Abstract
Traditional AI reading systems often process text sequentially, a method that diverges significantly from how humans interact with technical documentation. This paper introduces SPRITe (Structure to Plan the Reading of Instructional TExt), a "reading agent" that treats reading as an interactive, goal-driven activity. Integrated into the FLOABN architecture, SPRITe interleaves reading with physical action, using structural cues to navigate manuals efficiently. By mimicking human "skimming" behaviors, the system demonstrates how autonomous agents can use text to facilitate real-world problem-solving without the computational overhead of full-text comprehension.
Introduction: The Problem of Sequential Reading
Most early AI reading systems operated on a "start-to-finish" paradigm, ingesting text sequentially to build a complete internal model. However, humans rarely read technical manuals in this manner. For example, when assembling a complex device, a user typically looks for specific instructions only when they encounter a hurdle. This research addresses the gap between autonomous action and information retrieval, arguing that an intelligent agent should interact with text to facilitate a larger goal rather than merely "reading" it. The challenge lies in building a system that knows when to read, what to skip, and how to adapt its strategy when a task breaks down.
Methods: The SPRITe and FLOABN Architecture
The proposed system, SPRITe, is integrated into a larger adaptive planning architecture called FLOABN.
As shown in Figure 1, the system relies on the interaction between an "Action Agent" (Scavenger) and the "Reading Agent" (SPRITe). When the Action Agent encounters a failure or requires information, it sends a Request Frame containing the current goal and device state to SPRITe. SPRITe then processes the manual and returns a Response Frame.
Internally, SPRITe utilizes several key components to manage this goal-directed reading:
- Reading Plans: SPRITe selects from a library of specialized "reading plans" (e.g., "find-steps-short-text") that dictate how to navigate a document's superstructure—headers, lists, and formatting cues.
- Marker Passing for Coherence: To link text to physical activity, SPRITe uses marker-passing techniques to connect terms in the manual to the agent's internal knowledge of the device.
- Learning via Plan Memory: Successful reading strategies are saved as new, specialized reading plans, allowing the agent to improve its efficiency over time.
Evaluation and Data Sources
The research is grounded in two primary data sources:
- Human Protocol Study: A study of 12 subjects tasked with using a complex fax machine. Verbal protocols and line-by-line reading data confirmed that zero subjects read the entire manual; instead, they used structural clues to navigate and only read deeply during task "breakdowns."
- Simulated Airfone: The system was evaluated using a simulation of an Airfone (a 1990s-era airplane telephone), utilizing its actual 70-page instructional manual.
Key Insights and Results
The core finding is that reading is a behavior with goals, plans, and failures, much like physical action. By focusing on document superstructure, SPRITe avoids the computational overhead of full-text comprehension.
In the Airfone simulation, SPRITe successfully helped the action agent recover from mental model failures. For instance, when the agent incorrectly assumed a "pay phone" model (inserting a card after lifting the receiver), SPRITe identified the correct sequence in the manual and generated a new, more efficient reading plan for that specific device.
Limitations and Future Work
While effective for instructional text, SPRITe's reliance on rigid document superstructures may limit its application to narrative or persuasive texts. Additionally, the system currently depends on a pre-existing semantic network of devices. Future work aims to automate the construction of these device models and move toward agents that can improve general reading skills across diverse domains.
Summary updated