Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DMind-3: A Sovereign Edge--Local--Cloud AI System with Controlled Deliberation and Correction-Based Tuning for Safe, Low-Latency Transaction Execution

Enhao Huang, Frank Li, Tony Lin, Lowes Yang

TL;DR

DMind-3 tackles the challenge of safe, low-latency irreversible Web3 transactions by distributing intelligence across a sovereign Edge-Local-Cloud stack. Safety is anchored at the signing boundary via an edge intent firewall, while a private local verifier and a cloud-based macro-context synthesizer enable selective offloading under policy constraints. The approach is supported by two training objectives, Hierarchical Predictive Synthesis (HPS) and Contrastive Chain-of-Correction SFT (C3-SFT), which align cloud synthesis and local verification with the edge's safety role. Experimental results show a 93.7% multi-turn success on protocol-constrained tasks and favorable tradeoffs between latency and robustness, highlighting the practical viability of end-to-end sovereignty in high-stakes, latency-sensitive AI-assisted decision-making.

Abstract

This paper introduces DMind-3, a sovereign Edge-Local-Cloud intelligence stack designed to secure irreversible financial execution in Web3 environments against adversarial risks and strict latency constraints. While existing cloud-centric assistants compromise privacy and fail under network congestion, and purely local solutions lack global ecosystem context, DMind-3 resolves these tensions by decomposing capability into three cooperating layers: a deterministic signing-time intent firewall at the edge, a private high-fidelity reasoning engine on user hardware, and a policy-governed global context synthesizer in the cloud. We propose policy-driven selective offloading to route computation based on privacy sensitivity and uncertainty, supported by two novel training objectives: Hierarchical Predictive Synthesis (HPS) for fusing time-varying macro signals, and Contrastive Chain-of-Correction Supervised Fine-Tuning (C$^3$-SFT) to enhance local verification reliability. Extensive evaluations demonstrate that DMind-3 achieves a 93.7% multi-turn success rate in protocol-constrained tasks and superior domain reasoning compared to general-purpose baselines, providing a scalable framework where safety is bound to the edge execution primitive while maintaining sovereignty over sensitive user intent.

DMind-3: A Sovereign Edge--Local--Cloud AI System with Controlled Deliberation and Correction-Based Tuning for Safe, Low-Latency Transaction Execution

TL;DR

DMind-3 tackles the challenge of safe, low-latency irreversible Web3 transactions by distributing intelligence across a sovereign Edge-Local-Cloud stack. Safety is anchored at the signing boundary via an edge intent firewall, while a private local verifier and a cloud-based macro-context synthesizer enable selective offloading under policy constraints. The approach is supported by two training objectives, Hierarchical Predictive Synthesis (HPS) and Contrastive Chain-of-Correction SFT (C3-SFT), which align cloud synthesis and local verification with the edge's safety role. Experimental results show a 93.7% multi-turn success on protocol-constrained tasks and favorable tradeoffs between latency and robustness, highlighting the practical viability of end-to-end sovereignty in high-stakes, latency-sensitive AI-assisted decision-making.

Abstract

This paper introduces DMind-3, a sovereign Edge-Local-Cloud intelligence stack designed to secure irreversible financial execution in Web3 environments against adversarial risks and strict latency constraints. While existing cloud-centric assistants compromise privacy and fail under network congestion, and purely local solutions lack global ecosystem context, DMind-3 resolves these tensions by decomposing capability into three cooperating layers: a deterministic signing-time intent firewall at the edge, a private high-fidelity reasoning engine on user hardware, and a policy-governed global context synthesizer in the cloud. We propose policy-driven selective offloading to route computation based on privacy sensitivity and uncertainty, supported by two novel training objectives: Hierarchical Predictive Synthesis (HPS) for fusing time-varying macro signals, and Contrastive Chain-of-Correction Supervised Fine-Tuning (C-SFT) to enhance local verification reliability. Extensive evaluations demonstrate that DMind-3 achieves a 93.7% multi-turn success rate in protocol-constrained tasks and superior domain reasoning compared to general-purpose baselines, providing a scalable framework where safety is bound to the edge execution primitive while maintaining sovereignty over sensitive user intent.
Paper Structure (20 sections, 11 equations, 1 figure, 6 tables)

This paper contains 20 sections, 11 equations, 1 figure, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Security in Decentralized Systems
  4. Intelligence at the Edge and in Decentralized Environments
  5. Large Language Models and AI Agent Security
  6. Theory and Methods
  7. Setting and Edge-Local-Cloud System Model
  8. Policy-Constrained Routing and Risk Escalation
  9. Controlled Deliberation via Dual-State Inference
  10. Training Objectives: HPS and C3-SFT
  11. Experimental Evaluation
  12. Models, Baselines, and Protocol
  13. Backbone Capability on Knowledge and Reasoning Benchmarks
  14. Transaction-Time Structured Behavior at the Signing Boundary
  15. Additional Reproducible Checks: Scaling, Ablations, Latency, and Privacy
  16. ...and 5 more sections

Figures (1)

  • Figure 1: The DMind-3 Edge--Local--Cloud Architecture: A sovereign stack for Web3 execution. The system routes transactions based on privacy, latency, and uncertainty, ensuring that final enforcement (Policy Gate) happens at the deterministic Edge Tier.