🤖 AI

KNOWPLAN: Knowledge-Driven AI Agents for Smart Degree Pathway Planning

KnowPlan is a two-stage, knowledge-driven AI system that first reconstructs university curricula from heterogeneous sources using a proactive, extraction-first crawler (CatalogBrowse) and then optimizes personalized degree pathways via a certified hypergraph solver (DegreeMap), achieving high recall and feasibility while significantly reducing source access costs compared to end-to-end approaches.

Original authors: Shuheng Cao, Weijia Zhang, Jiaqi Wu, Xiyun Hu, Yat Yang, Juqy Chen, Zhaoxiang Feng

Published 2026-08-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shuheng Cao, Weijia Zhang, Jiaqi Wu, Xiyun Hu, Yat Yang, Juqy Chen, Zhaoxiang Feng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build the ultimate custom Lego castle, but the instructions are scattered across a million different websites, some written in code, some in PDFs, and some hidden behind doors you don't know exist. This is the world of academic advising, a field where computers try to help students figure out exactly which classes to take to graduate on time. For years, researchers have built smart systems that can suggest the perfect schedule if they already have the complete list of rules. But here's the catch: getting that list of rules is a nightmare. University websites are messy, inconsistent, and constantly changing. If a computer tries to "read" the rules while simultaneously deciding what classes to pick, it creates a confusing loop: it might skip a rule because it didn't think it was important yet, only to realize later that skipping that rule makes the whole plan impossible. This paper tackles that specific headache, asking: how can a computer gather all the messy rules first, without knowing the student's personal goals, and then use those rules to build a perfect, guaranteed plan?

Enter KnowPlan, a two-step robot team designed to untangle this academic knot. Think of the first robot, CatalogBrowse, as a super-organized librarian who doesn't care who you are or what you want to study. Its only job is to explore the university's digital library and find every single rule, course, and requirement, no matter how hidden. Instead of just clicking links randomly or following a simple map, this robot plays a high-stakes game of "what's next?" It calculates the value of every possible click, asking, "If I open this page, how much new information will I get compared to the effort?" It keeps going until it has a "closure certificate"—a digital seal of approval that proves it hasn't missed any critical pieces of the puzzle. It doesn't guess; it stops only when it is mathematically sure it has found everything it can find in that frozen snapshot of the website. Once it's done, it hands over three neat, organized digital files containing the entire curriculum.

The second robot, DegreeMap, is the master planner. It takes those three files and the student's personal profile (like their grades, interests, and how much work they can handle). It then builds a complex 3D map of requirements, where courses are connected by invisible strings of logic (like "you must take Math 101 before Math 102"). Using a powerful math engine, it doesn't just find a plan; it finds the best plan. It balances hard rules (you can't graduate without this class) with soft desires (you really love biology), ensuring the student finishes in the shortest time possible with the least stress. Crucially, because the first robot did its job perfectly, the second robot never has to guess if a rule exists. It just works with what it's given.

The results are impressive. When tested on 100 different universities, the first robot found 96.2% of all available courses and recovered 88.7% of the "hidden" pages that other methods missed, all while visiting 47% fewer web pages than a robot that just tries to read everything. The second robot then used this data to create plans that were 100% valid (no impossible schedules) and made students happier with their choices by a small but measurable margin compared to other top systems. The whole system successfully certified 99.5% of the requests, proving that by separating the "finding the rules" job from the "making the plan" job, we can build a much more reliable and personal guide for students. The authors show that this approach removes the confusion of circular logic, turning a chaotic search into a certified, step-by-step journey to graduation.

Technical Summary: KnowPlan

Problem Statement

Academic degree pathway planning is a constraint-driven decision problem requiring students to sequence courses to satisfy prerequisite logic, degree requirements, unit limits, and term offerings. While prior work has addressed course recommendation and curriculum optimization, these approaches largely assume a sufficiently complete curriculum representation is already available.

Constructing this representation from public university sources is inherently difficult. Institutional curricula are distributed across heterogeneous formats (webpages, JSON endpoints, PDFs) with no shared schema, containing nested logical clauses, requirement overlaps, and policy exceptions. A critical, unaddressed gap in existing literature is the acquisition–planning circularity: if a personalized planner controls source acquisition, the current student objective determines which pages are opened. Consequently, the resulting partial graph determines which plans appear feasible, and the same incomplete graph is used to validate the plan. Missing rules remain invisible because they were never acquired.

Methodology

The authors present KnowPlan, an extraction-first framework that separates curriculum reconstruction from personalized degree planning to break this circularity. The system operates in two distinct stages:

1. CatalogBrowse: Profile-Independent Extraction

CatalogBrowse is a web exploration agent designed to acquire curriculum data without access to any student profile.

State and Observability: The agent maintains a dynamic page graph and a frontier of unopened links. Crucially, page content is revealed only after an action is precommitted and paid for, preventing the agent from inspecting content to decide whether to access it.
Atomic Obligations: Instead of relying on weighted completion signals that encourage easy wins, the agent targets a finite set of atomic obligations ( $O$ ), such as discovering index entries, instantiating entities, filling required fields, and resolving references.
Policy: The agent selects actions using a lower-confidence bound (LCB) on the expected marginal obligation gain per unit of source access. This conservative policy prioritizes evidence already validated and ensures that unfamiliar actions are only pursued if the obligation itself generates positive expected gain.
Parsing and Fallback: The system uses deterministic platform adapters for APIs, DOMs, and PDFs. For complex logical clauses, it employs a span-constrained clause-to-AST model. The model is restricted to tokens literally present in the visible span and cannot emit arbitrary URLs or identifiers; outputs are re-located in the page text before acceptance.
Termination: The agent stops only when a closure certificate holds. This certificate verifies that every discovered index entry is either bound to an entity or recorded as explicitly failed, every schema-required field is bound or explicitly unresolved, and every provenance offset lies in bounds on a page the agent opened. The certificate locates evidence but does not adjudicate the facts themselves; it is snapshot-relative and makes no claim about the live Web.
Output: The stage emits exactly three provenance-linked JSON documents: courses/prerequisites, program requirements, and general-education frameworks.

2. DegreeMap: Personalized Planning

DegreeMap consumes only the three JSON documents produced by CatalogBrowse and the student's specific request.

Typed Requirement Hypergraph: The compiler maps the JSON documents into a typed requirement hypergraph. It preserves complex logical structures (e.g., CHOOSE_N, ALL_OF) as hyperedges rather than expanding them into pairwise clauses, ensuring exact enforcement of constraints like "choose 3 from 9."
Personalized Utility: The system calculates a score $S(c)$ $S (c)$ for each course based on:
- Direct semantic interest ( $I_0$ ).
- Typed graph diffusion ( $I_D$ ), propagating interest through fulfillment and cross-listing edges but not through prerequisite edges.
- Prerequisite unlock value ( $U$ ), propagating interest backward along dependency paths.
- Major/General Education gains, future option value, workload, risk, and aversion.
Lexicographic CP-SAT Optimization: The planner solves for a ranked set of plans using a lexicographic order:
1. Hard Feasibility (satisfying all constraints).
2. Minimum completion horizon.
3. Workload and scheduling risk.
4. Personalized utility.
5. Future option value.
  Each stage fixes the proven optimum of the previous stage as a hard constraint, preserving certifiability within a fixed solver budget.
Clarification: If the optimal plan varies across remaining preference hypotheses, the system asks targeted questions to reduce expected regret.

Key Contributions

Formulation: The paper formulates degree planning from public sources as a two-stage acquisition-and-optimization problem, making the acquisition–planning circularity directly measurable.
CatalogBrowse: Introduces a profile-independent agent featuring obligation-driven exploration, provenance-linked extraction, and a certificate-based stopping rule that ensures output completeness relative to a frozen snapshot by verifying that all discovered entries are processed or explicitly failed.
DegreeMap: Presents a planner that combines a typed requirement hypergraph with exact lexicographic optimization, evaluated through masked-source and paired gold-versus-extracted tracks.

Results

The system was evaluated on a Broad-100 track (100 universities) and a Dense-6 track (6 universities with diverse source formats).

Extraction Performance (CatalogBrowse):
- Achieved 96.2% inventory recall and 88.7% masked-source recovery (recovering entities behind hidden pages).
- Used 47% fewer source accesses than an exhaustive crawler while outperforming baselines like Static BFS and Go-Browse-style agents.
- Maintained high Typed F1 (79.4) and AST Semantic Equivalence (77.8) on the broad track.
Planning Performance (DegreeMap):
- Maintained 100.0% hard feasibility on the Dense-6 track.
- Improved personalized utility by +0.066 over the strongest baseline (LLMAP-style MSGS).
- Reduced the completion horizon from 6.7 to 6.3 terms.
End-to-End Performance:
- The full pipeline certified 99.5% of planning requests.
- The utility gap relative to planning over a privileged gold curriculum graph was only 0.015.
- Ablation studies confirmed that removing the fixed-point certificate reduced recall by 4.4 points, and using a weighted sum objective instead of lexicographic optimization reduced certified feasibility by 4.0 points.

Significance

The paper claims that KnowPlan removes acquisition–planning circularity by construction. By separating the acquisition of the curriculum (independent of student goals) from the optimization of the plan, the system makes the interface between data acquisition and usage measurable rather than assumed. The authors emphasize that this approach converts a modeling assumption (that the curriculum is known) into a measured quantity, demonstrating that reliable degree planning requires profile-independent curriculum acquisition before student-specific optimization. The framework is generalizable to any domain where an agent must acquire a world model to score against a later objective, provided the acquisition target is fixed before the objective is known.