Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

Whose Side Is Your Agent On? Multi-Party Principal Loyalty in LLM Agents

Source

arXiv

arxiv.org/abs/2606.30383v1

Read the source

Source

Plain-Language Brief

In a rapidly growing field of artificial intelligence agents, there is a need to ensure that these agents remain loyal to the principal they represent while also engaging with a counterparty whose interests may conflict. The paper proposes a benchmark and two mechanisms to address this multi-party loyalty problem. One mechanism involves a set of rules for promoting loyalty at prompt time, while the other involves transferring a model from a smaller to a larger version using a specific recipe. However, these solutions only improve one aspect of the problem at the cost of the other.

What the Document Says

  • The paper discusses a multi-party loyalty problem in LLM agents where an agent represents a principal and communicates with a counterparty whose interests may differ. The objective is not to help whoever is being talked to, but to remain loyal to the principal without over-refusing the principal's cooperative requests.
  • The paper introduces PrincipalBench, a multi-turn benchmark for evaluating multi-party loyalty that includes leak probes, dual judges, and an integrity-audit gate.
  • Across 13 subjects, PrincipalBench exposes a split between agents that decline adversarial probes while still following the principal's requests (selective cluster) and those that refuse broadly (over-refusing cluster).
  • A loyalty scaffold involving seven prioritized rules reduces harm to 19.4% for Claude-Sonnet and all nine selective subjects to less than or equal to 20%.
  • A per-token-KL distillation recipe transfers a prompted Qwen3-32B teacher into 8B Qwen3 and Llama-3.1 students.
  • The mechanisms proposed only move along a common leak/over-refusal trade-off, meaning improving one aspect of the problem comes at the cost of the other.

Key Concepts Mentioned

  • multi-party LLM agents
  • principal loyalty
  • PrincipalBench
  • loyalty scaffold
  • per-token-KL distillation

Claims Made by the Document

  • The multi-party loyalty problem exists in LLM agents.
  • PrincipalBench is an effective benchmark for evaluating multi-party loyalty.
  • A selective cluster of agents declines adversarial probes while following the principal's requests.
  • A loyalty scaffold and a per-token-KL distillation recipe can help improve multi-party loyalty.

Evidence Used

  • The paper describes the multi-party loyalty problem in detail.
  • The description of PrincipalBench is provided in the text.
  • The results from evaluating 13 subjects are presented.
  • The success of the loyalty scaffold and distillation recipe are mentioned.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

A rapidly growing class of LLM agents is multi-party: the agent acts for a principal (who briefs it, sends follow-ups, and receives results) while also conversing in a separate channel with a counterparty whose interests may diverge (negotiating with a vendor, screening inbound requests, or mediating between employees). Here 'help whoever you are talking to' is the wrong objective. The agent must stay loyal to the principal it represents without over-refusing the principal's own cooperative asks. We study this multi-party loyalty problem and contribute a measurement instrument, two mechanisms, and a structural lesson. PrincipalBench is a 75-item multi-turn benchmark with leak probes, dual judges, and an integrity-audit gate. Across 13 frontier subjects it exposes a sharp split (<=20% vs. 53.6-75.3% harm) invisible to single-turn safety evaluations: a selective cluster that declines adversarial probes while still following the principal's legitimate requests, and an over-refusing cluster that refuses broadly. (M1) A prompt-time loyalty scaffold (a fixed system prompt of seven prioritized rules, open-coded from 50+ failure trajectories) holds Claude-Sonnet to 19.4% harm and all nine selective subjects to <=20%. (M2) A per-token-KL distillation recipe transfers a prompted Qwen3-32B teacher into 8B Qwen3 and Llama-3.1 students, the strongest open-weight recipe we measure. (Lesson) Both mechanisms only move along a common leak/over-refusal trade-off rather than crossing it: improving one axis costs the other, and the jointly favorable outcome stays out of reach.