 2024, Self-RAG (The Twelfth International Conference on Learning Representations (ICLR 2024))

Subscriber deck — free during early access

2 cards in “3.3 Self-RAG Inference” · whole deck

3.3 Self-RAG Inference

  1. How can retrieval frequency be adjusted at inference?
    With a threshold: retrieval is triggered when the probability of Retrieve = yes exceeds it. Factual tasks can use a low threshold (retrieve often); open-ended tasks like personal essays can retrieve rarely.
    • Asai et al., 2024, Self-RAG (The Twelfth International Conference on Learning Representations (ICLR 2024)), §3.3
  2. How does Self-RAG choose between candidate segments generated from different passages?
    Segment-level beam search scoring each candidate by its generation probability plus a weighted sum of the normalized probabilities of the desirable critique tokens (e.g. IsRel = relevant, IsSup = fully supported); the weights can be changed at test time, or undesirable critiques used as hard filters.
    • Asai et al., 2024, Self-RAG (The Twelfth International Conference on Learning Representations (ICLR 2024)), §3.3

Source

  • Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil and Hannaneh Hajishirzi. 2024. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. In The Twelfth International Conference on Learning Representations (ICLR 2024). https://arxiv.org/abs/2310.11511