Complex Analogy Learning: Training Language Models to Solve Intricate Reasoning Puzzles through Analogies
Large language models (LLMs), such as GPT-3, PaLM, and potential future models like GPT-4, have demonstrated impressive natural language processing abilities. However, these models lack systematic reasoning abilities, unable to carry out the step-by-step deduction that allows humans to solve challenging problems.
Recent research from Yale University and the Chinese Academy of Science proposes an innovative technique called "Thought Propagation" to enhance LLMs' reasoning through analogical thinking. This technique, which involves three stages - proposing analogous problems, solving them, and aggregating the solutions - aims to boost the performance of LLMs on complex challenges requiring global optima or long-term planning.
The Thought Propagation (TP) model works by prompting LLMs to explore "analogous" problems related to the input before solving it. This approach allows LLMs to generalize novel ideas or innovations across different stages of a multi-step process, enhancing the transferability of innovations and enabling them to apply localized ideas more broadly.
One key innovation of Thought Propagation is the use of the Innovation Scatter Model. This model allows LLMs to generalize ideas across different stages, identifying the core innovation, abstracting it from specific contexts, determining its applicability, and systematically applying it to other similar stages. This method increases the reliability and robustness of LLMs' outputs.
Another innovation is Chain-of-Thought Prompting, which encourages the model to generate intermediate reasoning steps. This technique improves the model's performance on tasks requiring arithmetic and symbolic reasoning, allowing it to "think aloud" and explore multiple reasoning paths, leading to more accurate and reliable solutions.
Techniques like feedback-guided improvement and self-training paradigms further enhance LLMs' reasoning capabilities. These methods use feedback to refine answers iteratively, reducing dependency on human labels by generating and selecting the best traces for retraining.
The benefits of Thought Propagation are significant. It helps LLMs generalize beyond pretrained patterns, apply innovations across different contexts, and adapt more effectively to new tasks and domains, reducing the need for extensive retraining. Additionally, it increases the robustness of LLMs' outputs by exploring multiple reasoning paths.
Thought Propagation has demonstrated its power, exceeding sophisticated methods like chain-of-thought prompting. It has significantly boosted performance across different LLMs in shortest path tasks, story writing, and long-term planning for LLM agents. However, efficiently generating useful analogous problems is non-trivial, and chaining long analogical reasoning paths can become unwieldy.
With further development, induced analogical thinking could make LLMs' reasoning far more robust, demonstrating paths towards more human-like deduction in large language models. Solutions to these analogous problems can provide insights to either directly solve the input or extract useful plans to follow.
In conclusion, the Thought Propagation technique holds significant promise in overcoming the inherent limitations of large language models' logic capabilities, enhancing their reasoning capabilities beyond the limitations of their training data.
Artificial Intelligence (AI), in the form of Thought Propagation, is a technique that aims to integrate analogical thinking into large language models (LLMs), such as GPT-3, improving their reasoning and performing tasks requiring global optima or long-term planning. The Thought Propagation model enhances the transferability of innovations in LLMs by prompting them to explore analogous problems, generalize novel ideas, and apply localized ideas more broadly.
