Best Format for a University Level Marking Rubric

Discover the ideal formatting techniques to maximize LLM performance

Key Highlights

Markdown excels at overall structure – highly readable with natural hierarchy for LLM comprehension.
CSV vs Markdown for tables – while CSV is efficient for simple data, Markdown formats tables in a way that enhances context and readability.
Alternative options such as LaTeX – superior for complex mathematical expressions and intricate table structures, with different LLM models benefiting uniquely.

Detailed Analysis

Understanding the Importance of Format

When presenting a large university-level marking rubric to a Large Language Model (LLM), the format employed plays a pivotal role in how effectively the rubric is parsed and processed. It is crucial for the model to understand the rubric's requirements at its maximum potential in order to generate the best possible research output and, ultimately, achieve the highest possible marks. In this context, the discussion centers on which format maximizes that understanding, especially when it comes to formatting the table elements within the rubric.

Markdown as a Primary Format

Benefits for the Overall Rubric

Markdown is a lightweight markup language that offers numerous advantages:

Enhanced Readability: The simplicity of Markdown ensures that hierarchical structures like headings, bullet points, and tables are naturally integrated, allowing LLMs to discern the relative importance of content quickly.
Structure and Context: The explicit use of headers and formatting symbols assists LLMs in understanding and prioritizing the information presented.
Ease of Use: Writing in Markdown is intuitive, reducing the risk of formatting errors that could confuse the LLM during parsing.

Given these benefits, Markdown is generally recommended for the overall presentation of the rubric, ensuring a clear, structured, and accessible interpretation.

Evaluating Tabular Data Formatting: CSV vs Markdown

CSV: Advantages and Limitations

CSV (Comma-Separated Values) is a straightforward and widely accepted method for representing table data:

Efficiency: CSV files are very compact and work well when the data structure is simple and does not require additional context.
Data Parsing: LLMs that are specifically designed for data processing can utilize CSV effectively when performing tasks such as summarization or basic interpretation.

However, CSV has limitations especially in contexts where:

The data includes hierarchical relationships and nested information; or
Additional formatting cues (e.g., bold, italics, or visual separation) are necessary to comprehend the criteria fully.

Markdown Tables: A Superior Alternative

Markdown is not only effective for overall text, but its tables offer enhanced functionality:

Visual Clarity: Markdown tables allow for clearly defined headers, rows, and columns, presenting information in an organized manner. This facilitates an easier understanding of complex marking criteria.
Context Preservation: The use of Markdown retains key contextual elements such as table captions and inline comments that provide essential insights into the marks distribution.
Flexible Implementation: Markdown tables are easy to integrate into the overall rubric structure, providing a coherent flow between textual content and tabular data.

Despite CSV being useful for simple data, Markdown tables offer the benefit of enhanced readability and context—key factors in ensuring LLMs capture and process every nuance of a complex, university-level marking rubric.

Exploring Alternative Formats for Tabular Data

LaTeX for Complex Table Structures

In scenarios where the table includes intricate details such as mathematical notations, formulas, or multi-level criteria, LaTeX becomes an attractive option:

Complexity Handling: LaTeX is designed for technical and scientific documentation. Its robust syntax is ideal for formatting tables that include mathematical expressions, detailed formulas, and other complex elements, making it highly suitable for specialized marking criteria.
Precision: The ability to precisely control table layout and text formatting ensures that every detail of the rubric is accurately conveyed.
Structured Presentation: LaTeX’s precise control over numerical and symbolic data supports scenarios where a high level of accuracy is required.

Although LaTeX is highly powerful, its complexity means that it is best employed in situations where the data is mathematically or technically heavy, rather than for general rubrics. However, if your rubric includes heavy scientific content, its use can be justified.

Evaluating LLM Performance with Table Formatting

Performance of Leading LLM Models

When determining which tabular format is best, recent evidence suggests that leading LLMs demonstrate varied performance based on how tabular data is provided. Below is an analysis focusing solely on table formatting:

Model	Preference for Markdown Tables	Preference for CSV Tables	Preference for LaTeX Tables
ChatGPT (o1, o3-mini, deep research)	High – excellent at parsing structured Markdown tables; provides contextual hierarchy.	Moderate – effective when tables are simple, but lacks contextual intricacies.	Low – while capable of understanding LaTeX, not primarily optimized for it in general rubric use.
Claude (Sonnet 3.5/3.7)	High – performs well with structured Markdown, capturing detailed rubric requirements.	Moderate – can interpret CSV but may miss subtleties in complex table structures.	Moderate – benefits from LaTeX formatting when technical precision is needed.
Gemini (2.0 Pro, 2.0 flash-thinking)	High – readily understands Markdown tables, offering clear interpretation of multi-layer criteria.	Low – less effective with CSV for capturing context in complex academic rubrics.	High – performs admirably with LaTeX tables especially where mathematical or detailed technical information is involved.

This table demonstrates that while Markdown tables perform consistently well for most applications, LaTeX tables provide significant advantages in scenarios involving complex expressions and dense technical details, especially for models like Gemini. Meanwhile, CSV's simplicity can be useful in limited contexts, but it generally lacks the rich formatting cues found in Markdown or LaTeX.

Recommendations Based on Use Case

Choosing the Right Format

For overall maximum potential in understanding and execution of the marking rubric:

Markdown is recommended for the general structure: Its ability to provide a clear hierarchical structure makes it ideal for the complete rubric.
Use Markdown tables for clear, readable presentation of tabular data: For most rubric tables, Markdown provides the balance of clarity and contextual information necessary for accurate interpretation.
Consider LaTeX for exceptionally complex tables: If your rubric involves detailed mathematical formulas or technical data that require precision, LaTeX is advantageous, particularly when working with models like Gemini that have strong capabilities in managing detailed formatting.
Limit the use of CSV to simpler tables: While CSV format may be chosen for its simplicity, it is generally less effective for in-depth rubric tasks where context and structure are critical.

By balancing the strengths of these formats according to the specific demands of the rubric, you can ensure that the LLM interprets every element—from general guidelines to complex tabular data—with clarity and precision. Experimentation, along with targeted adjustments based on model feedback, can further refine this approach.