When one AI knows something the other can't see, how do you test whether it reasons or keeps quiet?

This explores how to build or test AI systems where one party knows something the other can't see, like a gift in a wrapped box, and the model has to reason or keep quiet across that gap. The corpus has no note that names hidden-gift problems or tests a fix for them, so I'm reading the term as a private-information task. The closest material is about what happens when information is unevenly shared. It gives a clear answer on the task side and only suggestive hints on the architecture side.

The most useful task variant is one where each agent holds its own secrets. LLMs look socially skilled when one model writes every side of a conversation, but they fail systematically when each agent has private information (Why do LLMs fail when simulating agents with private information?). In the one-model setup the model skips the grounding work of asking what the other party could actually know. Any hidden-gift solution should therefore be tested with separate agents and separate state that can't be peeked at. A method that only works when one model plays every role hasn't solved the problem.

On architecture, the best corpus example enforces the boundary structurally instead of asking the model to ignore what it knows. TiMoE pre-trains experts on separate two-year slices and masks any expert whose window postdates the query. That guarantees no future knowledge leaks, and it is an architectural property, not a prompt or retrieval patch (Can routing mask future experts to prevent knowledge leakage?). It's about time, not gifts, but the shape carries over: put the gift out of reach of the part of the system that has to guess. That transfer is my extrapolation, and the paper doesn't test it.

Multi-agent splitting cuts both ways. Role specialization means no single agent sees the whole objective. SafeFlow shows this lets a harmful goal fragment into steps that each look benign, with the harm appearing only when they combine (Can task decomposition hide harmful intent across agents?). So splitting knowledge across roles is a plausible way to keep a gift hidden from most agents, but what's hidden from each part can still be reconstructed from the whole. There is also a tension with the Thread Inference Model. It shows one model can run recursive subtask trees with pruned working memory and replace a multi-agent system (Can recursive subtask trees overcome context window limits?). Given the omniscience failure above, collapsing everything into one model would bring the problem back unless the pruning really drops what the other party shouldn't know. That's an open question the notes raise but don't settle.

One caution is that hiding something in a model's private reasoning is weaker than it sounds. Encrypted reasoning blocks are interchangeable across models within a provider, so a weaker, less-guarded model can decode a stronger model's hidden traces verbatim (Can cheaper models decrypt traces from stronger models?). If the gift lives only in hidden reasoning, it is only as hidden as the wrapper around it. The reward-hacking notes that surfaced in retrieval don't bear on this question, so I've left them out.