Inclusion-of-Thoughts: Mitigating Preference Instability via Purifying the Decision Space
A new self-filtering strategy helps LLMs ignore distracting wrong answers, boosting reasoning performance.
A team of researchers has introduced a novel method called Inclusion-of-Thoughts (IoT) designed to solve a critical weakness in large language models (LLMs): their susceptibility to 'preference instability' when faced with multiple-choice questions containing plausible but incorrect distractors. This instability causes models to oscillate between correct and incorrect answers, undermining reliable evaluation. IoT acts as a progressive self-filtering strategy, where the model systematically identifies and removes implausible options, effectively 'purifying' the decision space. By reconstructing the question with only the most viable choices, IoT reduces cognitive load and allows the model to focus its reasoning power more effectively.
Extensive testing demonstrates that IoT substantially boosts the performance of chain-of-thought reasoning across a range of challenging benchmarks, including arithmetic, commonsense reasoning, and educational tasks. Crucially, it achieves these gains with minimal added computational cost, making it a practical enhancement. Beyond raw performance, the method enhances transparency by explicitly documenting the filtering process, offering a clearer window into the model's internal decision-making pathway. This work, detailed in the arXiv preprint 'Inclusion-of-Thoughts: Mitigating Preference Instability via Purifying the Decision Space,' provides a simple yet powerful technique to make LLM reasoning more robust, stable, and interpretable in test settings riddled with tricky alternatives.
- IoT is a progressive self-filtering method that removes implausible distractors from multiple-choice questions, purifying the model's decision space.
- The technique substantially boosts chain-of-thought performance on arithmetic, commonsense, and educational benchmarks with minimal computational overhead.
- By documenting the filtering process, IoT also enhances the transparency and interpretability of the LLM's internal reasoning steps.
Why It Matters
Provides a simple, low-cost method to make LLM evaluations more reliable and reasoning more stable against deceptive answer choices.