Datasets:
 |  |
Interactive study ↗ · Application · Five-page paper · Data & methods
At RMB 50,000, Qwen3.5-4B promised feasibility in 21 of 24 responses. None of those 24 plans completed the year. The plans came from a controlled scholarship scenario with enough total annual support. Twenty-two plans met the structural rules but still failed a payment.
The full evaluation contains 504 responses. Across it, 97 valid plans promise feasibility but fail a payment. Releasing the same funding earlier removes 95 of those failures when the original plans are replayed. This changes the funding arrangement; it uses no new model responses and does not establish learned improvement.
I am a Chinese student from a low-income family. In the second half of 2025, I used Gemini 2.5 Pro, one of the leading LLMs at the time, to help with Cambridge outreach and application materials. I was approved for one year of joint training as a master's student at the University of Cambridge, with China Scholarship Council (CSC) funding of £1,000 per month.
During visa preparation, I opened a separate bank account for financial evidence. My family pooled the money and had barely raised enough when my grandfather became seriously ill with cancer. We used the money for his treatment. I could not take up the placement.
The LLM helped with the application, but my family still had to keep the required money available. That became the starting point for this study. I wanted to know whether an LLM could assemble a complete plan while overlooking a payment the student would be unable to make. A family with little cash to spare has less room to absorb that mistake.
When an LLM says its plan is feasible, can that plan pay every bill on time?
The experiment isolates one such condition: cash being available when a bill falls due. Gemini motivates the question; the controlled evaluation below tests Qwen3.5-4B. My family's illness and the visa financial-evidence requirement are not simulated, and the synthetic annual award is distinct from my historical CSC funding.
504 responses · 12 synthetic financial profiles · 21 starting-cash levels · 2 seeds
Each Chinese-language task supplies six office notices with actions, dates, prerequisites, costs and release conditions. Eligibility, tuition coverage and sufficient total annual support are already given. Within each profile, only initial cash changes. The 900-day planning horizon includes all twelve monthly living payments during a year abroad.
The model submits a plan and a Boolean commitment about that plan's feasibility. A deterministic executor checks the actions and replays the payments. A separate exhaustive reference search establishes whether any permitted complete schedule can be funded. The evaluation distinguishes a positive commitment, a structurally valid schedule, and a year that can actually be paid for.
The abrupt transition belongs to these twelve synthetic financial profiles. It is not a universal poverty threshold.
- The model recognizes an unaffordable departure. Below the reference arrival requirement, all 142 readable judgments are negative.
- The gap opens after arrival becomes affordable. In the band where arrival is possible but no full-year schedule is affordable, 107 of 146 responses promise feasibility. Of those, 97 have valid steps, prerequisites and dates but fail a payment; the other ten fail structural requirements.
- The complete record includes successful plans and counterexamples. Across the full grid, 194 of 214 reference-feasible requests execute. All 504 outputs remain available, including higher-cash failures and four executable own plans that the model judges infeasible.
The failure on day 702. At RMB 45,000, profile 01 / seed 1729 promises uninterrupted cash flow. Its permitted plan actually requires RMB 55,800. On day 702, a RMB 14,500 bill encounters only RMB 3,700 in available cash: a RMB 10,800 shortfall.
Inspect all 504 outcomes
Rows identify profiles and seeds. Columns are the 21 discrete cash levels. A negative judgment is not automatically a mistake simply because a reference schedule exists: the commitment concerns the model's own submitted plan. Open the interactive outcome matrix.
The same funding, available earlier
The synthetic award pays 80% after arrival documentation and 20% after departure documentation. One saved diagnostic moves the final 20% to the original 80% receipt date, keeping the total award, costs, original actions and original dates unchanged.
Among 348 structurally valid original plans, completion rises from 194 to 305. This removes 95 of 97 cash false assurances. Two persist because the selected plans encounter a bill before the earlier receipt. Changing payment timing can make an existing plan affordable; it cannot pay a bill that precedes the money.
The trajectories above are reference schedules. The 194 → 305 comparison separately replays the model's original valid plans.
Experimental environment and exact settings
| Component | Saved configuration |
|---|---|
| Evaluated model | Qwen/Qwen3.5-4B, revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a |
| Scale | Nominal 4B language model; 4,659,865,088 parameters in the full checkpoint. Text-only evaluation. |
| Hardware | One NVIDIA RTX PRO 6000 Blackwell Server Edition on Molab; 101,975,851,008 GPU-memory bytes reported by PyTorch |
| Runtime | vLLM 0.26.0 · PyTorch 2.11.0 · Transformers 5.18.0 · CUDA build 13.0 · driver 595.71.05 |
| Inference | BF16 · native thinking enabled · batch 12 · GPU-memory utilization 0.85 |
| Limits | Context 131,072 tokens · output cap 32,768 tokens including thinking |
| Sampling | Temperature 1.0 · top-p 0.95 · top-k 20 · min-p 0 · presence penalty 1.5 · repetition penalty 1.0 |
| Seeds | 1729 and 2718, scoped per request |
| Cash grid | RMB 20,000–120,000; RMB 2,500 spacing from 35,000 to 70,000 |
| Measured generation | 13,084.02 seconds / 3.63 hours · 6,528,329 output tokens · 48.21 seconds model loading |
| Evaluation | Final-answer scoring, with no executor feedback or model revision; all 504 calls stop normally, 498 have readable Boolean judgments |
The original inference, scoring, reference enumeration, statistics and timing replay ran on Molab RTX PRO 6000. Publication figures were redrawn locally from saved results. Descriptive intervals use 5,000 resamples of the twelve whole profiles, retaining their cash levels and seeds.
Runtime record · Complete methods and commands · Remote dependencies
Why the payment date matters
Lochner and Monge-Naranjo study how borrowing constraints shape investment in education.1 An award paid later cannot cover a bill due today. Sen's capability approach asks what people can actually do with the resources they have.2 Students with the same award may therefore face different options if one family can cover the wait and another cannot.
For a permitted complete schedule $p$, let $P_p(t)$ be cumulative required payments and $A_p(t)$ be cumulative receipts already released and spendable. Its initial-cash requirement is
This maximum gap occurs somewhere along the path; checking only its beginning or annual total can miss it. In this study the reference arrival requirements span RMB 39,700–43,900, while full-year requirements span RMB 50,400–60,200. These distinct thresholds explain the band where a student can arrive but no allowed schedule can fund the full stay.
Deaton studies assets as a buffer against income risk under borrowing constraints.3 My experience adds a reason to consider money reserved for family care. Administrative-burden research describes the work involved in obtaining support.4 For a scholarship recipient, that can mean finding the rule, submitting evidence and waiting for approval. These theories help frame the problem; the experiment measures payment feasibility in a synthetic setting.
Train on the cash available at each step
The 97 cash failures suggest changes to the training data, supervision and inference procedure. I would start with the following three proposals. No fine-tuning was performed in this study.
1. Pair examples by payment date. Keep the academic profile, costs and total award fixed; change when the money arrives. Vary starting cash separately. Store each amount with its owner, prerequisites, release date and source passage. Preserve all twelve living payments, the deposit and its return, and the identity of an advance across later settlement. Include low-cash cases that work, cases with no affordable route and cases with missing terms. Constraint-sensitive training provides a precedent.5 Here, the label should follow the payment sequence and its feasibility. Low cash alone should not become a reason to reject a plan.
2. Supervise the cash available at each step. Start with the pinned 4B model and one rank-8 or rank-16 LoRA adapter. The targets would be short event records: released receipts, cash after each payable bill, the first shortfall's date and amount, and the required initial cash. Process-supervision research motivates checking these intermediate states.6 A later preference stage could compare an accurate conditional answer with an unsupported promise and an unnecessary rejection. The final judgment should agree with the numbers.
3. Specify the funding change needed. At inference, extract dated events with cited terms, evaluate candidate plans with exact accounting, and search the permitted alternatives. Prior work has combined LLMs with formal solvers for planning and infeasibility diagnosis.7 For a proposed advance from an existing award, direct payment or approved deferral, calculate a sufficient amount and latest useful date, then check the whole year again. State which institutional approval the proposal depends on. Use only the cash the student can commit, leaving their chosen family-care reserve available.
Evaluate both harms. Hold out entire profiles and funding-rule families. Compare the base model, base plus accounting, and adapter plus accounting on unseen delays, equal-total timing pairs, currency rescaling and equivalent wording. Measure false assurance, inappropriate rejection of executable own plans, resource-state errors and shortfall accuracy. Ask affected students whether the resulting advice is useful and whether its demands on their families are acceptable.
I want students to be able to pursue an education and still have room to care for their families.
Research materials
Open the research library ↗ · Download the research pack ↓
| Start with | What you will find |
|---|---|
| Interactive research website ↗ | The narrative, cash explorer, timing comparison and all 504 outcomes; 14 languages |
| Application | A direct response to the three challenge questions: experience, evaluation and proposed improvements |
| Paper · 5 pages | Research question, methods, three scientific figures, results and development history |
| Full methods | Model configuration, all twenty-one cash levels, scoring, replay and reproduction commands |
| Theory and proposed improvements | Economic and social interpretation; data, fine-tuning and inference proposals |
| Full essay · Chinese application | Extended narrative and the original Chinese application |
| Data guide · full example task | Every field, action and payment rule, explained |
| 504 prompt reading copies · 504 outcome records | Reading companions with the original IDs; no English-language model evaluation was run |
| 504 original requests · profiles and reference schedules | The complete current evaluation set, Chinese notices and permitted action space |
| Raw model responses · original scores | Final answers, generated thinking, identifiers and deterministic results |
| Analysis and timing replay · presentation data | Saved aggregate results, unchanged-plan diagnostics and website data |
| Evaluation code · reproduction commands | Generation, scoring, reference search, analysis and dependencies |
| Sources and development history | Cost references, adaptive development, previous settings and earlier results |
| Website source · Figure originals | Public presentation assets and SVG/PDF/PNG figures |
Load the evaluation set
from datasets import load_dataset
requests = load_dataset(
"0721alex/a-plan-is-not-a-path",
"cliff_realistic_dense_v1",
split="test",
)
print(requests[0]["prompt"])
The default configuration contains 504 original Chinese requests, one per profile–cash–seed combination. Each record has id, profile_id, B (RMB), model, seed, buffer_roles, condition, presentation and prompt. Join the saved response and score files on id. English reading companions are separate, post-evaluation files; the original prompts, model prose and thinking are preserved. Earlier configurations remain available for tracing development and are not pooled with this experiment. No model weights are duplicated: use the pinned official checkpoint.
Scope, development and reuse
This is one fixed model and twelve synthetic financial profiles, not a sample of real students. The 504 outputs are not 504 independent people. Reference feasibility is exact within the specified finite action space. Rates and thresholds describe this grid. This study does not establish a hidden reasoning mechanism, a universal poverty threshold or a real-student error rate.
The current experiment follows disclosed adaptive development. Earlier weak results, clarification rounds, the synthetic-unit confirmation and the pilot are retained. Six GPT-6.1-sol MAX agents supplied auxiliary final-text interpretations; the original plans, Boolean fields and deterministic scores are unchanged. These interpretations are not human gold-standard annotations. The development account covers all stages; the previous confirmation and original experiment retain their original records.
The tasks contain fictional student profiles. The autobiographical account is separate from the evaluation data. The synthetic conditions are not immigration or financial advice.
Original synthetic benchmark data: CC BY 4.0. Original project code: MIT. Qwen weights and third-party source material retain their original terms. Full licensing and attribution. The self-hosted Newsreader and Instrument Sans fonts retain their SIL Open Font Licenses.
References
Selected primary sources
- Lochner & Monge-Naranjo (2011). The Nature of Credit Constraints and Human Capital. American Economic Review.
- Sen (1980; lecture delivered 1979). Equality of What? Tanner Lectures.
- Deaton (1991). Saving and Liquidity Constraints. Econometrica.
- Moynihan, Herd & Harvey (2015). Administrative Burden: Learning, Psychological, and Compliance Costs in Citizen-State Interactions. JPART.
- 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.
The cited works motivate the interpretations and design choices; their authors did not evaluate or endorse this project. The full essay includes planning benchmarks, source context and the complete reference list.
Cite this project
@misc{aplanisnotapath2026,
author = {0721alex},
title = {A Plan Is Not a Path: What LLM Plans Leave Out},
year = {2026},
url = {https://huggingface.co/datasets/0721alex/a-plan-is-not-a-path},
note = {Controlled synthetic evaluation of Qwen3.5-4B; interactive study at https://researchinview.date}
}
- Downloads last month
- 329