Multi-Hop Reasoning

Multi-hop reasoning is the ability to answer questions that require connecting facts from multiple documents. This is where standard RAG fails and where Vrin’s knowledge graph provides a structural advantage.

​

The Problem

Consider this question:

“Who is the CEO of the company that had the highest revenue growth last quarter?”

Answering this requires three hops:

1. Find all companies with revenue data from last quarter
2. Compare their growth rates to find the highest
3. Find who leads that company

No single document contains the full answer. Standard vector RAG retrieves chunks independently by cosine similarity to the query, but the intermediate facts (“which company had the highest growth”) aren’t in the query at all.

​

Why Transformers Struggle

Even with the entire context in the LLM’s window, transformers have an architectural ceiling on multi-hop reasoning. Research shows accuracy degrades significantly beyond 2-3 hops, because attention must implicitly chain multiple facts without explicit structure.

​

How Vrin Solves It

Vrin offloads multi-hop traversal to the knowledge graph:

1. Query decomposition: Complex queries are broken into sub-questions
2. Graph traversal: Each sub-question triggers Personalized PageRank from matched entities, following typed edges across hops
3. Iterative retrieval: Results from one hop inform the next retrieval step
4. Confidence scoring: Each hop’s results are scored, and low-confidence paths are pruned

The graph traversal is deterministic. It doesn’t rely on attention weights or context window position. A 5-hop query works as reliably as a 2-hop query.

​

Benchmark Results

On the MuSiQue benchmark (2-4 hop composed questions): Vrin achieves 27% improvement over baseline and outperforms the current state-of-the-art academic system on this benchmark.