[Paper Review] DyKnow: Dynamically Verifying Time-Sensitive Factual Knowledge in LLMs

Seyed Mahed Mousavi, Simone Alghisi+1 more|arXiv (Cornell University)|Apr 10, 2024

Topic Modeling4 citations

TL;DR

This paper introduces DyKnow, a dynamic benchmark that uses real-time Wikidata queries to detect outdated time-sensitive factual knowledge in LLMs, revealing that most models, including GPT-J and Llama-2, retain outdated facts. It evaluates knowledge editing methods and Retrieval-Augmented Generation (RAG), finding RAG outperforms editing techniques in alignment but does not correct model weights, while editing methods show limited scalability and strong model dependency.

ABSTRACT

LLMs acquire knowledge from massive data snapshots collected at different timestamps. Their knowledge is then commonly evaluated using static benchmarks. However, factual knowledge is generally subject to time-sensitive changes, and static benchmarks cannot address those cases. We present an approach to dynamically evaluate the knowledge in LLMs and their time-sensitiveness against Wikidata, a publicly available up-to-date knowledge graph. We evaluate the time-sensitive knowledge in twenty-four private and open-source LLMs, as well as the effectiveness of four editing methods in updating the outdated facts. Our results show that 1) outdatedness is a critical problem across state-of-the-art LLMs; 2) LLMs output inconsistent answers when prompted with slight variations of the question prompt; and 3) the performance of the state-of-the-art knowledge editing algorithms is very limited, as they can not reduce the cases of outdatedness and output inconsistency.

Motivation & Objective

• To address the critical gap in detecting outdated time-sensitive factual knowledge in LLMs, particularly the 'what' and 'when' of outdated facts.
• To develop a dynamic benchmark that avoids obsolescence by retrieving up-to-date answers from Wikidata at evaluation time.
• To evaluate the effectiveness of knowledge editing techniques and RAG as surrogates for aligning LLMs with current factual knowledge.
• To estimate the recency of training data used in LLMs by analyzing their factual responses to time-sensitive queries.
• To enable community extension of the benchmark to new domains via shared code, datasets, and evaluation scripts.

Proposed method

• DyKnow dynamically retrieves current factual answers from Wikidata at evaluation time, using unique identifiers and properties to extract time-stamped responses for each fact.
• The benchmark targets time-sensitive facts in politics (47 countries), sports (28 athletes), and organizations (23 CEOs/secretaries), with lexicalized prompts for model evaluation.
• Outdated knowledge is identified by comparing model responses to real-time Wikidata answers, with performance measured via exact match and harmonic mean across domains.
• Knowledge editing methods—ROME, MEMIT, SERAC, IKE—are applied to correct outdated facts in four selected LLMs, with performance evaluated on the same benchmark.
• Retrieval-Augmented Generation (RAG) is used as a surrogate alignment method, retrieving current facts and prompting the model to generate updated responses.
• Scalability of editing methods is evaluated by increasing the number of edits, measuring performance degradation across models like GPT-2 XL and Llama-2 Chat.

Experimental results

Research questions

• RQ1 What proportion of time-sensitive factual knowledge in LLMs is outdated, and when was the data used for pre-training/fine-tuning collected?
• RQ2 How effective are state-of-the-art knowledge editing methods in correcting real-world, time-sensitive facts in LLMs?
• RQ3 Can RAG serve as a reliable surrogate for knowledge editing to align LLMs with current factual knowledge?
• RQ4 How does the performance of editing methods scale with increasing numbers of edits across different LLM architectures?
• RQ5 What are the limitations of current editing and retrieval-based alignment techniques in maintaining up-to-date LLM knowledge?

Key findings

• GPT-J, Llama-2, Falcon, and Bloom returned outdated responses for Cristiano Ronaldo’s club history, with GPT-J incorrectly stating 'Real Madrid' despite being released in 2021.
• Models like GPT-3 and Vicuna were initially correct at release but became outdated over time, indicating knowledge decay post-deployment.
• ROME and SERAC showed performance degradation as the number of edits increased, indicating limited scalability for large-scale knowledge updates.
• IKE achieved high performance on GPT-J and Mistral Instruct, suggesting in-context learning can be effective for certain models, but failed to correct all outdated facts.
• RAG achieved over 80% alignment performance on Mistral Instruct and outperformed editing methods on some models, though it does not correct the model’s internal knowledge.
• MEMIT demonstrated stable performance across different edit counts on Llama-2 Chat, while ROME and SERAC showed declining performance with increasing edits.