"What External Knowledge is Preferred by LLMs? Characterizing and Exploring Chain of Evidence in Imperfect Context"

LLMs think better with connected evidence chains

Chain of Evidence (CoE) helps LLMs make better decisions by ensuring knowledge pieces are both relevant and mutually supportive, like evidence in a criminal case.

-----

https://arxiv.org/abs/2412.12632

🔍 Original Problem:

→ LLMs struggle with outdated knowledge and hallucinations when using external information

→ Current retrieval methods don't effectively handle irrelevant or misleading information

→ Complex queries requiring multiple knowledge pieces are particularly challenging

-----

🛠️ Solution in this Paper:

→ Introduces Chain of Evidence (CoE), inspired by criminal law principles

→ CoE requires knowledge pieces to show both relevance to the question and mutual support

→ Developed automated CoE discrimination approach to identify valid knowledge chains

→ Created ScopeCoE, a retrieval strategy that selects minimal sets of knowledge forming CoE

-----

💡 Key Insights:

→ LLMs show higher accuracy with CoE-structured knowledge

→ CoE helps resist misinformation and knowledge conflicts

→ LLMs exhibit strong faithfulness to CoE, even with factual errors

→ Fewer but well-connected knowledge pieces outperform larger sets

-----

📊 Results:

→ ScopeCoE improved accuracy by 10.4% on HotpotQA

→ Achieved 28.7% improvement on 2WikiMultihopQA

→ Required only 4.6-4.8 knowledge pieces vs standard 5 pieces

→ Maintained 85.4% faithfulness rate across tested models