How to Get LLMs to Accurately Infer Relative Days of the Week

Provide each date option with its explicit relative label (e.g., "Today at 12:00", "Tomorrow at 14:00", "This Wednesday at 16:00") rather than listing a separate calendar. This direct interpolation eliminates the hallucination of days of the week that occurs when the model must cross-reference a separate date list.

The Full Answer

When you ask an LLM to map a natural language phrase like "this Friday" or "next Monday" to an actual calendar date, the model must perform two tasks: (1) understand the current reference date (usually "today") and (2) compute the offset in days. Many LLMs, even advanced ones, struggle with this because they are not natively arithmetic engines. They pattern-match on language, and the arithmetic of day-of-week offsets is a weak point.

The Problem with a Separate Date List

A common initial approach is to give the LLM a list of candidate dates and a separate list of the next few days with their day names. For example:

Options:
03/03/2025 12:00
04/04/2025 14:00
05/03/2025 16:00

Next 3 days and their corresponding days of the week:
3 March 2025 - Monday
4 March 2025 - Tuesday
5 March 2025 - Wednesday

According to the accepted solution on Stack Overflow (source 2), this approach "typically led to hallucinated days of the week." The model would often say "this Friday" is an incorrect date, or get the date correct but assign the wrong day name. The reason is that the LLM must hold two separate pieces of information in its context window (the options list and the day-of-week mapping) and then perform a mental join. This is a fragile operation for a transformer-based model.

The Working Solution: Interpolate Relativity Directly

The accepted solution reports that a small change in formatting produced much better results. Instead of a separate mapping, each date option is annotated with its relative label:

Options:
03/03/2025 12:00 - Today at 12:00
04/04/2025 14:00 - Tomorrow at 14:00
05/03/2025 16:00 - This Wednesday at 16:00

By interpolating each date to explicitly declare the relativity next to each date, the model no longer needs to cross-reference a separate list. The relative label is directly attached to the date string. The model sees "05/03/2025 16:00 - This Wednesday at 16:00" and can immediately associate the date with the relative phrase. This reduces the cognitive load on the LLM and eliminates the hallucination of day names.

Why This Works

LLMs are sequence models. They predict the next token based on the tokens that came before. When you place the relative label immediately after the date, the model can learn a direct association between the two. In contrast, when the date and the day name are in separate sections, the model must attend to tokens far apart in the sequence, which is more error-prone.

Additionally, the explicit label "Today at 12:00" or "Tomorrow at 14:00" is a natural language phrase that the model has seen millions of times in its training data. It is a familiar pattern. The model does not need to compute the offset; it just needs to match the phrase to the date.

When to Use This Approach

Use this interpolation technique whenever your LLM application needs to resolve relative dates from a fixed set of options. Examples include:

• Scheduling assistants that must pick from a list of available appointment slots.
• Event planners that need to map "next Tuesday" to a specific date.
• Booking systems that accept phrases like "this weekend" or "the day after tomorrow."

If your application allows free-form date input (not a fixed list), you may need a different strategy, such as using a dedicated date parser library (e.g., Python's dateparser or parsedatetime) before passing the resolved date to the LLM. The interpolation technique is best when you control the set of options.