LLMs Fail to Adapt When Prompt Information Is Shared Incrementally

Research from Microsoft and Salesforce shows that popular AI models experience a 39 percent drop in accuracy when details are delivered across multiple chat turns instead of a single message. The study found that chatbots anchor heavily on early assumptions and struggle to revise their answers even when corrected later.