跳转至

文章背景与核心概要

长期的人机交互往往会失效，因为引导模型推理的内部信息是以隐式方式更新的，对用户而言始终是不透明的。为了解决这一痛点，研究人员引入了 Transfiver（交互式、可验证、可编辑表示的透明框架）。Transfiver 建立了一种架构，其中人类和人工智能通过单一的、持久的且可编辑的共享状态（\(S_t\)）进行协同推理。

这种设计使得人类的纠正可以直接改变计算状态，而不仅仅是作为一条独立的指令，从而弥合了隐式模型更新与显式用户控制之间的鸿沟。该研究不仅提升了人机交互的透明度与可控性，还为构建更加稳定、可干预的 AI 系统提供了全新的技术范式。

Transfiver: Human-AI Co-Inference through a Shared Editable State

Transfiver: Human-AI Co-Inference through a Shared Editable State

Summary

Summary

长期的人机交互往往会失效，因为引导模型推理的内部信息是由模型隐式更新的，用户无法直接检查或控制。为了解决这个问题，研究人员引入了 Transfiver（交互式、可验证、可编辑表示的透明框架，TRANSparent Framework for Interactive, Verifiable, Editable Representation）。Transfiver 建立了一种架构，使人类和 AI 能够通过单一的、持久的且可编辑的共享状态（\(S_t\)）进行协同推理。这使得人类的纠正可以直接改变计算状态，而不是作为单独的指令，从而弥合了隐式模型更新与显式用户控制之间的鸿沟。

Long-term human-AI interactions often break down because the internal information guiding the model's inference is updated implicitly and remains opaque to the user. To solve this, researchers introduce Transfiver (TRANSparent Framework for Interactive, Verifiable, Editable Representation). Transfiver establishes an architecture where both the human and the AI co-infer via a single, persistent, and editable shared state (\(S_t\)). This allows human corrections to directly alter the computational state rather than acting as a separate instruction, bridging the gap between implicit model updates and explicit user control.

Metadata

Metadata

  • arXiv ID: arXiv:2609.03797 [cs.AI]
  • Subjects: 人工智能 (cs.AI); 计算与语言 (cs.CL); 人机交互 (cs.HC)
    • Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Human-Computer Interaction (cs.HC)
  • Authors: Minji Park, Seunghyun Yoon, Hyuk Lim
    • Authors: Minji Park, Seunghyun Yoon, Hyuk Lim
  • Submitted: 2026年9月3日
    • Submitted: September 3, 2026
  • DOI: 10.48550/arXiv.2609.03797

Abstract

Abstract

长期的人机交互之所以困难，是因为引导推理的信息是由模型隐式更新的，用户既不能直接检查，也不能直接控制。我们介绍了 TRANSparent Framework for Interactive, Verifiable, Editable Representation (Transfiver)，这是一种通过共享可编辑状态进行人机协同推理的架构。其核心思想是：特定于交互的信息保存在一个单一的持久状态（\(S_t\)）中，模型和人类都可以对其进行更新。

Long-term human-AI interaction is difficult because the information that guides inference is updated implicitly by the model and is not directly inspectable or controllable by the user. We introduce the TRANSparent Framework for Interactive, Verifiable, Editable Representation (Transfiver), an architecture for human-AI co-inference through a shared editable state. Its central idea is that interaction-specific information is maintained in a single persistent state (\(S_t\)) that both the model and the human update.

Transfiver 区分了状态演化的两种模式： 1. 隐式流更新（Implicit Stream Update）： 模型解释正在进行的交互，并决定新信息是修改现有的状态项还是创建一个新项。 2. 显式定向编辑（Explicit Directed Edit）： 人类检查并修改指定的目标项。

Transfiver distinguishes two modes of state evolution: 1. Implicit Stream Update: The model interprets ongoing interaction and decides whether new information revises an existing state item or creates a new one. 2. Explicit Directed Edit: A human inspects and modifies an addressed item.

这两者都作用于相同的底层状态，因此人类的纠正会改变后续计算所读取的状态，而不是添加另一条指令或单独的记录。

Both act on the same underlying state, so a human correction changes the state that subsequent computation reads, rather than adding another instruction or separate record.

该架构将普通使用前预先学习的共享参数（\(\theta\)），与在部署过程中不断演化且无需重新训练参数的持久状态（\(S_t\)）分离开来。将 Transfiver 扩展到丰富的自然语言、关系型和大规模共享状态，目前仍然是一个开放的研究方向。

The architecture separates shared parameters (\(\theta\)), learned before ordinary use, from the persistent state (\(S_t\)), which evolves during deployment without parameter retraining. Extending Transfiver to rich natural-language, relational, and large-scale shared states remains open.