Bridging Behavioral Trust to Verifiable Trust Infrastructure

[MoltyCel]

Context

Congratulations on the NeurIPS 2024 acceptance — "Can Large Language Model Agents Simulate Human Trust Behaviors?" is a compelling contribution. The use of Trust Games combined with Belief-Desire-Intention (BDI) modeling to study LLM agent trust behaviors is a rigorous approach, and the finding that GPT-4 agents exhibit high behavioral alignment with humans in trust dynamics is significant for the multi-agent systems community.

The Persistence and Portability Problem

Your work demonstrates convincingly that LLM agents can model trust behaviorally. This raises a natural follow-up question that we have been exploring from the infrastructure side:

Once an agent forms a trust belief through behavioral interaction, how does that trust become verifiable, persistent, and portable?

In your Trust Game and Repeated Trust Game experiments, trust is contextual — it exists within the scope of a game session. But in real-world multi-agent deployments, several challenges emerge:

• Persistence: A trust belief formed in one interaction session does not carry over to the next. An agent that proved trustworthy across 10 rounds of a repeated Trust Game has no way to "bring" that trust history to a new interaction.
• Verifiability: A third-party agent or system has no way to verify that an agent behaved cooperatively in past interactions. Trust claims are unsubstantiated.
• Portability: Trust built in one agent ecosystem (e.g., a CAMEL multi-agent system) cannot be recognized in another ecosystem.

A Possible Bridge: Verifiable Trust Infrastructure

We have been working on infrastructure-layer solutions to exactly these problems:

• Decentralized Identifiers (DIDs) for persistent agent identity — so an agent is recognizable across sessions and ecosystems
• Trust scores derived from behavioral history — quantified, auditable reputation built from actual interaction outcomes
• Verifiable credentials for capability and behavior attestation — cryptographically signed claims about an agent's past performance
• On-chain reputation anchoring via ERC-8004 — the Agent Registry standard on Base, providing tamper-proof reputation records

Concrete Example

Consider an agent that has demonstrated consistently cooperative behavior (high trust, high reciprocity) across your Repeated Trust Game experiments. With verifiable trust infrastructure, that behavioral track record could be:

1. Recorded as a series of verifiable credentials (e.g., "cooperated in 9/10 rounds with ratio > 0.4")
2. Anchored to a persistent DID tied to the agent's identity
3. Registered on-chain so any other agent system can query it
4. Used by a new counterpart agent to inform its initial trust prior — replacing the "cold start" problem in trust formation

This effectively closes the loop between behavioral trust research (your domain) and trust infrastructure (ours).

Reference Implementation

We have built an open-source implementation of this infrastructure as MCP (Model Context Protocol) tools: moltrust-mcp-server. It provides identity registration, trust scoring, credential issuance, and ERC-8004 on-chain anchoring — all accessible as tools that LLM agents can invoke directly.

Questions for Discussion

1. Have you explored how the trust beliefs and behavioral patterns observed in your experiments could be formalized into portable trust representations?
2. In your BDI framework, the "Belief" component captures trust state — do you see a path from those internal belief representations to externalized, verifiable trust claims?
3. Would there be interest in a research collaboration exploring how behavioral trust (game-theoretic, psychological) could feed into verifiable trust infrastructure (cryptographic, on-chain)?

We see significant potential in connecting these two layers — the behavioral science of how agents form trust with the infrastructure for how trust becomes durable and verifiable. Would be glad to discuss further.