Reproduce “Evaluating the Sensitivity of LLMs to Prior Context” on OpenResearch

About this paper

Robert Hankache, Kingsley Nketia Acheampong, Liang Song, Marek Brynda, Raad Khraishi, Greig A. Cowan

This research systematically evaluates the sensitivity of large language models to prior context in multi-turn interactions, demonstrating significant performance degradation, with some models experiencing accuracy drops of up to 73% when context is present. The study also reveals that strategically repeating task instructions can recover up to 85% of this performance loss.

Key insights

• Single-turn benchmarks are insufficient for evaluating LLM performance in multi-turn interactive scenarios, as models proficient in single-turn tasks can fail drastically in complex, context-rich settings.
• LLMs struggle with maintaining task instructions and coherent behavior over long, dynamic contexts, experiencing 'task information dilution' that leads to performance degradation and output formatting errors.
• Strategic placement of task descriptions, specifically repeating instructions immediately before the target query, is a highly effective mitigation strategy for context-induced performance drops.

Summary by alphaXiv.