Title: Inclusion-of-Thoughts: Stabilizing Preferences by Refining the Decision Space

Multiple-choice questions (MCQs) serve as a standard metric for assessing the capabilities of large language models (LLMs). Nevertheless, these models often struggle when faced with plausible but incorrect distractors. Such options can misdirect the model’s focus, leading to erratic fluctuations between accurate and erroneous responses. To address this issue, we introduce Inclusion-of-Thoughts (IoT), a progressive self-filtering approach aimed at reducing cognitive load—specifically, the instability in model preferences caused by distractors—and allowing the system to concentrate more efficiently on viable options.

IoT functions by reconstructing the MCQ to include only the plausible choices. This creates a controlled environment for evaluating comparative judgments and assessing the robustness of the model’s internal reasoning against perturbations. Furthermore, by making the filtering mechanism explicit, IoT improves the transparency and interpretability of the model’s decision-making process. Comprehensive empirical tests reveal that IoT significantly enhances chain-of-thought performance across various arithmetic, commonsense reasoning, and educational benchmarks, all while incurring negligible computational costs.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC