From Knowledge Access to Source Learning: Developing Source-Specific Competence
The paper frames repeated source use as something an agent can learn from, not just retrieve from.
The authors propose SourceLearn, a system that builds a persistent model of an authoritative source and updates it over time. It uses self-directed passes to find weak spots in its understanding, plus task-guided feedback from downstream failures or recurring needs. In tests across five benchmarks and three LLM backends, it led 13 of 15 settings, with gains of up to 22.6 points over Hybrid RAG. ArXiv · AI/CL/LG's note
The authors propose SourceLearn, a system that builds a persistent model of an authoritative source and updates it over time. It uses self-directed passes to find weak spots in its understanding, plus task-guided feedback from downstream failures or recurring needs. In tests across five benchmarks and three LLM backends, it led 13 of 15 settings, with gains of up to 22.6 points over Hybrid RAG. ArXiv · AI/CL/LG's note
score 4