Knowledge cutoff
Temporal limit of a model's knowledge
From Wikipedia, the free encyclopedia
In machine learning, a knowledge cutoff (or data cutoff) is the point in time beyond which a model has not been trained on new data. The term is mostly used in reference to a large language model (LLM).[1] Any information about events after this date is absent from the model's training data.[1] It cannot access information about later events without a system for real-time data access like retrieval-augmented generation (RAG).[2] While useful for training and tuning LLMs, knowledge cutoffs introduce new limitations like hallucinations, information gaps, and temporal bias.[1]
Overview
A model with a fixed knowledge cutoff is unable to provide information on facts or developments that have emerged since that time, since the model is not connected to the internet.[1] Therefore, it may occasionally produce incorrect answers.[1] This is caused by the fact that training on newer data would cause a major price concern, given that training the most powerful large language models may soon cost over a billion dollars according to Time.[3]
Notable AI model cutoff dates include:
Effects of knowledge cutoffs
Knowledge gaps
Knowledge cutoffs create information gaps. The model lacks any knowledge of events or discoveries that are not included in its training data.[1] This can lead to hallucinations, where the model generates plausible but verifiably false statements. Such inaccuracies occur because LLMs are designed to predict and generate the most probable sequence of words based on their training patterns, which may result in confident but incorrect outputs when queried beyond their knowledge boundaries, otherwise known as hallucinations.[6]
Effective vs. reported cutoffs
A research paper on arXiv indicates that a model's functional knowledge may not be uniformly limited by its stated cutoff date. This effective cutoff often differs for various subjects and is influenced by the distribution of information within the training data itself, meaning some topics may reflect later knowledge than others while predating the cutoff may be absent.[7] Due to the high cost of retraining large language models, these models are rarely completely retrained to increase their knowledge cutoff.[8] Some models can also use integrated search tools to access more recent information, which blurs the line of their inherent knowledge base. For example, GPT-4, can access its search tool and give real-time info.[4]
Attempts to overcome knowledge cutoffs
Retrieval-augmented generation
RAG is a common technique used to overcome the limitations of a knowledge cutoff.[2] In a RAG system, the language model is connected to an external knowledge base or search engine to retrieve live data. This architecture allows the model to find current information relevant to a query and incorporate it into its response, often with citations.[2] Grounding a model in external data helps reduce the frequency of hallucinations and improves output accuracy. However, the external knowledge base might be outdated or contain biases, which may also lead to incorrect information or hallucinations.[9] For example, Google AI Overviews have created false claims, and the results are sometimes unreliable, since it either fail at interpreting the prompt correctly, or at pulling high quality sources.[9] However, a method to mitigate this is to apply techniques like RHLF (reinforcement learning from human feedback), which can enhance the quality and reliability of a large language model's responses.[9]
Continual learning
Another approach is continual learning, which involves methods like adapters and LoRA.[10] These fine-tuning techniques permit efficient, incremental updates to a model without the high cost of a full retraining cycle. However, this does not give real-time awareness, since adding modules to the system may result in algorithmic bias and catastrophic forgetting, as the weights in the model become biased towards the new set of data.[10]