From 509820f6dc12d12a88f2c3c16f309cf4e42c0678 Mon Sep 17 00:00:00 2001 From: Fan Pu Zeng Date: Tue, 22 Oct 2024 22:07:30 -0400 Subject: [PATCH] add summary --- ...ge-intensive-multi-step-questions.markdown | 29 ++++++++++++++++--- 1 file changed, 25 insertions(+), 4 deletions(-) diff --git a/_summaries/2024-10-03-interleaving-retrieval-with-chain-of-thought-reasoning-for-knowledge-intensive-multi-step-questions.markdown b/_summaries/2024-10-03-interleaving-retrieval-with-chain-of-thought-reasoning-for-knowledge-intensive-multi-step-questions.markdown index d50a1caf..bd897b8b 100644 --- a/_summaries/2024-10-03-interleaving-retrieval-with-chain-of-thought-reasoning-for-knowledge-intensive-multi-step-questions.markdown +++ b/_summaries/2024-10-03-interleaving-retrieval-with-chain-of-thought-reasoning-for-knowledge-intensive-multi-step-questions.markdown @@ -9,9 +9,30 @@ bib_id: 2212.10509v2 #### 1. Interleaved Retrieval guided by Chain-of-Thought -Overall idea seems pretty similar to the -[Enhancing Retrieval-Augmented Large Language Models with Iterative Retrieval-Generation Synergy]({% link _summaries/2024-10-03-enhancing-retrieval-augmented-large-language-models-with-iterative-retrieval-generation-synergy.markdown %}) -paper, except that instead of having a pre-defined number of iterations, the CoT-guided LLM decides when to stop. +The paper's insight is to use CoT to guide retrieval, and use the retrieved +contents to then guide CoT again. + +This is done as follows: + +- Generate one sentence of CoT +- Use CoT sentence to retrieve additional piece of context +- Using new context, repeat the previous steps until answer is provided, or reached max number of steps + +The retrieved context is ordered randomly at each step. As the LLM may output +multiple sentences of CoT each time, they just take one newly generated sentence +and drop the rest. + +Here's the overall structure of the prompt: + +``` +Wikipedia Title: + +... +Wikipedia Title: + +Q: +A: ... +``` {% include figure.html path="/assets/img/summaries/ir_cot_workflow.webp" @@ -40,4 +61,4 @@ Marginal novelty given other techniques like self-ask that came previously. ### Conclusions for Future Work Could use a LLM to drive querying of RAG datastores guided by CoT to resolve -complex queries. \ No newline at end of file +complex queries.