Mind Evolution: How Large Language Models Evolve from Shallow Thinking to Deep Planning

Forum topic · ✨步子哥 · 2025-12-28

Summary

Mind Evolution is a genetic search strategy that enables large language models (LLMs) to spend more inference-time computation solving natural language planning problems. Instead of formalizing the underlying reasoning task, the method evolves a population of candidate solutions directly in natural language, using an LLM for initialization, recombination, and island-based resets, guided by a programmatic fitness evaluator that scores candidates and produces textual feedback. Combining divergent exploration with convergent selection, it outperforms Best-of-N sampling and sequential revision baselines. On TravelPlanner, Gemini 1.5 Flash improved from 5.6% (1-Pass) to 95.6% success; on Natural Plan Trip Planning it reached 96.2%, and on Meeting Planning 85.0%, with a two-stage approach using Gemini Pro pushing results to nearly 100%—all without a formal solver. The paper also introduces StegPoet, a steganography benchmark demonstrating the method's applicability beyond planning. This post presents a full Chinese-language walkthrough of the arXiv:2501.09891 paper, including benchmark results, the island-model hyperparameter table, ablation studies showing the importance of critique steps and textual feedback, and cost-efficiency analysis versus API spend.

Introduction: An Evolutionary Adventure in the Language World

Imagine standing at the edge of a vast linguistic ocean, holding a fuzzy map, tasked with finding treasure hidden in complex problems. This is a fair picture of what large language models (LLMs) face with planning tasks. The paper reviewed here (arXiv:2501.09891, "Evolving Deeper LLM Thinking") proposes Mind Evolution, an innovative method that uses evolutionary search to let LLMs spend more compute at inference time, dramatically improving problem-solving.

Like Darwin observing variation across the Galápagos, Mind Evolution generates, recombines, and refines candidate solutions, simulating natural selection. Its core idea combines divergent thinking (freely exploring many ideas) with convergent thinking (evaluating and selecting the best)—a hallmark of intelligent human problem-solving. If you're unfamiliar with genetic algorithms, think of a family reunion: parents (candidate solutions) produce offspring (new solutions) through crossover and mutation, while the fitness function acts like a discerning elder ensuring only the best lineage continues.

Prior research explored chain-of-thought prompting, self-consistency, feedback-based sequential revision, and search guided by auxiliary verifiers. Best-of-N does broad search; tree search evaluates at step level. Mind Evolution differs: it is a genetic search over natural language, requiring only a solution evaluator—no formalization of the problem needed. Unlike Best-of-N (breadth only) or sequential revision (local fixes), it refines globally and parallelizes easily.

Benchmark Results: From Travel Planning to Meeting Scheduling

TravelPlanner expresses user preferences and constraints (lodging, food, transport, budget) in natural language. Baselines struggled: Gemini 1.5 Flash and o1-preview achieved only 5.6% and 11.7% success; even Best-of-N with 800 candidates reached just 55.6%. With comparable inference cost, Mind Evolution brought Flash to 95.6% on TravelPlanner validation and 85.0% on meeting planning; a two-stage approach (unsolved instances handed to Gemini Pro) pushed these to 100% and 98.4%—no formal solver required, comparable to prior GPT-4-based formalization-then-solve approaches (98.9%) but far more general.

On Natural Plan:

Trip Planning (320 validation / 1280 test): 96.2% / 94.1%; two-stage: 99.6% / 100%.
Meeting Planning (500/500): 85.0% / 83.8%; two-stage: 98.4% / 98.2%.

As problem complexity grows (more days, cities, or attendees), baselines degrade while Mind Evolution remains stable—and its advantage widens.

Default hyperparameters

| Parameter | Default Value | Description | |---|---|---| | N_gens | 10 | Maximum number of generations to search for a solution | | N_island | 4 | How many independent populations to evolve | | N_convs | 5 | Conversations per island | | N_seq | 4 | Turns per conversation | | N_reset_interval | 3 | How frequently to reset islands (in generations) | | N_reset | 2 | How many (lowest-scoring) islands to reset | | N_top | 5 | Starting parents transferred to islands on reset | | N_candidate | 15 | Candidate parents considered when resetting | | N_parent | 5 | Maximum parents per conversation | | P_r_no_parents | 1/6 | Probability of a conversation having no parents | | N_emigrate | 5 | Plans emigrated to the next island after each island | | N_retries | 5 | Generation attempts per turn before giving up |

The product N_gens × N_island × N_convs × N_seq gives the maximum candidate count (default 800).

Method: A Genetic Algorithm in Natural Language

Mind Evolution implements a language-based genetic algorithm:

1. Fitness evaluation: A programmatic evaluator parses solutions, returning scores plus textual feedback (constraint violations, event counts, format issues). Ablations show textual feedback is critical. 2. Population initialization: N_convs initial solutions are sampled; if N_seq > 1, they are refined through RCC—a critic–author dialogue where the critic analyzes inputs and feedback while the author proposes refined solutions. 3. Selection: Boltzmann tournament, sampling 0 to N_parent parents from softmax fitness. 4. Crossover + mutation: Performed as a single recombination step via RCC over parents. 5. Migration: The top N_emigrate plans are cyclically cloned to the next island. 6. Island reset: Every N_reset_interval generations, the N_reset lowest-scoring islands are rebuilt by using the LLM to select N_top diverse elites from N_top candidates.

Ablations confirm the critique step and textual feedback matter most; the island model and LLM-driven resets contribute significantly. Island models beat no-island setups, and more generations beat more candidates per generation.

Related Work

Earlier combinations of evolution with LLMs focused on program generation or numerical/combinatorial optimization, requiring formal spaces. Prompt-optimization methods like EvoAgent reached only 7.2% on TravelPlanner (with GPT-4), versus 95%+ here. Learned verifiers used in code generation and search tend to be noisy; Mind Evolution uses programmatic evaluators, though LLM-based evaluators could extend its reach.

Experiments

Experiments used Gemini 1.5 Flash by default, with Pro in the two-stage setup, against 1-Pass, Best-of-N (up to 800 candidates), and Sequential-Revision+ (10 threads, 80 turns) baselines. Metrics included success rate, LLM call counts, tokens, and API cost (October 2024 prices). Success rates climb steadily over generations, and Mind Evolution achieves better cost-efficiency curves than baselines. Results also replicate with GPT-4o-mini, where the improvement exceeds 100% relative to 1-Pass.

A new benchmark, StegPoet, encodes hidden messages into essays, stories, or poems—hard to formalize but programmatically checkable (message recovery scored by position plus Levenshtein distance). Gemini Pro reached 87%, demonstrating the method's generality beyond planning.

Limitations and Outlook

Mind Evolution depends on programmatic evaluators. Future work could substitute LLM-based evaluators. Overall, the method opens a new era of deeper LLM thinking—evolution granting models the adaptability to conquer complex puzzles.

---

References

1. Lee, K.-H., et al. "Evolving Deeper LLM Thinking." arXiv:2501.09891v1, 2025. 2. Wei, J., et al. "Chain-of-thought prompting elicits reasoning in large language models." NeurIPS, 2022. 3. Romera-Paredes, B., et al. "Mathematical discoveries from program search with large language models." Nature, 2024. 4. Xie, J., et al. "Travelplanner: A benchmark for real-world planning with language agents." arXiv:2402.01622, 2024. 5. Zheng, H. S., et al. "NATURAL PLAN: Benchmarking LLMs on natural language planning." arXiv:2406.04520, 2024.