Mind Evolution: How Genetic Search Helps LLMs Achieve Deeper Reasoning on Complex Planning Tasks

Forum topic · ✨步子哥 · 2025-12-28

Summary

This article explains the 'Mind Evolution' method (arXiv:2501.09891), a genetic search strategy that lets large language models (LLMs) solve complex planning problems by investing more compute at inference time. Instead of operating in a formal space, Mind Evolution evolves populations of natural-language candidate solutions using LLM-driven initialization, recombination, island-model migration, and reset, guided by programmatic evaluators that return both scores and textual feedback. On three challenging benchmarks, Gemini 1.5 Flash jumps from 5.6% to 95.6% on TravelPlanner, from 20.8% to 85.0% on Natural Plan meeting planning, and reaches 96.2% on Trip Planning. A two-stage variant with Gemini Pro pushes TravelPlanner to 100% and meeting planning to 98.4%. The approach outperforms Best-of-N sampling and sequential self-revision, scales with depth rather than breadth, and is easily parallelizable, demonstrating that evolutionary search over natural language unlocks deeper reasoning without formalizing the underlying problem.

Key points

Mind Evolution (arXiv:2501.09891) is a genetic search method that evolves natural-language candidate plans. It uses LLM prompts for initialization, parent selection (Boltzmann tournament), crossover/mutation via a Review–Critic–Critic (RCC) dialog, island-model migration, and periodic LLM-guided island resets, all guided by a programmatic evaluator.
The method needs only an evaluator, not a formalized problem. Programs score solutions and return textual feedback, which an ablation (Table 4) shows is critical to performance.
Benchmark gains are dramatic on hard planning tasks.
*TravelPlanner*: Gemini 1.5 Flash rises from 5.6% (1-Pass) and 55.6% (Best-of-N, 800 candidates) to 95.6% (validation) and 95.2% (test); a two-stage variant with Gemini Pro reaches 100% (val) and 99.9% (test).
*Natural Plan Trip Planning*: Flash reaches 96.2% / 94.1%, vs. Best-of-N 77.2% and Sequential-Revision+ 74.4%; two-stage 100% / 99.6%.
*Natural Plan Meeting Planning*: Flash reaches 85.0% / 83.8%, vs. Best-of-N 69.4%; two-stage 98.4% / 98.2%.
Depth beats breadth. Figure 6 shows success rate climbing steadily across generations, and Figures 7–9 show Mind Evolution is more compute-efficient than Best-of-N at matched candidate counts. Island models contribute over 10% (Table 5), and more generations help more than more candidates per generation.
Robustness across difficulty. As trip length grows (3, 5, 7 days) or city/participant count grows, baselines degrade while Mind Evolution stays stable (Figures 3–5).
Generality. On GPT-4o-mini (Table 7), Mind Evolution improves 1-Pass performance by 100%+. On the harder creative task *StegPoet* (hiding messages in essays/stories/poems), Gemini Pro reaches 87%, and Mind Evolution improves over baselines (Table 6), with evaluation via parsing and Levenshtein distance.
Limitations and outlook. The method depends on a programmatic evaluator; future work could swap in LLM-based judges. The paper positions Mind Evolution among prior LLM+evolution work (e.g., FunSearch, AlphaCode, prompt optimization) but argues it is unique in evolving directly in natural language, avoiding the need to formalize the underlying problem.

How it works (compact summary)

1. Sample N_convs initial plans per island across N_island islands; refine each for N_seq turns via RCC (critic + author dialog) using the evaluator's score and feedback. 2. Each generation: select 0–N_parent parents via Boltzmann tournament; recombine using RCC; parse and score offspring. 3. Every generation, migrate the top N_emigrate plans from island *i* to island *i+1*. 4. Every N_reset_interval generations, reset the N_reset lowest-scoring islands by prompting the LLM to pick N_top diverse elites from N_candidate global top solutions.

Default hyperparameters (from the paper)

| Parameter | Default | Description | |---|---|---| | N_gens | 10 | Max generations | | N_island | 4 | Independent populations | | N_convs | 5 | Conversations per island | | N_seq | 4 | Turns per conversation | | N_reset interval | 3 | Generations between resets | | N_reset | 2 | Islands reset each time | | N_top | 5 | Elite parents transferred on reset | | N_candidate | 15 | Candidates considered for reset | | N_parent | 5 | Max parents per conversation | | P(no parents) | 1/6 | Chance of a no-parent conversation | | N_emigrate | 5 | Plans emigrating per island per generation | | N_retries | 5 | Generation retries per turn |

Maximum candidate count per run ≈ N_gens × N_island × N_convs × N_seq = 800.

Takeaway

Mind Evolution reframes LLM inference-time scaling as evolutionary search in natural language: cheap, parallelizable, and broadly applicable whenever a programmatic checker (or future LLM judge) can score outputs, including travel planning, meeting scheduling, trip planning, and creative steganography.