0:00
/
0:00
Transcript

"Inference Scaling for Bridging Retrieval and Augmented Generation"

Generated below podcast on this paper with Google's Illuminate.

Smart passage shuffling reveals true importance, making RAG more reliable.

MoI (Mixture-of-Intervention) fixes position bias in RAG systems by reordering retrieved passages based on their true utility, improving answer quality by 7 points on benchmarks[1].

By understanding position bias, MoI helps LLMs see past the order of information.

-----

https://arxiv.org/abs/2412.10684

🤔 Original Problem:

→ RAG systems suffer from generator bias where better retrieval can actually hurt performance[1].

→ Current reranking approaches like RankGPT fail to improve RAG despite better retrieval quality[1].

-----

🔧 Solution in this Paper:

→ MoI observes how passages perform in different positions through multiple forward passes[1].

→ It separates passage utility from position bias using parallel observations[1].

→ The method aggregates outcomes from different permutations to estimate true passage importance[1].

→ MoI leverages retriever's prior knowledge to reduce computational costs[1].

-----

💡 Key Insights:

→ Position bias makes LLMs weigh passages differently based on their order[1]

→ Sometimes downranking relevant passages can improve overall performance[1]

→ Larger models like LLaMA-3 70B still suffer from position bias[1]

-----

📊 Results:

→ Improved ROUGE-L on MS MARCO by 7 points (44.30 vs 37.75 baseline)[1]

→ Boosted HotpotQA Exact Match score by 7 points (55.67 vs 48.54)[1]

→ Achieved 90% cost savings while maintaining 50% performance gains[1]

Discussion about this video