Original title: Transfiver: Human-AI Co-Inference through a Shared Editable State
Minji Park, Seunghyun Yoon, Hyuk Lim · September 2026
Working with an AI over a long time is hard. The AI updates its internal information invisibly, so you cannot see or change what it believes. To fix this, researchers built Transfiver, a system where a human and an AI share one single saved state. Think of this state as an open scratchpad that both sides can read and edit. It works in two ways. First, the AI reads the conversation and automatically adds new notes or updates old ones. Second, the user can directly open and edit any specific note on the pad. Any human edit changes the exact memory the AI reads next, instead of just tacking on extra instructions. The AI keeps its learned settings fixed without needing retraining.
You and a friend are following the same recipe while cooking together over a video call. Normally, if you spot a mistake and fix it on your copy, your friend might keep using the old version, leading to mismatched results. Here, it's as if both of you edit the same single recipe document in real time, so a correction by one person instantly appears for the other to follow.
This makes long-term AI conversations more reliable and correctable. When users can see and edit what an AI remembers, mistakes get fixed at the source, which could improve trust in tools like personal assistants or collaborative software.
The abstract does not provide performance results or test scalability; it notes that extending to rich natural-language, relational, and large-scale states remains open work. The architecture may not yet handle complex real-world scenarios.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Transfiver: Human-AI Co-Inference through a Shared Editable State — Minji Park, Seunghyun Yoon, Hyuk Lim
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.