Public articles linked to the same research event.
arXiv The work formulates source learning, represents reusable understanding of a persistent authoritative source as a persistent revisable source model, and refines it through Self-Directed Source Learning (an Inspect-Study-Consolidate cycle with adaptive Deepen or Connect actions) and Task-Guided Source Learning (failure-guided local refinement plus cross-task representation-policy recalibration), with persistent updates always reconstructed from the authoritative source; across five benchmarks and three LLM backends, SourceLearn is best in 13 of 15 settings, improving over Hybrid RAG by 14.3, 4.9, and 13.4 points on average, with gains up to 22.6 points.
The work formulates source learning, represents reusable understanding of a persistent authoritative source as a persistent revisable source model, and refines it through Self-Directed Source Learning (an Inspect-Study-Consolidate cycle with adaptive Deepen or Connect actions) and Task-Guided Source Learning (failure-guided local refinement plus cross-task representation-policy recalibration), with persistent updates always reconstructed from the authoritative source; across five benchmarks and three LLM backends, SourceLearn is best in 13 of 15 settings, improving over Hybrid RAG by 14.3, 4.9, and 13.4 points on average, with gains up to 22.6 points.
The work formulates source learning, represents reusable understanding of a persistent authoritative source as a persistent revisable source model, and refines it through Self-Directed Source Learning (an Inspect-Study-Consolidate cycle with adaptive Deepen or Connect actions) and Task-Guided Source Learning (failure-guided local refinement plus cross-task representation-policy recalibration), with persistent updates always reconstructed from the authoritative source; across five benchmarks and three LLM backends, SourceLearn is best in 13 of 15 settings, improving over Hybrid RAG by 14.3, 4.9, and 13.4 points on average, with gains up to 22.6 points.
The work formulates source learning, represents reusable understanding of a persistent authoritative source as a persistent revisable source model, and refines it through Self-Directed Source Learning (an Inspect-Study-Consolidate cycle with adaptive Deepen or Connect actions) and Task-Guided Source Learning (failure-guided local refinement plus cross-task representation-policy recalibration), with persistent updates always reconstructed from the authoritative source; across five benchmarks and three LLM backends, SourceLearn is best in 13 of 15 settings, improving over Hybrid RAG by 14.3, 4.9, and 13.4 points on average, with gains up to 22.6 points.