Appier Research Targets AI Reliability Through Reasoning and Limits

When an AI faces a query unsupported by its knowledge base, it often forces an answer rather than flagging a lack of information. Testing 28 leading models, Appier researchers found that accuracy plummeted by 30% to 50% when a 'none of the above' option was the correct response. By employing Direct Preference Optimization, the team improved model performance by 30 percentage points, demonstrating that targeted training can enforce honesty in AI responses. This shift is critical for high-stakes enterprise tasks, such as business ethics, where the ability to escalate to a human is safer than generating misleading output.

Beyond managing uncertainty, the research highlights that reasoning language dictates logical and safety outcomes. Models frequently default to English even when processing tasks in other languages, often missing local nuances or cultural risks. Appier’s 'text prefilling' technique allows systems to route reasoning to a specific language suited to the task, such as utilizing local languages for cultural context while maintaining English for technical or mathematical precision. CEO Chih Han Yu noted that these advancements are essential for transforming AI from a basic instruction-follower into a reliable decision-making system capable of navigating real-world complexities across global markets.