Beyond Citations: Why AI Truthfulness Lies in the Traversal Path, Not Just the Evidence
Have you ever had a very knowledgeable friend explain a subtle physics phenomenon and confidently list several reference books — only for you to discover that while the books do mention the phenomenon, the way your friend derived the conclusion is completely different? This pattern of "correct citations, crooked reasoning" is evolving into a sophisticated form of hallucination in the world of large language models.
For a long time, we have judged whether an AI is "confidently making things up" primarily by its citations. As long as its claims can be located in the cited documents, we considered it honest. But a recent arXiv paper — "Why Neighborhoods Matter: Traversal Context and Provenance in Agentic GraphRAG" — reveals a surprising truth: the secret of AI truthfulness lies not in the final list of citations, but in the "neighborhood footprints" the model traverses while searching for an answer.
The "Neighborhoods Matter" Effect
Feynman once said: "What I cannot create, I do not understand." By the same logic, if an AI merely "pastes" a piece of evidence at the last moment, it has not necessarily understood the evidence's context.
The paper studies Agentic GraphRAG: letting an AI act like an explorer navigating a vast map of "entities" and "relations" (a knowledge graph) in search of answers.
The authors observed something deeply counterintuitive:
Even if you remove the core documents the AI ultimately cites, it may still produce a correct answer.
Strange, right? How can it answer correctly without "reading the book"? The answer lies in the neighbors it "passed by but never cited."
The AI's "Subconscious" Map
Breaking this down with Feynman's logic:
1. Scenery passed along the way (visited-but-uncited): When the AI searches for "Zhang San's occupation," it may traverse "Zhang San's university" and "Zhang San's advisor." Although the final answer only cites "Zhang San's onboarding announcement," those background details already helped it eliminate wrong distractors subconsciously. 2. Structural hints: The topology of the knowledge graph itself is a hint. What a node's "neighbors" are shapes how the AI interprets that node.
The paper demonstrates through ablation studies that forcibly erasing the traversal context — showing the AI only the final cited evidence — causes reasoning accuracy to collapse like an avalanche.
Why This Matters
The paper exposes a huge blind spot in current AI auditing: we audit the result (citations), not the process (footprints).
It is like a student who cheats on an exam: they list the correct bibliography, but actually assembled the answer from snippets overheard in the hallway. Such answers are unreliable because they lack genuine logical support.
The paper calls for future AI systems to provide not just a "reading list" but a complete "provenance trajectory" — a map of the entire reasoning expedition.
Summary
The truth of a fact depends not only on the evidence itself, but on the environment in which that evidence sits.
This paper tells us that an AI agent's honesty is not a static state but a dynamic journey. Along that journey, the unmentioned "neighbors" that were merely passed by are the unsung guardians of truth.
The next time an AI presents a long, perfect list of citations, don't award full marks so quickly. Instead, ask it: "Besides these, whom did you meet along the way?"
Truth is not in the pages of the book — truth is on the path leading to those pages. That is the deepest reflection on "evidence and context" that graph retrieval theory offers.