The experiment identifies a gap between an LLM's commitment about its plan and the cash needed to execute it. It does not show that low-cash students always receive bad advice. The model rejected every readable arrival-unaffordable case. Its false assurances concentrated after arrival became affordable but before a full year could be funded.
Why enough annual support can still be insufficient
Lochner and Monge-Naranjo's work on borrowing constraints and education investment provides the economic starting point.[1] If borrowing is unavailable, a future receipt cannot finance a bill due now. For a valid schedule π, let Kπ(t) be cumulative obligatory payments and Aπ(t) be cumulative money already received, owned by the student and available to spend. Then
Here [x]+ means max(x,0), and P is the set of complete schedules permitted by the scenario. R(π) is the requirement of a particular schedule; R* is the smallest requirement across alternatives. R_H is the minimum to complete arrival. The interval R_H ≤ B < R* therefore has a precise interpretation: the student can arrive, but cannot complete the year on any permitted schedule.
The twelve realised R* values lie between RMB 50,400 and 60,200. Combining profiles with thresholds in this relatively narrow band produces an abrupt aggregate transition. The location of that transition follows the specified costs and rules. It is not a universal poverty threshold, and two generation seeds do not turn twelve financial variants into a sample of independent students.
The model's intermediate behaviour matters. Across 146 requests in the interval above, it made 107 unconditional positive commitments. Ninety-seven had valid schedules but failed a payment; the other ten had structural errors. All ninety-seven occurred where no affordable full-year schedule existed. The gap is therefore between commitment and execution under the given terms, rather than failure to find a cheaper hidden alternative.
From an award to an opportunity a student can use
Sen's capability approach asks what people are actually able to do with their resources.[2] Applied here, identical qualifications and nominal awards can leave different feasible options when families have different capacities to cover a delay. The amount awarded and the opportunity the student can take up are related through payment conditions.
Deaton's work on saving under liquidity constraints treats assets as a buffer against risk.[3] My family's experience adds a reason to ask which part of a balance can be committed to education without taking away money needed for care. The experiment contains no illness shock and does not estimate that trade-off. It shows why a planner's use of available cash needs to be explicit.
Administrative-burden research distinguishes the work of learning the rules, complying with them and bearing the associated psychological costs.[4] In this task, the award requires separate agreement, arrival, travel-settlement and departure actions. Free paperwork can still determine whether a large payment is released. The model must preserve those conditions in its schedule.
The FAFSA experiment found that practical application assistance combined with information changed education-related outcomes, whereas information alone did not have the same effect.[5] That was a study of human assistance, not LLMs. It nevertheless suggests a useful question for evaluating advice: does the help carry a person through the conditions needed to act?
What the saved timing diagnostic tells us
All ninety-seven cash false assurances reached arrival. Ninety-five first failed after the initial award receipt; two failed on a monthly bill before a late first receipt. Moving the same final 20% to the original first-receipt date removed ninety-five failures, with no change to total funding, costs or actions. The two earlier bills remained unpaid. Among all 348 valid original plans, completion rose from 194 to 305.
This is evidence about a payment arrangement, not a trained model improvement. It also makes a practical intervention specific: an advance must arrive before the bill it is meant to cover. Earlier disbursement transfers financing and settlement risk to the institution, even when the award total is unchanged.
Proposal 1 — Examples that isolate the changed condition
Curate published, dated award terms and consented, anonymised accounts of completed and interrupted applications. Each amount needs a recipient, release prerequisites, payment date, currency and source passage. Preserve the identity of a travel advance through settlement, the deposit's occupation and return, and all twelve living payments. A later settlement cannot create a second copy of the same money.
Pair examples that keep qualifications, costs and total support fixed while varying receipt dates. Vary starting cash separately. Include low-cash plans that work through confirmed direct payment or advances, high-cash plans with structural errors, settings with no affordable route and settings with missing terms. Labels should reflect the complete sequence, rather than teach a general association between poverty and refusal.
CARL studies constraint-aware training through comparisons of constrained and unconstrained outputs.[6] Here, a timing change should affect the answer only when it changes the computed consequences. The training target must retain the distinction between the model's own schedule and the existence of any affordable schedule.
Proposal 2 — Learn payment states, not just the final judgment
Begin with supervised fine-tuning of the pinned 4B checkpoint using a rank-8 or rank-16 LoRA adapter. Suitable targets are short event records: obligations, released receipts, cash after each payable event, the first shortfall's date and amount, R(π), R* and the final judgment. Labels can be generated from the supplied finite model of payments and checked independently of the response's prose.
Process-reward-model research motivates checking intermediate reasoning rather than only the apparent correctness of a final answer.[7] In this application, the intermediate objects are externally checkable payment states. A proposed joint objective is
The terms supervise the action sequence, payment states and final feasibility judgment. Loss weights and the value of an auxiliary state-prediction head remain design choices to test. There is no demonstrated need for a larger architecture in the current results. A later preference stage could favour a correct conditional answer over an unsupported assurance or an unnecessary rejection; it should follow, rather than replace, a supervised baseline.
Proposal 3 — Find the institutional change that would help
At inference, extract obligations and releases from documents into dated events linked to their source passages. Use exact accounting to evaluate a candidate schedule and a finite search to compare the permitted alternatives. Prior work has combined LLMs and formal tools for planning and infeasibility diagnosis.[8] The LLM's explanation can then identify the first unpaid bill, the date and the responsible condition in ordinary language.
Where no affordable schedule exists, compute a conditional remedy. If support is an unrestricted, non-repayable addition available initially, the minimum amount is
A later advance or repayable loan requires a new payment sequence, including the repayment obligation. For a direct payment, partial advance or approved deferral, calculate a sufficient amount and latest useful date, then recheck the entire year. The explanation should state which approval or missing term the proposal depends on. It should not assume extra family money, unconfirmed credit or access to a family-care reserve.
A funding office could confirm one set of documents, pay a provider directly or give a dependable disbursement date. The planner should help the student make that specific request. It cannot turn an unapproved institutional option into money already available.
Evaluation that protects affordable opportunities
Compare the base model, base plus accounting, adapter plus accounting and, if useful, a later preference-trained adapter. Hold out whole profiles and funding-rule families. Test unseen delays, equal-total timing pairs, currency rescaling and equivalent wording. Related small-model planning work reports transfer and representation failures; success on familiar schedules would not establish generalisation.[9]
Report cash false assurance, unnecessary rejection of an executable own plan, payment-state error, first-shortfall accuracy, R* prediction and correctness of conditional remedies. Retain the four executable plans that the current model incorrectly rejected. Ask students affected by financing constraints to assess whether the explanations and proposed requests are useful, and whether the family obligations they assume are acceptable.
These proposals have not been trained or evaluated in this study. The results identify what needs to be represented and what an intervention must change; they do not demonstrate the proposed training gains.
References
- Lochner & Monge-Naranjo (2011). The Nature of Credit Constraints and Human Capital. American Economic Review.
- Sen (1980). Equality of What?. Tanner Lectures.
- Deaton (1991). Saving and Liquidity Constraints. Econometrica.
- Moynihan, Herd & Harvey (2015). Administrative Burden. Journal of Public Administration Research and Theory.
- Bettinger, Long, Oreopoulos & Sanbonmatsu (2012). The Role of Application Assistance and Information in College Decisions: Results from the H&R Block FAFSA Experiment. Quarterly Journal of Economics.
- Qi et al. (2026). CARL: Constraint-Aware Reinforcement Learning for Planning with LLMs. Findings of ACL.
- Zhang et al. (2025). The Lessons of Developing Process Reward Models in Mathematical Reasoning. Findings of ACL.
- Hao et al. (2025). Large Language Models Can Solve Real-World Planning Rigorously with Formal Verification Tools. NAACL.
- Belcamino, Attolino, Capitanelli & Mastrogiovanni (2026). On the Generalization Gap in LLM Planning: Tests and Verifier-Reward RL. arXiv preprint. The study's full reference context is also available in the Research essay.