Mind Evolution: Evolving Large Language Model Inference #deepmind #ucsandiego #universityofalberta

arxiv: https://arxiv.org/pdf/2501.09891
"How can a large language model (LLM) be guided to think deeper about a complex problem and leverage inference time compute to improve its problem solving ability?"
This research paper introduces Mind Evolution, a novel evolutionary search strategy designed to enhance the problem-solving capabilities of large language models (LLMs). The method uses LLMs to generate, refine, and recombine potential solutions iteratively, guided by feedback from an evaluator. Experiments on established natural language planning benchmarks (TravelPlanner and Natural Plan) demonstrate significant performance improvements compared to existing methods, even surpassing approaches using formal solvers. Furthermore, the researchers introduce a new benchmark, StegPoet, to showcase the versatility of Mind Evolution in tackling less formalizable problems. The results highlight the potential of combining evolutionary search with LLMs for complex problem-solving.
Mind Evolution is an evolutionary search strategy that uses LLMs to solve complex natural language planning problems by scaling inference-time compute for stochastic exploration and iterative refinement. It is analogous to combining divergent thinking (free-flowing parallel idea exploration) with convergent thinking (idea evaluation and selection).

Mind Evolution leverages LLMs in multiple ways:

* It uses LLMs to **generate, recombine, and refine candidate solutions** based on feedback from an evaluator. The LLM is guided to think more deeply about the problem, exploring a diverse set of candidates and refining the most promising alternatives.
* The process is driven by **a tailored set of prompts**, which allows Mind Evolution to efficiently search for solutions to natural language planning tasks.
* **Refinement through Critical Conversation (RCC)** is a key component where the LLM is prompted to organize a conversation between a “critic” and an “author.” The critic analyzes candidate solutions, interprets evaluation feedback, and suggests improvements, while the author proposes refined solutions based on the critic’s analysis.
* During **island reset** events, the LLM can be prompted to select the top performers from a pool of candidates based on their fitness and diversity.

The effectiveness of Mind Evolution is demonstrated in its performance on three benchmark natural language planning domains: TravelPlanner, Trip Planning, and Meeting Planning.

* **Mind Evolution consistently outperforms baseline strategies** like Best-of-N and sequential revision, achieving higher success rates and requiring fewer candidate solutions to reach a specified level of performance.
* **The two-stage approach, which uses a more powerful LLM like Gemini 1.5 Pro to tackle unsolved problems, achieves even higher success rates.**
* **Ablation studies demonstrate the contribution of various components,** such as the critic step in RCC, textual feedback from evaluation functions, and island reset with LLM selection.

In conclusion, **Mind Evolution effectively leverages LLMs for complex problem-solving by combining the LLM's language understanding and generation capabilities with evolutionary search principles.** It achieves this through a carefully designed process involving prompt engineering, critical conversation, and iterative refinement guided by an evaluator. This approach enables LLMs to tackle complex planning tasks without relying on formal solvers, showcasing the potential of LLMs for solving challenging real-world problems.
Created with NotebookLM

Видео Mind Evolution: Evolving Large Language Model Inference #deepmind #ucsandiego #universityofalberta канала Srikanth Bhakthan