A correct answer may have no retrieved support
The FoG paper separates two roles for a knowledge graph: supplying facts an agent needs and supplying evidence that makes an answer traceable. An agent can answer correctly without retrieving the supporting triples, especially when training rewards only the final answer.
That is a problem for a workflow that needs to inspect where a conclusion came from. Answer accuracy and evidence retrieval must therefore be measured separately, rather than treating one as proof of the other.
Build the evidence into training
FoG combines evidence-grounded data synthesis, a knowledge-graph interaction agent and an evidence-coverage process reward. The synthesis step recursively replaces explicit entities with verifiable subquestions, so each generated question carries supporting triples by construction.
During interaction, the agent records triples returned by the graph. The process reward credits each supporting triple to the turn that first retrieves it. That gives the training procedure a signal about gathering evidence along the way, rather than waiting to score only the final answer.
Evaluate coverage as well as accuracy
The authors report results on three knowledge-graph question-answering benchmarks. In those experiments, 3B–4B open-source agents outperform larger baselines in answer accuracy and evidence coverage, with zero-shot transfer across knowledge graphs.
Those are benchmark findings under the paper’s setup. They do not establish that every retrieved triple is true, that the graph is complete or that all real-world answers will be adequately supported. Coverage is measured against the evidence available for the benchmark questions.
For readers reproducing the work, the important comparison is whether an agent retrieves the needed support on the same questions where it answers correctly. Inspecting the first-retrieval reward and synthesized evidence chain can help explain the result. FoG’s contribution is to make support gathering part of the learning objective, rather than infer an audit trail from a plausible final response.
Comments