SheetCompass is an agentic framework designed to improve how large language models (LLMs) reason about and automate complex spreadsheet tasks. By moving away from traditional methods that flatten spreadsheets into simple text, the framework uses hierarchical graphs to preserve the spatial and structural relationships inherent in grid-based data.
Addressing Structural Information Loss
Existing spreadsheet automation methods often convert multi-dimensional tables into sequential strings like Markdown or JSON. According to researchers from the State Key Laboratory of Cognitive Intelligence at the University of Science and Technology of China, this process destroys the "orthogonal positioning system" of rows and columns, causing LLMs to lose track of spatial neighborhoods and cross-table dependencies. SheetCompass addresses this by modeling spreadsheets as a hierarchical graph, which explicitly maps table-level and column-level nodes to maintain the original structural topology.
The Multi-Agent Workflow
SheetCompass coordinates three specialized agents to manage the reasoning process:
Navigational Explorer: Identifies the relevant parts of the spreadsheet by extracting topological features from the hierarchical graph, ensuring the model focuses on the correct data without processing unnecessary information.
Logical Programmer: Generates executable code while adhering to strict constraints, ensuring that all variables are anchored to verified nodes within the spreadsheet graph.
Critical Reflector: Validates the output against a checklist derived from user instructions. If the execution state does not match the requirements, the reflector provides feedback to trigger an iterative refinement process.
Dual-Level Memory Mechanism
To improve reliability, the framework employs two types of memory. The "expert knowledge memory" acts as a permanent repository for domain-specific rules, common formulas, and lessons learned from past code-execution errors. The "reasoning experience memory" is dynamic, recording the specific steps, errors, and feedback from the current task. Successful strategies from the reasoning experience memory can be transferred to the expert knowledge memory, allowing the system to improve over time and avoid repeating past mistakes.
Performance and Results
In empirical evaluations across the SCB, SB, and SheetRM benchmarks, SheetCompass achieved state-of-the-art performance compared to existing methods. For instance, when using GPT-5 as a backbone, SheetCompass reached a Pass@1 score of 71.3% on the SCB benchmark and 52.3% on the SB benchmark, outperforming other models like SheetAgent and SheetCopilot. The researchers attribute these gains to the framework's ability to maintain the intrinsic grid properties of spreadsheets, which enables more accurate cross-table reasoning and complex task execution.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!