# Indistinguishable Agents: Training Agent Privacy as Twin-World Indistinguishability

DevFeed: [Indistinguishable Agents: Training Agent Privacy as Twin-World Indistinguishability](<https://devfeed.tech/articles/indistinguishable-agents-training-agent-privacy-as-twin-world-indistinguishability-93433.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.25161>)

Author: Jiaming He, Guanyu Hou, Yi Yu, Xudong Jiang

Published: 2026-10-21T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Security & Privacy](<https://devfeed.tech/topics/security-privacy.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [grpo](<https://devfeed.tech/topics/grpo.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>)

Tags: [agent-privacy](<https://devfeed.tech/tags/agent-privacy.md>), [agents](<https://devfeed.tech/tags/agents.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [grpo](<https://devfeed.tech/tags/grpo.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-agents](<https://devfeed.tech/tags/llm-agents.md>), [privacy](<https://devfeed.tech/tags/privacy.md>), [training](<https://devfeed.tech/tags/training.md>)

## AI overview

The paper proposes training LLM agents to limit what observers can infer from their actions, not just what they explicitly disclose. It creates paired "twin worlds" differing only in a private fact; a frozen distinguisher scores each observer's view by its log-likelihood ratio for the true world, and paired GRPO optimizes task success minus that score. The authors say the objective controls every observer's Bayes success across channels. In shopping and messaging tasks, the method reportedly reduces leakage missed by disclosure judges and query-level rewards while preserving task completion.

## Source excerpt

arXiv:2610.25161v1 Announce Type: new Abstract: LLM agents search, shop and send messages for their users, and every action is seen by some party. Existing privacy training rewards agents for not mentioning protected items, yet an observer can infer a private fact from what the agent does. Since an observer learns a fact only if its view changes with the fact, we propose \methodfull (\method), which trains on twin worlds that differ only in the private fact. A frozen distinguisher scores each observer's view by its log-likelihood ratio for the true world, and paired GRPO trains the agent on task success minus this ratio. We prove that this objective controls the Bayes success of every observer on every channel. On shopping and messaging tasks \method removes the leakage that disclosure judges and query-level rewards miss, while the agent still completes its tasks.