RecEvolve: Knowledge-Driven Recommender Evolution¶
Core mechanism evidence. Paper and local evaluations are separate.
Paper information¶
| Field | Value |
|---|---|
| Paper | RecSys 2026 |
| Affiliations | |
| First public release | 2026-07-20 (arXiv v1) |
| Authors' code | No public author code found as of 2026-09-06. |
| Adapter / method | recevolve |
| Local code | src/auto_research/reproductions/recevolve/ |
Original paper¶
Background and contribution¶
Propose variants from experiment knowledge and select champions on isolated validation. Version state for rollback and reject reward hacking such as extreme temperature, carrying forward auditable knowledge.
flowchart LR
K[Experiment knowledge] --> P[Proposal] --> G[Critic gate] --> V[Isolated validation] --> C[Promote or roll back] --> KKey figure from the paper¶
Original RecEvolve figure. From the paper PDF; copyright belongs to the authors. Not a local result.
Core mechanism¶
Compare candidate \(c_t\) with champion \(c^*\) using independent validation utility \(J\) only after the critic gate. If \(J(c_t)>J(c^*)\), atomically set \(c^*\leftarrow c_t\); otherwise roll back. Reject hacking via \(g(c_t)\).
Paper results¶
41 autonomous experiments; offline NDCG@50 0.4796→0.5751 (+19.9%). Live satisfaction +3.77%, unique content +7.44%, cold-start discovery time −16.50%.
Local reproduction¶
Fixed MovieLens 100K candidate space, actual research control flow. NDCG@10 0.05401→0.05431 (+0.56%) after validation selection. Three-seed metrics.
Reproduction limits¶
Not Google's private two-tower metrics, logs, TPUs, coding-agent infrastructure, or live attribution. Adapter is a recommendation Evolve controller operator.