spb/wp13_uqo

1% ============================================================================2\section{Related Work}3\label{sec:literature}4% ============================================================================56VQuant sits at the intersection of four research strands: (i)~large language models and7reasoning, (ii)~tool-augmented and agentic systems, (iii)~retrieval augmentation and grounding,8and (iv)~the application of LLMs to finance. We review each in turn and position our9contribution accordingly.1011\subsection{Language models and reasoning}12\label{subsec:lit_lms}1314The transformer architecture of \citet{vaswani2017attention} and the demonstration of15in-context, few-shot learning at scale by \citet{brown2020language} established the modern16paradigm of general-purpose language models. A complementary line of work showed that17intermediate ``chain-of-thought'' reasoning improves performance on multi-step problems18\citep{wei2022chain}, motivating a generation of models that allocate explicit computation to19deliberation before answering. VQuant builds directly on this development: it targets20reasoning-augmented models that can ``think'' under an adaptive token budget prior to21responding, and it surfaces this intermediate reasoning to the user. We treat the foundation22model as a fixed, externally provided capability \citep{bommasani2021opportunities} and study23the \emph{system} that must be engineered around it.2425\subsection{Tool use and agentic systems}26\label{subsec:lit_agents}2728A model confined to its parametric memory cannot retrieve live prices or perform exact29arithmetic. The tool-use literature addresses this by teaching models to call external30functions. \citet{schick2023toolformer} showed that models can learn to invoke APIs, while the31ReAct framework of \citet{yao2023react} interleaved reasoning traces with actions in an32observe--think--act loop that has become the template for modern agents. Program-aided33approaches \citep{gao2023pal} delegate computation to a code interpreter, offloading exact34arithmetic from the language model to a deterministic runtime---a principle that VQuant adopts35as a first-class design constraint. Industrial function-calling protocols36\citep{anthropic2024tooluse} standardized the schema-driven invocation of tools, and the Model37Context Protocol generalized tool exposure across providers. VQuant operationalizes these ideas38at scale: rather than a handful of tools, it exposes a curated catalogue of more than two39hundred, and it contributes practical mechanisms---prompt-level routing, execution batching,40and a single-call computation paradigm---for managing the resulting orchestration complexity.4142\subsection{Retrieval augmentation and grounding}43\label{subsec:lit_rag}4445Retrieval-augmented generation \citep{lewis2020retrieval} grounds model outputs in external46documents, mitigating hallucination and staleness. VQuant generalizes grounding beyond text47retrieval: its ``retrieval'' surface includes structured market-data endpoints, semantic and48keyword web search, and on-demand numerical computation. Every quantitative claim in a VQuant49report is traceable to a deterministic source, which we view as the financial-domain analogue50of citation-grounded generation. This emphasis on auditability connects to broader concerns51about reliability and provenance in machine learning for finance \citep{lopezdeprado2018advances}.5253\subsection{Large language models in finance}54\label{subsec:lit_finance}5556Domain-adapted models such as BloombergGPT \citep{wu2023bloomberggpt} and open frameworks such57as FinGPT \citep{yang2023fingpt} demonstrated the value of finance-specific training data.58On the application side, \citet{lopezlira2023chatgpt} found that LLM-derived sentiment from59news headlines predicts subsequent returns, and \citet{kim2024financial} showed that LLMs can60analyze financial statements at a level competitive with human analysts. These studies largely61treat the LLM as a \emph{forecasting or classification} engine operating on text. VQuant takes62a different stance: it uses the LLM as an \emph{orchestrator} that assembles live data,63delegates numerical work to deterministic code, and composes the results into a report. The64classical quantitative methods it automates---option pricing \citep{black1973pricing,65merton1973theory}, conditional-volatility modeling \citep{engle1982autoregressive,66bollerslev1986generalized}, Monte Carlo valuation \citep{glasserman2004monte}, and67Value-at-Risk \citep{jorion2007value}---are well established; the contribution is the68\emph{architecture} that makes them accessible through natural language while preserving69reproducibility.7071\subsection{Positioning}72\label{subsec:lit_positioning}7374Relative to prior work, VQuant is distinctive in three respects. First, it integrates an75unusually broad and heterogeneous tool ecosystem under one protocol, and it contributes76concrete methodology for routing and rate-managing that ecosystem. Second, it enforces a strict77separation between reasoning (performed by the LLM) and computation (performed by deterministic78code), which we argue is the appropriate design for numerically sensitive domains. Third, it79treats the \emph{deliverable}---a citable, exportable, reproducible document---as a primary80artifact rather than an afterthought. The remainder of the paper develops each of these points.81