Singapore – September 11, 2026 -- Appier's AI Research team has found that large language model accuracy falls by 30% to 50% when the correct response is that no valid answer exists, a gap the company says undermines trust in enterprise Agentic AI deployments.
Testing of 28 LLMs exposes a critical blind spot in autonomous decision-making
In a paper titled "None of the Above, Less of the Right: Parallel Patterns between Humans and LLMs on Multi-Choice Questions Answering," Appier's research team introduced "None of the Above" (NA) options into multiple-choice tests to simulate scenarios where retrieval fails to surface a valid answer. Across 28 leading LLMs of varying sizes, models tended to select suboptimal or incorrect options rather than flag insufficient information, even when they possessed relevant knowledge. The team identified business ethics as a domain where this weakness is most pronounced, since it requires holistic evaluation of multiple plausible options rather than a single verifiable answer.
Targeted training lifts accuracy by nearly 30 percentage points
Appier applied two training methods to address the gap: Supervised Fine-Tuning (SFT), which trains models on correct examples, and Direct Preference Optimization (DPO), which exposes models to both correct and incorrect responses. DPO improved accuracy in identifying "none of the above" scenarios by nearly 30 percentage points. The researchers caution that the NA approach only works when answer options are clearly defined and mutually independent, meaning enterprises must build task-specific checkpoints that confirm information sufficiency before an AI agent acts, escalating to further search or human review when gaps remain.
Reasoning language shifts logic, safety judgments and cultural accuracy
A second paper, "Language Matters: How Do Multilingual Input and Reasoning Paths Affect Large Reasoning Models?", found that the language a model uses internally to reason can directly affect its logical output, safety judgments and cultural understanding. Using a "text prefilling" technique to force models to reason in a specified language, Appier's team found that for some models, the reasoning language diverged from the final response language in over 90% of cases, with models defaulting to high-resource languages such as English regardless of the input language.
English reasoning wins on math, local language wins on safety and culture
High-resource languages like English produced stronger results on mathematics and knowledge-based tasks. For tasks requiring cultural context, reasoning in the local language captured nuances more effectively, and safety testing showed local-language reasoning was better at identifying harmful or illegal queries. Appier frames this as the basis for future "reasoning-language routing," where Agentic AI systems dynamically select a reasoning language based on task type, market and cultural context while still responding in the user's preferred language.
Appier CEO ties the findings to enterprise adoption confidence
"These two papers redefine the standard for evaluating AI. As AI moves from answering questions to making autonomous decisions, measuring whether a model produces the correct answer is no longer enough," said Chih Han Yu, CEO and Co-founder of Appier. Appier said its AI Research team will continue applying these findings across its three product lines — Ad Cloud, Personalization Cloud and Data Cloud — as it develops Agentic AI capable of judging when information is insufficient and which reasoning approach fits a given market or task.