TRUST: A Decentralized Framework for Auditing Large Language Model Reasoning

TL;DR

Trust addresses the unsolved problem of auditing large language model reasoning by decentralizing semantic verification through TRUST, a framework combining Hierarchical Directed Acyclic Graphs (HDAGs) for modular reasoning, a three-tier auditor network (Computer, LLM, Human) with commit–reveal voting, and a blockchain-backed, privacy-preserving audit trail. It provides theoretical guarantees of safety and profitability for honest auditors while deterring malicious behavior, and demonstrates empirical gains in correctness, safety, and bias mitigation across multiple models and domains, including human-in-the-loop validation. The work enables scalable, transparent, and privacy-preserving auditing of proprietary reasoning traces, paving the way for safer deployment of LRMs in high-stakes settings. Overall, TRUST integrates robust consensus, structured reasoning decomposition, and economic incentives to deliver verifiable, public accountability without exposing model internals.

Abstract

Large Language Models generate complex reasoning chains that reveal their decision-making, yet verifying the faithfulness and harmlessness of these intermediate steps remains a critical unsolved problem. Existing auditing methods are centralized, opaque, and hard to scale, creating significant risks for deploying proprietary models in high-stakes domains. We identify four core challenges: (1) Robustness: Centralized auditors are single points of failure, prone to bias or attacks. (2) Scalability: Reasoning traces are too long for manual verification. (3) Opacity: Closed auditing undermines public trust. (4) Privacy: Exposing full reasoning risks model theft or distillation. We propose TRUST, a transparent, decentralized auditing framework that overcomes these limitations via: (1) A consensus mechanism among diverse auditors, guaranteeing correctness under up to $30\%$ malicious participants. (2) A hierarchical DAG decomposition of reasoning traces, enabling scalable, parallel auditing. (3) A blockchain ledger that records all verification decisions for public accountability. (4) Privacy-preserving segmentation, sharing only partial reasoning steps to protect proprietary logic. We provide theoretical guarantees for the security and economic incentives of the TRUST framework. Experiments across multiple LLMs (GPT-OSS, DeepSeek-r1, Qwen) and reasoning tasks (math, medical, science, humanities) show TRUST effectively detects reasoning flaws and remains robust against adversarial auditors. Our work pioneers decentralized AI auditing, offering a practical path toward safe and trustworthy LLM deployment.

Figures (10)

• Figure 1: Reasoning traces (left) are decomposed into hierarchical segments (middle) and verified by a heterogeneous, multi-tier auditor network (right). Audit outcomes are recorded on-chain, while raw trace is stored off-chain for privacy.
• Figure 2: Overview of TRUST. The TRUST framework decomposes Chain-of-Thought reasoning traces into hierarchical directed acyclic graphs (DAGs) for modular verification across a three-tier auditor system. The process begins with a reasoning query (left panel) that generates intermediate reasoning steps, which are then decomposed into graph components and distributed across automated computers, LLM-based, and human auditors. TRUST utilizes IPFS for decentralized storage of reasoning traces and blockchain technology for immutable audit records, vote aggregation, and consensus mechanisms. Auditors verify reasoning segments independently, while cryptographic protocols ensure the privacy preservation of proprietary model internals. The final verification result provides confidence guarantees for reasoning trace faithfulness and correctness for the end user.
• Figure 3: Example of an HDAG. Node color indicates the level of difficulty assigned to different types of auditors. Node shape denotes the type of step and edges represent relationships between nodes.
• Figure 5: Correctness of Single, Ensemble (Centralized) with decentralized TRUST framework.
• Figure 6: The parameters $\epsilon_{\mathrm{C}}=0$, $\epsilon_{\mathrm{L}}=0.05$, $\epsilon_{\mathrm{H}}=0.30$, and $\rho_{\mathrm{H}}=0.1$. (Left) Repupation scores. (Right) Profit curves.

Theorems & Definitions (9)

• Theorem 4.1: Safety–Profitability Guarantee
• Lemma 4.1: Exact pass probability for type $t$
• proof
• Proposition 4.1: Hoeffding and Chernoff bound
• proof
• Lemma 4.2: MGF of a Bounded Centred R.V.
• proof
• Theorem 4.1: Safety–Profitability Guarantee
• proof : Proof of \ref{['thm:safty_profit']}