KnowPlan AI agents plan degree pathways with 99.5% certification rate

KnowPlan, introduced by Shuheng Cao and six researchers on arXiv, tackles the messy reality of university degree planning. Most AI planners fail because they couple information extraction with path optimization, so the system only learns facts its current plan needs. KnowPlan enforces an extraction-first boundary: CatalogBrowse explores university sources with no user profile, scoring legal actions by expected marginal gain across atomic catalog obligations. It parses deterministically via platform adapters, with a span-constrained clause-to-AST model as fallback, and terminates on a closure certificate over index, schema, provenance, and reference completeness—not a reward threshold. Its output is three provenance-linked JSON documents.

DegreeMap consumes only those documents, compiling them into a typed requirement hypergraph and optimizing lexicographically with CP-SAT across hard feasibility, completion horizon, load, risk, personalized utility, and option value. Each optimization stage stays inside the previous stage's proven optimum. In benchmarks across a 100-university broad track and a six-school dense track, CatalogBrowse achieved 96.2% inventory recall and 88.7% masked-source recovery at 47% less source access than an exhaustive crawler. DegreeMap maintained 100% hard feasibility while improving personalized utility by +0.066 over the strongest baseline. The full pipeline certified 99.5% of requests with a utility gap of only 0.015 against a privileged gold graph—demonstrating that decoupled, knowledge-driven agents can handle real-world institutional data at scale. This research is relevant to any system that must ingest schemaless, distributed documents before optimizing decisions.