A paper from Microsoft has some rather worrying findings.Frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT-5.4) lose...

A paper from Microsoft has some rather worrying findings.Frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT-5.4) lose on average 25% of document content over twenty interactions. The all-model average is 50% with “Catastrophic corruption” occurring in 80%+ of model/domain combinations.All AI providers are pushing hands-off multi-step multi-agent work flows. This paper makes a strong case that that approach is currently not fit for purpose. #AI #AIAgent #LLMhttps://arxiv.org/abs/2604.15597

Read Original

Related