AI thesis tools generate proposal reports, but half the references open to blank pages. In 2026, should you fix the candidate pool first or add export validation first?

Conclusion first: in 2026, when building AI thesis tools, if references in a proposal report open to blank pages or cannot be found, it is usually not a regression in model capability, but rather a missing "citation verifiability" gate in the generation pipeline. The model should only be responsible for organizing language; references must first land in a searchable candidate pool, then the business layer performs existence and consistency checks, and one more spot check is done before export. Simply writing "do not fabricate references" in the prompt basically cannot stop it, and the further it gets into a long piece, the more easily it is ignored.

Why references are often the first thing to go wrong when AI writes a proposal report

Generative models continue text by probability. When they reach the "references" section, they output a string of characters that "look like references"—author names look plausible, journal names look plausible, the year is reasonable—but the combination may not actually exist. Unfortunately, in the proposal report scenario, users will click links and copy titles to search. If a search returns nothing, they immediately conclude "this tool is unusable." Errors in the body text may go unnoticed, but citation errors are easily discovered.

A common situation in projects is: tight constraints—launch in two weeks, limited budget, and the only materials at hand are the school-provided proposal template and a list of past references. The approach is to not touch the model first; instead, turn the past list plus searchable sources into a candidate pool, and during generation only allow selections from the pool. The result is a clear improvement in first-version deliverability. Based on projects we have encountered, for scenarios that launch within two weeks and only cover an in-school template, the typical range for first-version deliverability can rise to around 70% (experience range, fluctuating with literature pool coverage). The trade-off is narrower coverage: research directions not in the pool require manual additions, and revisions are mostly concentrated in the addition step rather than logic rework.

A nameable framework: the four-layer citation trust chain

We recommend splitting it into four layers: capability layer—business layer—carrier layer—data and risk-control layer, treating "citation" as data that runs through the entire chain, rather than a field added after generation. Each layer manages a different thing: the capability layer manages whether generation is possible, the business layer manages what is allowed to be generated, the carrier layer manages where users see it, and the risk-control layer manages how to fall back after an error.

1. Capability layer (model/multimodal): responsible for understanding the topic, organizing chapter structure, and rewriting to reduce duplication; references are only matched and ranked, not left to free improvisation.
2. Business layer: reference candidate pool, structured fields (title, author, year, source, link or identifier), and one-to-one correspondence between in-text citation numbers and the reference list.
3. Carrier layer (website/mini-program/APP/H5): before exporting Word or PDF, perform one more rendering validation—links clickable, numbers continuous, layout following the school template.
4. Data and risk-control layer: logging and traceability, spot-check queue, manual review entry point, and a feedback loop for users reporting "cannot open."

The reason for this division is that citation errors basically happen at the handoffs between layers—the pool has the reference, but the business layer does not number it; the number is correct, but it gets overwritten by the template during export. Therefore, each layer must leave a verifiable artifact, rather than just showing a line in the UI saying "AI-generated, for reference only."

How to intercept on the generation side: turn "fabricating" into "selecting"

In 2026, a common approach is retrieval-augmented generation (RAG): first use topic keywords to recall a batch of candidates from an in-house literature library or publicly searchable sources, then let the model select from the candidates and write them into paragraphs. Models such as GPT, Claude, Gemini, Tongyi, and DeepSeek can all handle this ranking and rewriting step. The differences are mainly in their feel for Chinese academic corpora and long-context performance; there is no need to agonize over specific versions here.

• Only allow "selection": if the candidate pool cannot provide it, leave the output blank and prompt the user to add it, rather than adding something that looks similar.
• Structured output: have the model return fields by title, author, year, source, link or identifier—not a whole line of string.
• Flagging uncertainty: the model can mark "to be verified," but the UI must explicitly distinguish verified from unverified; the two must not share the same style.
• Counterexample: relying only on a system prompt constraint across the whole piece—"do not fabricate references"—decays in long text; it looks compliant but does not actually stop the problem.

How to catch it on the validation side: a three-level validation gate

After generation, do not directly allow export. Run through three levels; the cost is not high, but it can block most complaints. The first level runs by machine, the second requires retrieval or semi-automation, and the third uses the model plus rules. Reversing the order is not recommended; otherwise, human reviewers will be flooded by a large amount of noise.

1. Format validation (millisecond-level): continuous numbering, complete fields, and formatting according to the reference style specified by the school or journal. A common practice is that the school designates a national-standard-style format; just check against the template.
2. Existence validation (API or semi-automatic): for references with links, check reachability; for those without links, run title searches. Those that cannot be found go into a manual review queue; do not silently let them through.
3. Consistency validation (model plus rules): check whether in-text citation numbers exist in the list, and whether the point at the citation is obviously mismatched with the reference topic.

The pass line can be made a bit more specific: in exported documents, the proportion of links that cannot be opened or titles that cannot be found should be controlled to within single-digit percentages (experience range, fluctuating with literature library coverage); in-text citation numbering misalignment should be zero. If these two cannot be achieved, do not rush to launch "one-click export."

Solution comparison: direct generation, retrieval-augmented generation, and private deployment

• Pure model direct generation: development cycle is typically one to three weeks, billed by call volume (experience range: monthly cost for small-to-medium traffic is several hundred to several thousand yuan). The citation error rate is relatively high; suitable for internal drafts and demos.
• Retrieval-augmented generation plus validation gate: add another one to three weeks on top of direct generation; requires building a literature pool or connecting to retrieval sources, with a higher cost (experience range: after adding retrieval and storage, monthly cost is typically several thousand to several tens of thousands of yuan). Suitable for external delivery.
• Private deployment: suitable for sensitive data or high-frequency calls; one-time investment in hardware and operations is typically in the range of one hundred thousand to several hundred thousand yuan (experience range, fluctuating with concurrency and model scale), and the cycle is mostly measured in months. Not recommended as a first step.

The selection criteria are clear: if you cannot accept users not being able to click links, do not build an external version. If you want an external version, first get the three-level validation running, then discuss private deployment. Private deployment solves data not leaving the domain and per-call cost; it does not solve citation accuracy. Do not conflate the two.

Applicable scenarios and boundaries

Suitable scenarios: organizing the first draft of proposal reports and literature reviews, suggesting chapter structure, formatting and layout, rewriting to reduce duplication, and organizing reference formats. In these steps, AI is a speed-up tool; users already need to check it themselves, and the cost of errors is controllable.

Unsuitable scenarios: submitting generated results directly as a final draft; content requiring strong academic conclusions, or involving medical or legal judgment; and situations where the school explicitly prohibits the use of generative tools. What AI thesis tools can do is speed up frameworks and drafts. Verification responsibility and authorship responsibility remain with the author and advisor—writing this boundary into product descriptions and UI copy saves far more trouble than explaining it item by item afterward.

Frequently asked questions

If the model is only allowed to select from a literature candidate pool, will it never fabricate at all?

It can be greatly reduced, but not to zero. Research directions not covered by the candidate pool will still leave gaps. The acceptable practice is to leave them blank and prompt for additions, rather than letting the model add something that looks similar.

For building a thesis tool, must you connect to a database API like CNKI?

Not necessarily. An in-house literature list, open retrieval sources, and materials uploaded by users can all serve as the candidate pool. APIs only improve recall; just verify against sources permitted by the school or institution.

If a reference opens to a blank page, was it fabricated by the model or lost by the export template?

Both are possible. First check whether the reference exists in the candidate pool, then see whether the link field was overwritten by the template during export. In most cases, existence validation was not done, and the template silently dropped the field.

Before launching a proposal report tool, what level of reference handling counts as acceptable?

In exported files, the proportion of links that cannot be opened or references that cannot be found is controlled to within single-digit percentages, in-text citation numbering has zero misalignment, and unverified items are clearly marked in the UI. Only when all three are met does it pass.

Can private deployment make references more accurate?

The impact is limited. Private deployment solves data not leaving the domain and call-volume cost. Citation accuracy mainly depends on the candidate pool and validation gate; these two layers should not be conflated.

If you plan to launch an AI proposal report or thesis tool in 2026, it is recommended to first get the literature candidate pool and three-level validation running before opening export. The cost is controllable and rework is lower. Following enterprise project delivery habits, teams implementing this usually fill in the data and risk-control layers first. If data is sensitive or there is already stable high-frequency usage, then evaluate private deployment—do not reverse the order. This applies to tool-type products and campus internal systems; it does not apply to scenarios that require AI to directly produce a submittable final draft.