Learnijoy
ResearchAI ResearchAI Investing

RAG vs. Deterministic Tax Engines in Financial Advisory AI

Aryan Brar, Justin Du, Avery Lor, Kylie Seto, Eric Taylor· August 26, 2026 View original

Key takeaways

  • Deterministic tax engines surprisingly reduced tax savings in an AI advisory system.
  • RAG-only conditions achieved the highest tax savings.
  • LLMs' internalized financial knowledge can be sufficient for complex tasks like tax-loss harvesting.
  • Augmenting LLMs with domain-specific engines doesn't guarantee improvement and can introduce conflicts.

Who benefits

BFSIWealth ManagementFintechInvestment Banking

Summary

A 2x2 factorial experiment compared Retrieval-Augmented Generation (RAG) and a custom capital gains calculation engine in a multi-agent financial advisory system for tax-loss harvesting. The study found that enabling the deterministic tax optimization engine surprisingly reduced tax savings, while RAG alone performed best, suggesting LLMs' internalized financial knowledge can be sufficient.

Researchers conducted a 2x2 factorial experiment to evaluate the effectiveness of different context providers in a multi-agent financial advisory system focused on tax-loss harvesting. The study compared the impact of a custom, deterministic capital gains calculation engine against a Retrieval-Augmented Generation (RAG) system, which provided context from market advisory reports. The findings revealed a significant and counterintuitive result: enabling the deterministic tax optimization engine actually led to a reduction in tax savings by approximately 55 percentage points compared to conditions where it was not used. Conversely, the RAG system alone achieved the highest descriptive mean tax savings, outperforming even the baseline condition where no explicit tooling was provided. This suggests that the pre-trained language model's inherent financial knowledge might be sufficient for competent tax-loss harvesting recommendations, and augmenting it with a domain-specific computation engine does not guarantee improved performance. In some cases, such augmentation might even introduce conflicting optimization signals, hindering the overall effectiveness of the advisory system.

Why it matters

This research challenges the assumption that more explicit, deterministic tooling always improves AI performance in complex financial tasks. It highlights the potential for LLMs' inherent knowledge to be highly effective and the need to carefully evaluate augmentation strategies to avoid unintended negative impacts.

How to implement this in your domain

  1. 1Re-evaluate the necessity and impact of integrating deterministic computation engines with LLM-based financial advisory systems.
  2. 2Conduct rigorous A/B testing or factorial experiments when augmenting LLMs with external tools to measure actual performance improvements.
  3. 3Prioritize leveraging and fine-tuning the LLM's internalized domain knowledge before adding complex external computational components.
  4. 4Develop robust evaluation metrics for financial advisory systems that go beyond simple accuracy to include real-world outcomes like tax savings.
  5. 5Investigate potential "conflicting optimization signals" that might arise when combining LLM reasoning with deterministic engines.

Original post by Aryan Brar, Justin Du, Avery Lor, Kylie Seto, Eric Taylor

"arXiv:2608.23908v1 Announce Type: new Abstract: Tax-loss harvesting demonstrates consistent benefits to long-term portfolio growth; yet implementing it efficiently often involves complex considerations that are specific to the holdings within that portfolio and the individual who…"

View on X

Originally posted by Aryan Brar, Justin Du, Avery Lor, Kylie Seto, Eric Taylor on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevToolsAI Investing

FraudBench Benchmarks Adversarial Robustness in Financial Risk Assessment

This paper introduces FraudBench, a protocol-sensitive benchmark for evaluating the adversarial robustness of machine learning models in financial fraud and credit-risk detection. It demonstrates that robustness conclusions are highly dependent on how domain-specific constraints and attacker capabilities are incorporated into the evaluation protocol.

Xitong Zeng, Zhaoge Bi, Yitian Yang, Huaming Chen, Quan Z. ShengAug 26, 2026
AI ResearchAI Engineering & DevTools

Persistent Cross Entropy Extends Topological Data Analysis

This paper introduces Persistent Cross Entropy (PCE), a novel extension of cross-entropy to persistence diagrams, which are used in topological data analysis. PCE bridges different event spaces of diagrams using an induced probability, enabling new applications like distinguishing diagrams with similar persistent entropy and separating causal directions in dynamical systems.

Sijin Yeom, Jae-Hun JungAug 26, 2026
AI ResearchAI Engineering & DevTools

Bridging Numerical PDE Solvers and Neural Emulators for Faster Simulation

This thesis explores the deep connections between traditional numerical solvers for Partial Differential Equations (PDEs) and neural emulators, arguing that they are more alike than different. It proposes that insights can flow profitably in both directions, leading to faster and more efficient scientific and engineering simulations.

Felix KoehlerAug 26, 2026