Financial assistants often have to infer unstated premises—such as what a data figure measures, which product variant is intended, and the investor's personal situation and applicable tax rules—when answering seemingly factual questions. If the assistant ignores these premises, it can still output a real number from its knowledge base, but the answer actually addresses a different question. We call this phenomenon frame collapse.
We study frame collapse using Indian mutual funds, where product variants and tax regulations are formally defined, allowing a unique correct answer for every possible interpretation. Three challenge sets were constructed:
- Ambiguously labelled figures – models typically use the figure’s value and assume the user’s label.
- Figures with multiple labels – models list all possible readings, yet the order of the final answer follows the visual layout rather than any logical priority.
- Missing investor facts – weaker agents fall back on a default or outdated tax rule, and explicitly requesting the fact does not guarantee its use.
Because these answers match the stored data, conventional data‑consistency checks pass them. An AI‑based judge only catches mislabelled figures when each field’s meaning is explicitly provided, and it still accepts answers that silently adopt an alternative reading.
The paper illustrates how these failures arise, without quantifying their prevalence, and proposes control mechanisms for financial assistants.
Review