Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Memorization vs. Reasoning: Updating LLMs with New Knowledge

Aochong Oliver Li, Tanya Goyal

TL;DR

This paper introduces Knowledge Update Playground (KUP), a framework to automatically generate realistic knowledge updates and an evaluation suite (KUPEval) to test LLM memorization and reasoning over updated facts. It also proposes Memory Conditioned Training (MCT), a lightweight CPT method that prefixes memory tokens generated from the base model to training data to encourage surface and reasoning over updates. Empirical results on two open-source LLMs show KUP is challenging for standard CPT approaches, with indirect reasoning remaining particularly hard, while MCT significantly improves direct memorization and benefits from chain-of-thought strategies. The work highlights that current CPT methods largely memorize updates and struggle to reason about their implications, suggesting a need for future methods that better integrate updated knowledge into inferential processes. The authors provide open-source code and data to facilitate further research in continual knowledge updating for LLMs.

Abstract

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 corpora. KUP's evaluation framework includes direct and indirect probes to both test 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 strong LLMs 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 $25.4\%$.

Memorization vs. Reasoning: Updating LLMs with New Knowledge

TL;DR

This paper introduces Knowledge Update Playground (KUP), a framework to automatically generate realistic knowledge updates and an evaluation suite (KUPEval) to test LLM memorization and reasoning over updated facts. It also proposes Memory Conditioned Training (MCT), a lightweight CPT method that prefixes memory tokens generated from the base model to training data to encourage surface and reasoning over updates. Empirical results on two open-source LLMs show KUP is challenging for standard CPT approaches, with indirect reasoning remaining particularly hard, while MCT significantly improves direct memorization and benefits from chain-of-thought strategies. The work highlights that current CPT methods largely memorize updates and struggle to reason about their implications, suggesting a need for future methods that better integrate updated knowledge into inferential processes. The authors provide open-source code and data to facilitate further research in continual knowledge updating for LLMs.

Abstract

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 corpora. KUP's evaluation framework includes direct and indirect probes to both test 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 strong LLMs show that (1) KUP benchmark is highly challenging, with the best CPT models achieving 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 .

Paper Structure

This paper contains 52 sections, 3 figures, 23 tables.

Table of Contents

  1. Introduction
  2. Knowledge Update Playground (KUP)
  3. Issues with Existing Benchmarks
  4. Notation
  5. Curating KUP's Training Corpora
  6. Step 0: Identify Candidate Entities
  7. Step 1: Generate Knowledge Update
  8. Verifying $\bm{f^{\mathrm{old}}_e}$ and $\bm{f^{\mathrm{new}}_e}$
  9. Step 2: Generating Training Corpora
  10. KUP's Continued Pre-Training Setup
  11. KUP Dataset Analysis
  12. KUP includes richer knowledge update phenomenon than prior benchmarks beyond simple entity substitutions
  13. Analysis of phenomenon
  14. Memory Conditioned Training (MCT)
  15. Motivation
  16. ...and 37 more sections

Figures (3)

  • Figure 1: Example of LLM that is continued pre-trained on updated knowledge surfacing updates in the direct probing but failing under indirect probing settings
  • Figure 2: Our KUP data curation pipeline. We omit details about verification of $f^{\mathrm{old}}_e$ (in test models' parametric knowledge?) and $f^{\mathrm{new}}_e$ (contradictory to models' parametric knowledge?). The final training dataset comprises of fictitious evidence documents and real axillary documents for our entity set.
  • Figure 3: Illustration of Memory Conditioned Training