ATA: A Neuro-Symbolic Approach to Implement Autonomous and Trustworthy Agents

TL;DR

This paper tackles the trustworthiness gaps of large language models in high-stakes tasks by introducing Autonomous Trustworthy Agents (ATA), a neuro-symbolic framework that splits work into offline knowledge ingestion and online task processing. A transpiler converts informal T&Cs into a formal, human-verifiable many-sorted first-order logic theory $T$, while an input encoder axiomizes each claim into the same formal language and a symbolic decision engine checks $T$-validity to deliver a provable, auditable result. The approach yields perfect determinism, robustness to input perturbations, and immunity to prompt injection, and it can surpass larger models when the knowledge base is corrected by humans. The claims-processing demonstration shows that grounding decisions in symbolic reasoning enables transparent explanations via minimal unsat cores or counterexamples, offering a practical path toward trustworthy autonomous agents with controllable, auditable behavior.

Abstract

Large Language Models (LLMs) have demonstrated impressive capabilities, yet their deployment in high-stakes domains is hindered by inherent limitations in trustworthiness, including hallucinations, instability, and a lack of transparency. To address these challenges, we introduce a generic neuro-symbolic approach, which we call Autonomous Trustworthy Agents (ATA). The core of our approach lies in decoupling tasks into two distinct phases: Offline knowledge ingestion and online task processing. During knowledge ingestion, an LLM translates an informal problem specification into a formal, symbolic knowledge base. This formal representation is crucial as it can be verified and refined by human experts, ensuring its correctness and alignment with domain requirements. In the subsequent task processing phase, each incoming input is encoded into the same formal language. A symbolic decision engine then utilizes this encoded input in conjunction with the formal knowledge base to derive a reliable result. Through an extensive evaluation on a complex reasoning task, we demonstrate that a concrete implementation of ATA is competitive with state-of-the-art end-to-end reasoning models in a fully automated setup while maintaining trustworthiness. Crucially, with a human-verified and corrected knowledge base, our approach significantly outperforms even larger models, while exhibiting perfect determinism, enhanced stability against input perturbations, and inherent immunity to prompt injection attacks. By generating decisions grounded in symbolic reasoning, ATA offers a practical and controllable architecture for building the next generation of transparent, auditable, and reliable autonomous agents.

Figures (4)

• Figure 1: An overview of the proposed approach called Autonomous Trustworthy Agents (ATA).
• Figure 2: A graphical illustration of the claims processing task. A policyholder has signed or agreed to certain T&Cs from a policy provider. The task is to decide if a certain claim is covered by the contract or not.
• Figure 3: Concrete implementation of ATA for claims processing with knowledge ingestion and task processing: The T&C formalization is done only once per T&C optionally with human expert corrections, while the claim axiomization is done fully autonomously for each claim. The final decision is made by an automatic theorem prover that checks the validity of the claim w.r.t. the T&C. The overview also shows how human experts can (1) correct the T&C formalization and (2) review certain claims if certain conditions are met.
• Figure 4: The effect of the thinking budget on the accuracy and standard deviation evaluated on the travel insurance dataset from $0$ to its maximum of $25\text{k}$ tokens. The blue line shows the accuracy of gemini-2.5-flash with reasoning and a fixed thinking budget. The red line shows the accuracy of gemini-2.5-flash with reasoning and a dynamic thinking budget.