EvoLattice #55
Description
Hi Robert! we briefly met at GECCO. I’ve really enjoyed digging into ShinkaEvolve; it’s a very clean and thoughtfully designed framework.
I wanted to share my preprint research called EvoLattice that can naturally extend ShinkaEvolve, without changing its core evolution logic.
https://www.arxiv.org/abs/2512.13857
Instead of representing a candidate as a single program, EvoLattice represents it as a DAG where each node contains multiple alternatives. Every valid path in the graph is executable, so one “program” compactly represents a combinatorial family of programs as parents/elites.
Each alternative has explicit performance statistics (mean / std / best path score and age) based on various paths that they're involved in.
LLM makes its own decisions which alternatives to add or remove at which nodes, based on these statistics and observed internal diversity.
There is no quality-diversity ranking, embeddings-based retrieval, etc. LLM observes stats and diversity directly and reasons about what's meaningful to keep, extend (mutate, cross-over, etc) or forget during the course of the evolution, w.r.t. qualitative instructions/constraints.
Thus, exploration vs. exploitation is therefore can be controlled entirely by the LLM’s own reasoning, rather than defining external heuristics.
The compact representation is both LLM-friendly but also highly-computational efficient through subpath memoized execution at run-time.
Mutations are minor and much less risky due to being granular on alternative level; whilst a diff may break the execution of a whole program.
A deterministic repair step enforces acyclicity and consistency (only if that'd be desired); hence the LLM focuses on local, meaningful edits.
EvoLattice representation is applicable to AI agents as well, which is my main area of interest as I'm working on an AI agent framework firm.
I actually already obtained good results in AIME task in ShinkaEvolve as well; but saving those now probably for a follow-up publication later.
If this sounds interesting, I’d be happy to chat further. Thanks again for publishing/open-sourcing ShinkaEvolve — great work! Best, Kamer
P.S. I'm planning to attend EvoStar'26 this year as I've submitted many relevant work to EvoMusArt - let me know if you'll also be around.