New Framework Uses Multiple AI Agents to Improve Evidence Retrieval for Complex Questions

Answering intricate questions often requires piecing together information from multiple sources—a challenge known as multi-hop question answering. Researchers have introduced PRISM, a system that harnesses large language models in an agent-based framework to gather the right evidence while minimizing noise.

How PRISM Works#

The system divides the retrieval task across three specialized agents operating in a coordinated loop:

• Question Analyzer: Takes a complex query and breaks it down into simpler sub-questions that can be tackled independently
• Selector: For each sub-question, identifies the most relevant available context, prioritizing precision to avoid irrelevant material
• Adder: Reviews what information may be missing and retrieves additional evidence to improve recall

This division of labor lets the agents iterate back and forth, refining the evidence set until it becomes both compact and comprehensive. Rather than dumping all available context onto a downstream model, PRISM filters aggressively while ensuring nothing crucial gets overlooked.

Why This Matters#

Typical retrieval approaches either chase completeness at the cost of including distracting content, or focus narrowly and risk missing key pieces. PRISM's balanced approach allows the final question-answering model to achieve better accuracy while using significantly fewer irrelevant passages. This reduces computational overhead and helps models stay focused on what matters.

Validation and Performance#

The team tested PRISM on four established benchmarks: HotpotQA, 2WikiMultiHopQA, MuSiQue, and MultiHopRAG. Across all four, the system outperformed existing strong baselines, demonstrating that the multi-agent approach generalizes well to different types of complex questions.