PAPT
A post-training method for LLM-based graph agents that introduces evidence-source annotations into training signals, enabling the model to learn whether different paths point to the same underlying source and thereby avoid redundant exploration of the same evidence.
Timeline (last 90 days)
The paper proposes a source-aware post-training method (PAPT) that makes agents aware of the provenance of evidence during training.
On synthetic data, PAPT reduces repeated visits and improves accuracy, but covers fewer distinct sources.
On real scientific claim tasks, PAPT consistently reduces repetition rates but accuracy declines.
Training signals designed for controlled synthetic environments often fail to generalize to the complex evidence networks found in real-world data.
All Facts (4)
The paper proposes a source-aware post-training method (PAPT) that makes agents aware of the provenance of evidence during training.
50%UnverifiedOn synthetic data, PAPT reduces repeated visits and improves accuracy, but covers fewer distinct sources.
50%UnverifiedOn real scientific claim tasks, PAPT consistently reduces repetition rates but accuracy declines.
50%UnverifiedTraining signals designed for controlled synthetic environments often fail to generalize to the complex evidence networks found in real-world data.
50%