Diagnosing Retrieval Bias Under Multiple In-Context Knowledge Updates in Large Language Models

A research team led by Boyu Qiao has published a groundbreaking study diagnosing how large language models (LLMs) handle multiple sequential updates to the same factual information within a single context. Unlike previous research focusing on single updates, this work introduces the Dynamic Knowledge Instance (DKI) framework to simulate real-world scenarios where facts evolve over time—like a person's job title changing multiple times in a long document. The study tested models including GPT-4, Claude 3.5, and Llama 3, revealing a troubling pattern: as updates accumulate, models become increasingly biased toward retrieving the earliest version of a fact rather than the most recent.

The researchers discovered that while models maintain high accuracy (80%+) on initial facts, their ability to retrieve the latest updated information drops dramatically—sometimes to near 50% accuracy after multiple changes. Diagnostic analysis of attention patterns, hidden states, and output logits showed these signals become "flatter" and less discriminative as updates increase, making it difficult for models to distinguish between historical and current information. Even cognitively-inspired intervention strategies yielded only modest improvements, suggesting this is a fundamental architectural limitation rather than a simple training issue.

This research has significant implications for applications relying on accurate knowledge retrieval in long contexts, such as legal document analysis, medical record review, or financial reporting where facts frequently update. The findings suggest current LLM architectures may need fundamental redesigns to properly handle temporally evolving information, potentially requiring new attention mechanisms or memory systems that can better track information recency and relevance in dynamic contexts.