OPEN LINK ↗
// 75d agoRESEARCH PAPER
ShinkaEvolve tops AlphaEvolve on sample efficiency
Sakana AI's open-source ShinkaEvolve framework uses LLMs as evolutionary mutation operators to automatically discover and improve scientific programs, reproducing AlphaEvolve's circle-packing result with orders of magnitude fewer evaluations. Accepted at ICLR 2026 and installable via PyPI, it adds a bandit-based LLM ensemble that dynamically picks the best model mid-run.
// ANALYSIS
ShinkaEvolve is the most credible open-source answer to AlphaEvolve yet — it not only matches the results but makes them accessible to researchers without Google-scale budgets.
- –Where AlphaEvolve needs a human to hand it the right problem, ShinkaEvolve co-evolves problems alongside solutions — a qualitatively different approach borrowed from POET and MAP-Elites
- –The bandit-based LLM ensemble (GPT-5, Sonnet 4.5, Gemini) solving the credit-assignment problem mid-run is a practically useful contribution beyond the evolutionary theory
- –Concrete wins are impressive: SOTA circle packing in ~150 evals, 2nd-place equivalent on AtCoder, a novel MoE loss function beating DeepSeek's approach, and helped win the 2025 ICFP Programming Contest
- –The honest caveat from author Robert Lange — "nothing interesting happens" when LLMs run fully autonomously — keeps expectations grounded: this is a co-pilot for researchers, not autonomous science
- –Apache 2.0 license and PyPI availability lower the barrier considerably compared to AlphaEvolve, which remains closed
// TAGS
shinka-evolvellmopen-sourceagentresearchbenchmark
DISCOVERED
75d ago
2026-03-15
PUBLISHED
76d ago
2026-03-14
RELEVANCE
8/ 10
AUTHOR
44th--Hokage
