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

1 Introduction

• Existing access methods and agent memories address different objects, leaving an opportunity to improve how a source itself is represented and used across tasks.Access methods support query-time use of external knowledge, while agent memory preserves knowledge derived from prior interactions.Source: §1 · p. 1
• Source learning develops reusable understanding of how source knowledge is structured, interpreted, and applied, complementing rather than replacing the authoritative source.The model can capture relationships, applicability conditions, procedures, and how information combines for future reasoning and action.Source: §1 · p. 2
• Source learning must identify broadly reusable knowledge before future tasks are known and use task feedback to repair source-understanding gaps rather than retain task-specific answers.A source may contain too much information to retain, and one-pass representations may miss distributed connections or implicit conditions.Source: §1 · p. 2
• SourceLearn maintains a persistent source model, revisiting incompletely understood content and using failures and recurring task demands to refine knowledge organization.Both mechanisms use observations to guide learning while reconstructing persistent updates from the authoritative source.Source: §1 · p. 2
• The contributions include a source-learning formulation and representation, two complementary learning mechanisms, and evaluation across document QA, code QA, and tool use.The listed evaluation scope also includes interactive environments.Source: §1 · pp. 2–3
• SourceLearn improves same-source task performance over task-local retrieval, static source representations, and experience-based memory, with gains from both learning mechanisms.The introduction reports results in interactive environments.Source: §1 · p. 3

2.1 External Knowledge Access and Source Representations

• Access and representation methods improve retrieval, long-context use, or source structure, while SourceLearn progressively refines a reusable model of a persistent authoritative source.Examples include retrieval approaches [1] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20], long-context methods [2] [21], hierarchical summaries [3], graph structures [4] [22] [23] [24], associative retrieval [5] [25], and task-conditioned representations [26].Source: §2.1 · p. 3

2.2 Agent Memory and Experiential Learning

• Agent-memory systems preserve interaction-derived knowledge, whereas SourceLearn uses task experience to reveal source-understanding gaps and reconstructs persistent updates from the authoritative source.Prior systems include conversational, episodic, and hierarchical memories [31] [32], as well as experience-learning and structured-memory approaches [7] [8] [9].Source: §2.2 · p. 3

3.1 Knowledge Access over a Persistent Source

• The formulation considers a persistent authoritative source and a sequence of tasks grounded in that same source.Source: §3.1 · p. 3
• Existing source-grounded systems apply a task-solving agent to source context made available for the task; that context may come from retrieval, long-context access, or structured representation.Source: §3.1 · p. 3
• Repeated access does not necessarily improve source understanding, because reusable knowledge may remain fragmented or need reconstruction across tasks.Source: §3.1 · p. 3

3.2 Source-Specific Competence

• Source-specific competence is reusable understanding that supports new source-grounded tasks and captures how source knowledge is structured, interpreted, and applied.Source: §3.2 · p. 4
• A persistent, revisable source model carries understanding from earlier study and use into future tasks.Source: §3.2 · p. 4
• Accumulated source understanding complements task-specific source information rather than replacing direct access to it.Source: §3.2 · p. 4
• The original source remains authoritative for factual content, and source learning aims to improve future same-source task performance through stronger reusable understanding.Source: §3.2 · p. 4

4.1 Overview

• SourceLearn carries forward an organized model of the same authoritative source, refining it through self-directed study and task-guided learning.The stages proceed from initial construction through self-directed learning to a final model refined using guidance tasks grounded in that source.Source: §4.1 · p. 4

4.2 Source Model Construction and Activation

• The source model organizes reusable entity knowledge and higher-level structure; at inference, activated model context complements direct retrieval of precise evidence.The model is compact rather than a reproduction of the source, which remains available as the authoritative factual reference.Source: §4.2 · pp. 4, 6, Figure 2 · p. 5

4.3 Self-Directed Source Learning

• Self-Directed Source Learning inspects the current model for gaps, adaptively deepens single-entity understanding or connects entities, then consolidates observations into the model.DEEPEN investigates an unresolved aspect of one entity; CONNECT studies two entities when an important dependency or shared structure remains insufficiently understood.Source: §4.3 · pp. 6–7

4.4 Task-Guided Source Learning

• Task-Guided Source Learning uses task requirements and failures to repair missing source understanding, while recurring representation lessons guide source-grounded recalibration.Reference outcomes are used to locate authoritative support and derive requirements; answers and supporting evidence are not stored as persistent knowledge.Source: §4.4 · pp. 7–8

5 Experiments

• The experiments ask whether same-source gains span sources and backends, how the learning mechanisms contribute, and whether learned understanding goes beyond static compression.These are the three questions guiding the evaluation.Source: §5 · p. 8

5.2 Main Results

• SourceLearn is best in 13/15 settings and second-best in one, beating Hybrid RAG by +14.3, +4.9, and +13.4 average points with GPT-5.6-Luna, gpt-oss-120b, and DeepSeek-V4.1-Flash.Table 1 reports task performance across document QA, code QA, tool use, and interactive environments, with changes relative to Hybrid RAG.Source: §5.2 · p. 8, Table 1 · p. 9

5.3 Ablation Study

• On two sources per QA benchmark, full MT is most accurate; removing either learning mechanism or either task-guided pathway lowers accuracy, and the pathways contribute different information.Failure-guided local refinement repairs local gaps, while cross-task representation learning changes how the whole source is represented.Source: Figure 3 · p. 9, §5.3 · p. 9

5.4 Source Learning Beyond Compression

• From M0 to MT, representations shift from isolated lookup content toward rules and procedures, while units with explicit applicability conditions rise from roughly 30% to 65–82%.The reported condition-unit percentages compare M0 with MT across sources.Source: §5.4 · p. 9, Figure 4 · p. 10
• Test-required claims represented in the model rise from 23.2% in M0 to 40.0% in MT; accuracy is associated with coverage, ranging from 67.5% with no required claims represented to 86.6% when all are represented.The coverage increase is reported across QA benchmarks; the accuracy comparison is between no required claims and all required claims represented.Source: Figure 4 · p. 10, §5.4 · p. 10

6 Conclusion

• SourceLearn develops reusable source-specific competence through a persistent, revisable source model and complementary learning mechanisms grounded in the original source.The conclusion distinguishes cumulative competence development from repeated source access or accumulation of task experience.Source: §6 · p. 10
• Across five benchmarks and three LLM backends, SourceLearn improves overall same-source task performance over task-local retrieval, static source representations, and experience-based memory.The conclusion also reports that analyses show greater explicitness of conditions, procedures, exceptions, and relations, and increased coverage of knowledge required by future tasks.Source: §6 · p. 10
• The study focuses on persistent, authoritative, relatively stable sources and repeated same-source tasks; evolving, noisy, or conflicting sources remain future work.Source: §6 · p. 10

A.1 Dataset Details

• For three QA benchmarks and APIBench, sources are split into 30% guidance and 70% test tasks under three fixed seeds, keeping related tasks on the same side; methods share the splits.QA accuracy is macro-averaged over sources and then over splits, while APIBench and AppWorld use official evaluators.Source: §A.1 · p. 15
• MultiDoc2Dial uses government-domain documents as sources; NarrativeQA uses complete stories as sources, with 27–30 questions per story.The MultiDoc2Dial domains each contain 92–149 documents and 60 questions; NarrativeQA uses 12 stories from the official test split.Source: §A.1 · p. 15
• SWE-QA evaluates repository-level questions, APIBench evaluates API calls against hub documentation, and AppWorld tests code-driven tasks against simulated-app APIs.AppWorld uses 90 training guidance tasks and 168 test-normal tasks, reporting official task goal completion.Source: §A.1 · p. 15

A.2 Counterpart Details

• Counterparts share the backend, embedding model, and answer or agent prompt with SourceLearn; AWM uses successful guidance-task experiences while receiving the same test-time source evidence as Hybrid RAG.The methods otherwise follow their official or reproduction settings unless specified otherwise.Source: §A.2 · pp. 15–16
• Hybrid RAG uses fused BM25 and dense retrieval, RAPTOR [3] uses a hierarchical summary tree, and HippoRAG 2 passes retrieved passages to the solver.Hybrid RAG passes its top eight elements; HippoRAG 2 also passes eight passages to match the evidence budget.Source: §A.2 · pp. 15–16

A.3 Implementation Details

• GPT-5.6-Luna and gpt-oss-120b use medium reasoning effort; DeepSeek-V4.1-Flash uses high effort on QA and medium effort on APIBench and AppWorld.Within each configuration, the same backend handles construction, learning, indexing or workflow induction, and task solving; all methods use text-embedding-3-large.Source: §A.3 · p. 16
• SourceLearn builds M0 once per source and backend, runs one Inspect–Study–Consolidate cycle with K = 20 study actions, then processes each split’s guidance tasks in one pass.The resulting MS is shared across three splits; task-time source-model context has a budget of B = 24k tokens.Source: §A.3 · p. 16
• Ablations use two sources per QA benchmark; AppWorld methods use the official ReAct-style code agent with a 50-step limit and method-specific documentation access.The implementations also access GPT-5.6-Luna and DeepSeek-V4.1-Flash through hosted APIs and serve gpt-oss-120b locally with vLLM on six NVIDIA A100 80GB GPUs.Source: §A.3 · p. 16

• All experiments in this section use GPT-5.6-Luna and the sources, guidance/test splits, and judge of the main results.Source: B Additional Results · p. 16

B.1 Robustness to Incomplete Retrieval

• With no retrieved evidence, the source model alone performs comparably to Hybrid RAG with eight elements; removing the gold document lowers Hybrid RAG by about 25 points versus about 8 for SourceLearn.SourceLearn also degrades more slowly as retrieval is reduced.Source: §B.1 · p. 17, Figure 5 · p. 17

B.2 Comparison at Matched Context Budgets

• At matched per-question context budgets, SourceLearn remains best on three of four sources and second on Sphinx.The comparison controls context amount without changing retrieval ranking or index.Source: §B.2 · p. 17, Table 2 · p. 17

B.3 Grounding and Source Authority

• Contradictions remain at or below 1%, and 93–100% of dropped claims are preserved, though some Task-Guided units on document and code sources are partially supported or unsupported.The authors conclude that the source model is reusable understanding, not a replacement for the authoritative source.Source: §B.3 · p. 18, Table 3 · p. 18, Table 4 · p. 18

B.4 Transfer Beyond Guidance-Task Regions

• SourceLearn improves over Hybrid RAG by 14.9 points on seen regions and 12.1 points on unseen regions.Seen regions overlap with guidance tasks; unseen regions do not.Source: §B.4 · p. 18, Table 5 · p. 18

B.5 Number of Self-Directed Learning Cycles

• Additional self-directed study cycles expand the source model and generate observations, but accuracy shows no consistent improvement beyond the first cycle.This motivates the single-cycle setting used in the main experiments.Source: §B.5 · pp. 18–19, Table 6 · p. 19

C Prompt Details

• The prompt appendix covers initial construction, model-conditioned inspection, adaptive study planning, evidence consolidation with a grounding gate, and cross-task representation learning.Source: C Prompt Details · p. 19

C.1 Initial Source Model Construction (Construct)

• Construction builds entity profiles with attributes, mechanisms, distinctions, and decisive details, while leaving lookup facts to the raw source.The prompt favors structural entries and requires source element identifiers to be copied verbatim.Source: §C.1 · p. 19

C.2 Model-Conditioned Source Inspection (Observe)

• Inspection records temporary, evidence-backed observations about what the current model explains poorly, including rules, conditions, dependencies, procedures, exceptions, and connections.A full-entity review checks whether the model represents the source text’s rules and cross-references without fragmentation.Source: §C.2 · p. 20

C.3 Adaptive Source Study Planning (PlanStudy)

• The planner selects consequential gaps for one focused DEEPEN or CONNECT action, asking a study question to resolve each gap.It favors entities or pairs not yet studied and avoids repeating actions.Source: §C.3 · p. 20

C.4 Evidence Consolidation and Grounding (Consolidate, GroundApply)

• Consolidation rewrites a source-model region from current units, study notes, and evidence; a grounding gate and preservation check govern whether the rewrite is committed.The rewrite retains source-supported meaning while merging fragments or making conditions and relations explicit.Source: §C.4 · pp. 20–21

C.5 Cross-Task Representation Learning (Lesson, Aggregate)

• Each guidance task can yield one general representation lesson, which is aggregated into a policy appended to writer prompts during recalibration.Lessons specify what kind of source content to represent, at what granularity, and which distinctions, relations, or conditions to preserve.Source: §C.5 · p. 21