A small finance model learns from its own mistakes: +3 pp without a teacher

Financial institutions often cannot send client data, positions and internal analytics to external APIs. That leaves them with open models they can run on their own servers. Such models are usually weaker than frontier closed systems at multi-step calculations, where you first have to understand accounting conventions and only then plug numbers into a formula. Researchers at the University of California, Berkeley (the work was done during an internship at CTGT and accepted to ICAIF'26 in Milan) proposed a method called Fin-SD. It raised the accuracy of GPT-OSS-20B on the hard FinanceReasoning Hard set by 3.03 percentage points. It needed neither an external teacher nor reference solutions.

How the model learns from its first error

Standard "on-policy distillation" works like this: the student solves a problem itself, and a teacher scores each of its tokens. The downside is that you need a strong external teacher. The OPSD variant does without one but requires a full reference solution.

Fin-SD asks for even less. When the model gets a problem wrong, a frozen copy of it receives the problem statement, the student's wrong answer and the correct answer. It returns three things:

• a quote of the first erroneous step;
• a short critique along the lines of "your logic diverges where you define the vacancy loss base; compute it only from apartment rent";
• the error type: conceptual or arithmetic.

Arithmetic slips and rounding errors are discarded: no gradient is computed from them. Such errors are fixed with a calculator in the scaffolding. In a manual check, the classifier got 48 of 50 examples right (96%). In the final accuracy score, these tasks still count as wrong.

Next, the solution is cut off right before the error. The student and the frozen "teacher" see the same correct prefix, but the teacher also reads the hint. The model pulls its next-token distribution toward the teacher's over a 100-token segment (the reverse KL divergence is minimized). At the end of each epoch, the list of failures is rebuilt, and the teacher is updated to the student's latest weights.

The hint barely matters

The main result is that the content of the critique turned out to be almost irrelevant.

• If you substitute a template like "Your error is conceptual" for the full analysis, the gain over the base model is +3.36 pp, versus +3.03 for full Fin-SD.
• Hints from GPT-OSS-120B (the same family, markedly larger) did not produce a materially bigger effect than the self-written ones. In this corpus, the small model mostly errs when choosing an accounting convention, and the 20B model, with the known answer in hand, locates such a spot accurately enough.
• The authors cautiously conclude that most of the gain comes from applying the signal at the exact place of the error. The hint text is secondary.

Compared with OPSD, Fin-SD scored 72.19% versus 70.42%. At evaluation time, it generates 9.5% fewer tokens. Its share of answers that fit within the token budget is also slightly higher, though the authors call this only a trend.

Data, side-effect checks and limitations

The authors built the training corpus themselves. They took a library of 3 133 FinanceReasoning financial functions, rated the difficulty of each on an 8-point scale (the rater was GPT-OSS-120B) and kept the hardest ones. DeepSeek-V4-Flash then checked whether a task with a single answer could be built, leaving 295 functions. For these, GPT-5.5 wrote scenarios in which the inputs are not given directly. They have to be derived from conventions: annual rent, vacancy, "actually invested money". Only then is a financial metric such as cash-on-cash return computed.

The main risk of fine-tuning in a bank is damaging behavior that has already been tuned. So the authors tested the post-Fin-SD model on two benchmarks:

• MMLU-Pro (general knowledge and reasoning): no noticeable drop;
• FinTrust, an audit across fourteen compliance metrics in a financial context: also no significant degradation.

Some caveats are in order. The experiments ran on a single model (GPT-OSS-20B) and on one specific corpus, where errors are mostly tied to the choice of convention. The evaluation set is small, at 238 tasks, so a difference of 1–3 pp lies close to noise. The method's author himself calls the conclusion that the hint content is unnecessary a hypothesis.

What this means in practice

For teams that keep a model inside the bank's perimeter, the takeaway is simple and inexpensive. If you have answers for the training tasks and the model itself, you can isolate the first conceptual error, cut the solution there and train on the short segment. No external teacher or fine-grained manual labeling is needed. If the gain really is determined by where the signal is applied, there is no need to pay for expensive hints from a large model. It remains to be checked whether this holds for other models and other domains.