The grounding problem
A model can fluently state facts that are simply wrong. Grounding means tying every claim in the answer to evidence the agent actually retrieved, such as a search snippet or a database row.
How grounding works
- The agent calls a tool and gets back concrete data.
- It quotes or cites the relevant fields in its reasoning.
- The final answer references only what the tool returned.
When the model sticks to retrieved text, it cannot drift into invented details as easily.
Reinforcing it
- Instruction: tell the model to answer only from tool results and to say it does not know otherwise.
- Citation: require a source tag next to each claim.
- Verification: a second pass checks that each sentence maps to a returned snippet.
Why it fails
If a tool returns nothing useful, a weak agent fills the gap with a guess. Strong grounding makes the agent admit the gap and either call another tool or report uncertainty rather than fabricate.
Key idea
Tool result grounding anchors each claim to data the agent actually retrieved, using instructions citations and verification so the agent reports uncertainty instead of inventing facts when evidence is missing.