Memorization vs. Reasoning: Updating LLMs with New Knowledge

Large language models (LLMs) encode vast amounts of pre-trained knowledge in their parameters, but updating them as real-world information evolves remains a challenge. Existing methodologies and benchmarks primarily target entity substitutions, failing to capture the full breadth of complex real-world dynamics. In this paper, we introduce knowledge update playground (KUP), an automatic pipeline for simulating realistic knowledge updates reflected in an evidence corpus. KUP's evaluation framework includes direct and indirect probes to test both memorization of updated facts and reasoning over them, for any update learning methods. Next, we present a lightweight method called memory conditioned training (MCT), which conditions tokens in the update corpus on self-generated "memory" tokens during training. Our strategy encourages LLMs to surface and reason over newly memorized knowledge at inference. Our results on two LLM families show that (1) KUP benchmark is highly challenging, with the best CPT models achieving < 2% in indirect probing setting (reasoning) and ( 2 ) MCT training significantly outperforms prior continued pre-training (CPT) baselines, improving direct probing (memorization) results by up to 25.4%. step. Our aim is to generate updates that are realistic and contradictory to the prior fact f old e . We find that strong LLMs like GPT-4O perform better at proposing realistic and logical updates when prompted to additionally generate fictitious event sequences that realize the update from f old e to f new e . Verifying f old e and f new e In order to align with the goals of KUP, we need to ensure that LLM M T 's parametric knowledge includes f old e and contradicts f new e . We probe both our test LLMs to guarantee this. Concretely, we generate answers to True/False questions using test models M T for both f old e and f new e facts. We only retain pairs where it generates the expected label (True for f old e