AI Agents Revise IPO Prospectuses in Response to SEC Comments: Learning Adds 3.7 Percentage Points

When a company prepares for an IPO, it files registration documents with the US Securities and Exchange Commission (SEC). Commission staff read them and send back comments, for example: "you write that revenue grew 40% because of demand; please explain in more detail." The company's lawyers gather the facts, rewrite the section, and send a response.

This goes on for several rounds. Researchers from Nanyang Technological University (Singapore) and the Hong Kong University of Science and Technology (Guangzhou) tested whether AI agents built on large language models can play both roles. To do this, they built the FilingLoop benchmark: 380 tasks covering 96 companies, mostly IPOs.

The best agent variant, which learns between tasks, met 27.96% of the regulator's requirements. The same agent without learning met 24.

28%.

How this task differs from familiar finance questions

Existing financial benchmarks, such as FinQA or FinanceBench, hand the model a ready-made question and ask it to find the answer in a report. Here there is no question, and that is the first difficulty. The Reviewer has to work out for itself which parts of the document raise questions. The second difficulty: the necessary figures may not be in the public files at all. The Filer is required to request them from the company, and in the test its access to such facts is limited to what actually existed in the historical record.

The third difficulty concerns verification. A lawyer can write in a response, "we have fixed everything," while nothing in the document itself changes. For that reason, the evaluator looks at both the final version of the document and the written response. A single SEC comment usually contains several requirements, and a careful edit to one paragraph does not close the others. The authors broke the comments down into atomic requirements, 6 099 in total. A separate evaluator that takes no part in the agents' work checks each one. It also looks for unsupported claims and for edits that damage a section that was perfectly fine before.

Two roles and instructions that accumulate between tasks

Each task reproduces one historical review round. The agents see the document under review and the prior history. The actual SEC comments, the real responses, and subsequent revisions are hidden.

• The Reviewer receives the documents and must formulate on its own what is missing.
• The Filer receives the comments, gathers facts within the available data, edits the document, and writes a response.
• Both agent variants, the fixed one and the learning one, can improve their work during a single task.

The difference lies in what happens after a task. For the learning agent, the evaluator first records the scores. Then, based on the requirements that were not met, text instructions are generated for each role. Examples include how to check a revenue breakdown or when an explanation needs to be moved into the document itself. These instructions apply only to subsequent batches of tasks and are frozen within a batch. The model weights do not change; only the set of rules the agent reads before working changes. Facts about specific companies do not make it into the instructions: only procedures that are suitable for other cases go in.

Scoring uses the first attempt on each task, with no retries. So the gain cannot be explained by the agent simply retaking the same exam.

The numbers: a small gain, but almost free

The authors compared eight configurations across all 380 tasks. The best results came from FilingLoop-Evolving, the version that learns between tasks.

• Requirement fulfillment: 24.28% → 27.96% (about 3.7 percentage points, roughly +15% relative).
• Fully closed SEC comments: 16.90% → 20.75% (about 3.9 pp, roughly +23% relative).
• Total cost of a run rose by 5.8%.
• Cost per fulfilled requirement fell, because results increased while expenses rose only slightly.

The authors write that learning improved two skills: finding gaps and conditional editing, where a document must be changed only under certain circumstances.

What this means

For practitioners, the key number is not 28% but how little the agents can do in this work. Even the best variant meets only slightly more than a quarter of the regulator's requirements and fully closes roughly one comment in five. Such a system cannot yet replace lawyers and analysts. It is useful as a draft that a human must double-check.

There are three takeaways for those building agents for regulated fields. First: check the final document, not the confident tone of the response. Second: the evaluator must be separate, otherwise the agent will praise its own work. Third: accumulating procedures without retraining the model delivers a noticeable gain for little money. For companies that cannot fine-tune models on proprietary data, this approach is simpler and safer. A caveat: this is the result of a single benchmark on historical SEC data, and the authors did not test transfer to other regulators.