EvoSkillRec: Skill-Genome Evolution for Recommender Architecture Discovery¶
L1 mechanism diagnostic. Typed genomes, validation, promotion, and reuse execute. No external LLM invents arbitrary code and no large-scale search gain is claimed.
Paper information¶
| Field | Value |
|---|---|
| Paper | arXiv 2609.34552 |
| Affiliation | City University of Hong Kong |
| First public release | 2026-09-28 (arXiv v1) |
| Authors' code | EvoSkill-Rec |
| Adapter | evoskillrec |
| Local code | evoskillrec |
Original paper¶
Background and contribution¶
Split a recommender into executable skills with typed inputs and outputs. The controller composes within constraints; only validated innovations advance to the reusable library.
flowchart LR
I[Public inputs or state] --> M[Typed skill genome] --> A[Auditable intermediates] --> O[Outputs or state update]Key figure from the paper¶
Original components and connections from the paper; copyright belongs to the authors.
Core equation and evidence¶
A genome \(g=(s_1,\ldots,s_n)\) is executable only if adjacent types satisfy \(O(s_i)=I(s_{i+1})\). Promote to the reusable library only when \(V(g)>V(g_0)+\epsilon\). The local adapter runs type checking, composition, validation, promotion, and reuse. Paper benchmarks, speed, and learning curves are separate from the local three-seed mini-suite and imply no online gains here.
Local reproduction¶
Compare a fixed genome without skill transformations to typed composition with validation-gated promotion. Three-seed metrics are diagnostic_only=true, not relative capability scores or rankings.
Reproduction limits¶
The controller executes the constrained workflow, not arbitrary LLM-generated code or the paper's large-scale search.