A study by researchers affiliated with Nvidia’s AI Technology Center shows that a small amount of misleading passages in a long prompt can cause a sharp decline in answer accuracy. When these hard distractors made up 10% of a 128,000-token context, attention to the passage containing the correct answer fell by 76%. Researchers tested Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct on question-answering benchmarks using both single-hop and multi-hop questions. Masking a few attention heads responsible for processing irrelevant passages restored much of the lost accuracy. The study also finds that downstream document filtering offers limited benefits and that retrieval precision upstream should be the priority.
Source: Read the original article
Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
