Extent of grounded reasoning in financial documents for investment insight

Determine the extent to which large language models can ground their reasoning in financial documents to uncover new insights for investment decision-making, assessing whether their conclusions are supported by evidence within the documents themselves.

Background

The paper motivates a knowledge discovery dimension where LLMs must derive non-trivial investment insights from long, hybrid-content financial documents (e.g., text and tables). The authors note that despite growing attention to LLMs in investment contexts, it remains unclear how well these models can ground their reasoning in actual financial documents rather than relying on unsupported internal inferences.

FinTrust includes tasks that require cross-document reasoning and numerical calculations using 10-K reports and other financial materials, intended to probe whether models can produce evidence-backed insights. The open question frames the need for a systematic determination of the degree and reliability of such grounding.

References

Third, despite the growing attention on using LLMs to advance investment decision-making , it is unclear to what extent these models can ground their reasoning in financial documents to uncover new insights.

— FinTrust: A Comprehensive Benchmark of Trustworthiness Evaluation in Finance Domain  (2510.15232 - Hu et al., 17 Oct 2025) in Section 1, Introduction

The causal identification of the transmission channels comes from one model family, whose hybrid design makes the recurrent memory state separable from the attention store; the behavioral dissociation replicates in two other families, and the collapse of decision-time reading replicates in a pure-attention architecture, but how the disclosure travels in architectures without a recurrent channel remains to be mapped.

— Reading Is Not Using: Retrieval, Judgment, and the Design of AI Financial Research Workflows  (2608.24842 - Liu et al., 25 Aug 2026) in Section 6, “Conclusion”