From Knowledge Access to Source Learning: Developing Source-Specific Competence

From Knowledge Access to Source Learning. Developing Source-Specific Competence. By Lucheng Fu and twelve others, from the Georgia Institute of Technology, the University of Illinois Urbana-Champaign, and one other institution. Published on the first of October. Its code repository is worked out: about fifty source files, with a readme and a licence.

Agents built on large language models increasingly lean on persistent external sources to work through sequences of knowledge-intensive tasks. Existing methods improve how a source's content is accessed and organised, and agent-memory systems keep reusable knowledge from earlier interactions. Yet repeated use of the same source is still mostly treated as repeated access. Each visit is a lookup, not a chance to understand the source better than the last time. An agent can consult the same manual again and again and come away from the latest visit no wiser about how the manual is put together.

The authors call the alternative source learning: building reusable, source-specific competence over a persistent and authoritative source. That competence lives in a persistent source model, which captures reusable understanding of the source, including how its knowledge is structured, how it is interpreted, and how it is applied. The model persists between tasks, so what was understood while solving one task is still there for the next.

SourceLearn constructs and progressively refines these models with two complementary mechanisms. Self-Directed Source Learning identifies what is still incompletely understood and adaptively revisits the source. Task-Guided Source Learning uses downstream experience to expose local gaps in representation, and recurring needs in how the source's knowledge should be organised.

One principle holds across both. Learning signals decide what should be reconsidered, but the persistent updates themselves are reconstructed from the authoritative source. A signal can say that something is missing or poorly organised. What gets written into the source model is drawn back from the source, not from the signal. The two mechanisms differ in where the signal comes from. One looks inward, at what remains unclear. The other looks outward, at what the tasks keep needing.

Testing spans five benchmarks and three language model backends, fifteen settings in all. SourceLearn achieves the best performance in thirteen of them, with gains of up to twenty-two point six points over Hybrid RAG. The authors also report substantial overall improvements over static source representations, and over memory baselines built on experience.

One way to hear the distinction: a memory system keeps what was learned along the way, while source learning keeps returning to the source itself. Experience only points at what to revisit, the source stays the authority for every update, and in most of the settings tested that pairing comes out on top.