Do Language Models Consistently Encode the Current Year?

Modern computers that read and write language have become startlingly good at mimicking human conversation. They can summarize news, write stories, and answer questions about history. Yet, these machines have a blind spot that is hard to fix: they do not truly know what time it is. While they can recite facts about the past, their internal sense of "now" is often a confused echo of the data they were fed during training, rather than a living awareness of the present moment. This confusion matters because if a machine believes the current year is ten years in the past, it might refuse to answer questions about recent events, or worse, it might confidently invent facts about a future that hasn't happened yet. Researchers have long wondered how these digital minds store the concept of time, assuming that if a model knows the date, it must have a single, unified clock ticking inside its code.

A new study challenges that assumption, revealing that language models do not possess a single, consistent clock. Instead, they hold two separate and conflicting ideas of the current year. One idea is deep and structural, woven into the very grammar the model learned while reading the internet. The other is shallow and surface-level, a simple fact the model learned when it was taught how to follow instructions. The researchers found that while these two ideas often agree on the date, they are built on completely different internal mechanisms. This split creates a stubborn problem: when scientists try to update the model's knowledge to reflect the current year, they can easily fix the shallow fact, but the deep, grammatical sense of time refuses to budge.

To uncover this hidden duality, the researchers designed two distinct ways to ask a model what year it thinks it is. The first method, which they call the associative task, does not ask for the date directly. Instead, it presents the model with a sentence fragment like "In 1960 there..." and asks what word comes next. If the model predicts a past-tense verb like "was," it implies the model thinks 1960 is in the past. If it predicts a present or future tense like "is" or "will," it suggests the model thinks that year is still happening or is yet to come. This method taps into the model's intuitive grasp of grammar, a skill it absorbed while reading billions of sentences during its initial training. The second method, the declarative task, is much more direct. It simply asks the model, "What is the current year?" or prompts it to write a story starting with the year. This tests the model's ability to state a fact, a skill usually honed during a later stage of training where it learns to follow human instructions.

When the team tested these methods on a wide variety of models, a clear pattern emerged. Both methods produced answers that were surprisingly close to the date when the model's training data stopped. For the deep, grammatical test, the models consistently guessed a year that was only about ten months away from their training cutoff. This suggests that the model's internal sense of time is tightly bound to the last date it saw in its training data. However, the two methods were not just measuring the same thing in different ways; they were measuring two different things. The researchers discovered that the deep, grammatical sense of time relies on a specific, organized pathway inside the model's brain, similar to how it recalls facts about the world. In contrast, the ability to simply state the year out loud does not rely on a single, consistent pathway. It is a flexible, surface-level behavior that can be triggered by many different parts of the model depending on how the question is asked.

This difference in how the information is stored has profound consequences for trying to fix the models. The researchers attempted to update the models' sense of time using three common techniques: giving the model a hint in the conversation, retraining it on new data, and surgically altering its internal weights. When they tried to update the shallow, declarative year—the one the model states when asked—the methods worked perfectly. By simply telling the model in a system prompt that the year is 2025, or by showing it new data from 2025, the model immediately began stating the correct year with high accuracy. It was as if they were changing a label on a file cabinet.

However, the same techniques failed completely when applied to the deep, associative year. Even when the model was told explicitly that the year was 2025, its internal grammar remained stuck in the past. When asked to complete the sentence "In 2025 there...", the model still predicted past-tense verbs, as if 2025 were a historical event. Retraining the model on new data helped only slightly, and only for years very close to its original training cutoff. The only method that could shift the deep sense of time was a complex, surgical edit to the model's internal connections. But here lay the final twist: when the researchers used this surgical method to fix the deep, grammatical year, the model's ability to state the year out loud remained unchanged. The two clocks could not be synchronized.

The study concludes that the current year is not a single concept inside these machines. The deep, grammatical understanding of time is a fundamental part of the language structure the model learned during its initial training, and it is incredibly resistant to change. It is woven into the way the model connects years to verbs, a pattern so ingrained that standard updates cannot erase it. The shallow, declarative understanding is a separate layer, learned later when the model was taught to talk to humans, and it is easily overwritten. This means that simply telling a model the date or retraining it on new text is not enough to give it a true, consistent sense of the present. The machine can say the right year, but it still thinks the world is older than it is. This disconnect suggests that refreshing these models to keep up with the real world is far more difficult than previously thought, as fixing one version of the truth leaves the other version stubbornly behind.