% ============================================================================
\section{Related Work}
\label{sec:literature}
% ============================================================================

VQuant draws on four strands of research: the design-science tradition in information systems,
which supplies the paper's genre; the literature on tool-augmented language models and LLM
agents, which supplies the mechanisms; work on the evaluation of such agents, which supplies the
standard the paper does not yet meet; and the rapidly growing literature on LLMs in finance and
economics, which supplies the application and its risks.

\subsection{Design science and system descriptions}
\label{subsec:lit_ds}

Information-systems research distinguishes behavioural studies, which test theories about the use
of artifacts, from design-science studies, which build and evaluate artifacts that solve
identified problems \citep{hevner2004design}. \citet{peffers2007design} give the canonical
process---problem identification, objectives, design and development, demonstration, evaluation,
communication---and \citet{gregor2013positioning} classify contributions by their maturity, from
situated implementations to nascent design theories. In those terms the present paper reports a
\emph{situated implementation} together with a \emph{demonstration} (a walk-through of one
request); the \emph{evaluation} step is specified but not performed. We adopt this vocabulary
deliberately: it lets us describe a working system in detail while making clear which of the
process's steps remain open.

\subsection{Tool-augmented language models and agents}
\label{subsec:lit_agents}

A model confined to its parameters cannot retrieve a current price or perform exact arithmetic.
Chain-of-thought prompting \citep{wei2022chain} improved multi-step reasoning but left
computation inside the model; program-aided language models \citep{gao2023pal} and
program-of-thoughts prompting \citep{chen2023pot} moved it out, having the model write code that a
deterministic interpreter executes, and showed large gains on numerical benchmarks including
financial question answering. Toolformer \citep{schick2023toolformer} taught models to insert API
calls into their own text, and ReAct \citep{yao2023react} interleaved reasoning traces with
actions in an observe--think--act loop that has become the template for agents. The survey of
\citet{mialon2023augmented} names this family \emph{augmented language models} and organises it
along the two axes---reasoning and tool use---that VQuant combines; \citet{wang2024agents} survey
the resulting agent architectures. Retrieval-augmented generation \citep{lewis2020retrieval} is the
special case in which the tool is a document retriever; VQuant's ``retrieval'' surface is
structured (market-data endpoints), textual (web search and extraction), and computational
(code), and the auditability it seeks is the financial analogue of citation grounding.

Two practical problems arise when the tool surface grows. The first is selection: Gorilla
\citep{patil2024gorilla} and ToolLLM \citep{qin2024toolllm} fine-tune models to choose among
thousands of APIs and introduce retrievers over tool documentation, whereas VQuant transmits its
entire 245-tool catalogue on every call and relies on schema descriptions and a routing prompt.
The second is context: tool results are verbose, and \citet{liu2024lost} show that models use
information in the middle of long contexts poorly even when it fits. Provider-level
function-calling protocols \citep{anthropic2026tooluse} and the Model Context Protocol
\citep{mcp2025spec} standardise how tools are declared and invoked but leave both problems to the
application. VQuant's prompt-level routing (Section~\ref{subsec:routing}) and three-stage
context management (Section~\ref{subsec:context}) are its answers; neither is claimed to be
optimal, and the evaluation protocol of Section~\ref{subsec:protocol} is designed to test them.

\subsection{Evaluating LLM agents}
\label{subsec:lit_eval}

The agent literature has converged on benchmark suites with executable environments and
automatic scoring: AgentBench \citep{liu2024agentbench} spans eight environments, Reflexion
\citep{shinn2023reflexion} evaluates self-correction over repeated episodes, and WebGPT
\citep{nakano2021webgpt} pioneered human-preference evaluation of browser-assisted answers. For
open-ended outputs such as reports, model-based grading has become standard, with
\citet{zheng2023judge} documenting both its agreement with human raters and its biases.
\citet{chang2024survey} survey the resulting evaluation landscape. In finance specifically,
FinanceBench \citep{islam2023financebench} supplies questions over public filings with
human-verified answers and reports that unaided retrieval-augmented models answer a majority of
them incorrectly, and FinBen \citep{xie2024finben} assembles a multi-task financial benchmark that
includes agent and quantitative-reasoning tasks. These resources are the natural starting point
for an evaluation of VQuant; the protocol in Section~\ref{subsec:protocol} builds on them.

\subsection{Large language models in finance and economics}
\label{subsec:lit_finance}

Text has long been a source of financial signals: \citet{loughran2016textual} survey
dictionary-based textual analysis, \citet{gentzkow2019text} review text-as-data methods for
economists, and \citet{li2021culture} show that embedding-based methods recover latent constructs
such as corporate culture from earnings calls. LLMs extend this line. \citet{lopezlira2026chatgpt}
find that ChatGPT-derived headline sentiment predicts subsequent returns; \citet{chen2023expected}
use LLM representations of news to forecast returns across sixteen markets; and
\citet{hansen2023fedspeak} show that GPT models classify the stance of FOMC communications more
accurately than dictionary and smaller supervised methods.\footnote{A widely circulated study of
LLM-based financial statement analysis \citep{kim2024financial} was withdrawn from arXiv by its
authors in February 2025 pending a review of data inconsistencies; we therefore do not rely on it.}
Domain-adapted models such as BloombergGPT \citep{wu2023bloomberggpt} and FinGPT
\citep{yang2023fingpt} demonstrate the value of finance-specific training data, and
\citet{li2023survey} survey the field. On the applied side, \citet{fatouros2025marketsenseai}
build an LLM-driven stock-selection framework from news, fundamentals, and prices and evaluate it
on realised performance, while \citet{korinek2023generative} and \citet{dell2025deep} assess
what generative and deep-learning methods offer to economic research.

Most of this work treats the LLM as a \emph{forecasting or classification} engine operating on
text. VQuant takes a different stance: the model is an \emph{orchestrator} that assembles live
data, delegates numerical work to deterministic code, and composes the results into a document.
The classical methods it automates---option pricing \citep{black1973pricing,merton1973theory},
conditional volatility \citep{engle1982autoregressive,bollerslev1986generalized}, Monte Carlo
valuation \citep{glasserman2004monte}, and Value-at-Risk \citep{jorion2007value}---are
established; the broader concern with reproducibility and provenance in financial machine
learning is articulated by \citet{lopezdeprado2018advances}. The point of contact with the
forecasting literature is the risk, documented there, that LLM outputs about numbers are
confidently wrong; VQuant's architecture is an attempt to remove numbers from the model's remit
altogether.

\subsection{Positioning}
\label{subsec:lit_positioning}

Relative to this work, the artifact described here is distinctive in three respects, none of
which is an empirical claim. First, it exposes a broad and heterogeneous catalogue of tools under
one protocol and documents the practical mechanisms---routing, batching, truncation,
summarisation---needed to operate it. Second, it enforces a strict separation between reasoning
(performed by the model) and computation (performed by deterministic code), a separation that
\citet{gao2023pal} and \citet{chen2023pot} showed to be beneficial on benchmarks and that VQuant
carries into a production setting with live data. Third, it treats the deliverable---an
exportable, shareable document---as a primary artifact. What the paper does not establish is
whether these choices improve accuracy, latency, or cost relative to simpler designs; that is the
question Section~\ref{subsec:protocol} is written to answer.