ShinkaEvolve: Towards Open-Ended and Sample-Efficient Program Evolution

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

ICLR 2026; ShinkaEvolve: LLM-driven scientific discovery via open-ended program evolution

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

ShinkaEvolve is an LLM-driven evolutionary program synthesis framework that achieves state-of-the-art sample efficiency by combining adaptive parent sampling, code novelty rejection-sampling, and bandit-based LLM ensemble selection. It surpasses AlphaEvolve's circle-packing solution in only 150 evaluations (versus thousands) and generalizes to agentic scaffold design, competitive programming, and MoE loss discovery.

Problem

Existing LLM-guided evolutionary code frameworks (AlphaEvolve, OpenEvolve, FunSearch) suffer from severe sample inefficiency, requiring thousands of program evaluations due to naive exploration strategies that fail to exploit accumulated evolutionary knowledge. This makes them computationally prohibitive for broad scientific discovery.

Method

ShinkaEvolve maintains an island-based archive of programs with fitness scores and iterates three phases:

1. Adaptive parent and inspiration sampling. Programs are selected from island subpopulations using either power-law sampling \(p_i \propto r_i^{-\alpha}\) (rank-based, \(\alpha\) controls exploitation) or weighted sampling \(p_i \propto w_i / \sum_j w_j\) where \(w_i = \sigma(\lambda(F(P_i) - \alpha_0)) \cdot \frac{1}{1+N(P_i)}\), combining fitness sigmoid and inverse offspring count. The latter is the default and balances performance with novelty.

2. Novelty rejection-sampling for mutations. After an LLM proposes a diff-edit, full rewrite, or crossover mutation, the mutable code is embedded and cosine similarity is computed against the island archive. If \(\max_j \text{sim}(e_i, e_j) > \eta\) (default \(\eta=0.95\)), a second LLM judges whether the proposal is meaningfully different; otherwise it is rejected and resampled.

3. Bandit-based LLM ensemble selection (UCB1). Each LLM in the ensemble (GPT, Gemini, Claude, DeepSeek) is scored by relative improvement \(F(P_i)^u = \exp(\max(F(P_i) - F(P_i)^b, 0)) - 1\), where \(F(P_i)^b\) is the max of the parent and the initial program score. The exponential term rewards high-risk, high-reward mutations; UCB1 statistics are maintained per model and updated each generation.

A meta-scratchpad periodically (every \(T\) generations) synthesizes insights from high-performing solutions and appends design principles to mutation prompts for knowledge diffusion across generations.

Key Contributions

• Adaptive parent sampling combining fitness sigmoid and inverse offspring frequency for exploration-exploitation balance.
• Code novelty rejection-sampling via embedding cosine similarity + LLM novelty judge, preventing near-duplicate proposals.
• UCB1 bandit LLM ensemble selection using relative improvement rewards to dynamically allocate mutation budget to the best-performing model.
• Meta-scratchpad for online synthesis and diffusion of discovered optimization strategies.
• State-of-the-art circle-packing solution surpassing AlphaEvolve in ~150 evaluations.
• Evolved agentic scaffold for AIME reasoning featuring expert persona diversity, peer review, and synthesis stages.

Results

• Circle packing (26 circles in unit square): ShinkaEvolve finds a new SoTA solution (combined fitness 2.6+) in 150 evaluations vs. AlphaEvolve's thousands; cumulative API cost under ~$10.
• AIME 2024: Evolved scaffold achieves 34.4% accuracy vs. 23.0% base agent and 24.4% majority@5; uses only 7 LLM calls per problem (budget = 10). On AIME 2025, ShinkaEvolve reaches 25.6% vs. 20.0% base and 24.4% majority@5. Transfers to gpt-4.1-mini (65.6%), gpt-4.1 (80.0%), and o4-mini (94.4%), compared to base agents at 60.0%, 65.6%, and 88.9% respectively.
• ALE-Bench LITE (10 competitive programming tasks): Mean score improves from 1879.3 (ALE-Agent) to 1932.1 (ShinkaEvolve public Top-1, ~2.3% gain); on ahc039, solution moved from 5th to 2nd place on AtCoder leaderboard.
• MoE load balancing loss: ShinkaEvolve discovers novel loss functions (results for this domain are cut off in the provided text).

Limitations

• Circle-packing experiments use only 150 samples on one canonical benchmark; generalization to harder or higher-dimensional problems is untested.
• ALE-Bench solutions tend to stay close to the initialization (ALE-Agent), indicating susceptibility to overfitting to the seed solution.
• AIME 2023 shows smaller gains, likely due to training data contamination in the base LLM; the framework cannot correct for memorization.
• The meta-scratchpad and multi-objective feedback add implementation complexity; hyperparameter sensitivity (η, T, α, λ) is not fully characterized.
• MoE loss discovery results are not fully reported in the provided text, limiting assessment of that domain.
• Relies on external API providers (GPT, Gemini, Claude, DeepSeek), introducing cost and rate-limit constraints at scale.

Relevance to Agentic AI / LLM Agents

ShinkaEvolve directly addresses the automated design of agentic scaffolds — its AIME application evolves multi-agent architectures (expert generation → peer review → synthesis) rather than hand-crafting them, which is a core open problem in the field. The framework instantiates LLMs as meta-level operators that iteratively self-improve agent designs, connecting to lines of work on automated agent search (ADAS, EvoPrompting) and program synthesis. The bandit-based LLM ensemble selection is a principled approach to multi-model orchestration, relevant to any system that must route tasks across heterogeneous LLMs. Its sample efficiency gains are practically significant: reducing costs from thousands to ~150 evaluations makes evolutionary agent design feasible for resource-constrained research.