Reproduce “Reversal of Thought: Enhancing Large Language Models with Preference-Guided Reverse Reasoning Warm-up” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiahao Yuan, Dehui Du, Hao Zhang, Zixiang Di, Usman Naseem
The Reversal of Thought (RoT) framework improves Large Language Models' logical reasoning by eliciting and leveraging their intrinsic "cognitive preferences" through a reverse reasoning warm-up and an adaptive Cognitive Preference Manager. This approach consistently outperformed state-of-the-art baselines in reasoning accuracy, achieving up to a 17.15% gain on tasks like Game of 24, while maintaining competitive efficiency.
Key insights
- Large Language Models exhibit inherent "cognitive preferences" and "cognitive patterns" (shaped by pre-training and RLHF) that can be effectively elicited and leveraged for enhanced logical reasoning.
- A "reverse reasoning" approach allows LLMs to autonomously generate optimal, task-specific prompts by articulating their own preferred logical structures and pseudocode.
- Dynamically assessing an LLM's "knowledge boundary" for a given task and adaptively merging prompt components significantly improves flexibility and performance across diverse reasoning challenges.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.