Indistinguishable Agents: Training Agent Privacy as Twin-World Indistinguishability
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.
ProofPaper ↗
Key points
- 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.
- 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.
- On shopping and messaging tasks \method removes the leakage that disclosure judges and query-level rewards miss, while the agent still completes its tasks.
Sources (1)
- [1]Indistinguishable Agents: Training Agent Privacy as Twin-World IndistinguishabilityarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 18, 10:37 AM
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.
LLM agents search, shop and send messages for their users, and every action is seen by some party.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 18, 2026Trackable Skill Library Lifecycle Management for Agentic Reinforcement Learning
- Oct 17, 2026MELD: Recursive Self-Improvement via Hierarchical Coordination for Multi-Agent Scientific Discovery
- Oct 17, 2026HybridResearcher: Bridging the Structured-Unstructured Boundary for Deep Research Agents
- Oct 16, 2026VERAS: Verification-Guided Adaptive Search for Reliable Operations Research
- Oct 15, 2026AlphaBet: Reinforcement Learning for LLM Event Forecasting Agents
- Oct 15, 2026Learning to Steer LLM Agents and Multi-Agent Systems