EvoSkillRec: Skill-Genome Evolution for Recommender Architecture Discovery
Xiaopeng Li, Kuo Cai, Bo Chen, Wenlin Zhang, Mengyang Ma, Yingyi Zhang, Zichuan Fu, Yu Yang, Qidong Liu, Yiyu Wang, Ruiming Tang, Wenwu Ou, Jiang Wu, Zhanbo Xu, Xiangyu Zhao
cs.IR
2026-09-28
CityU and Kuaishou evolve recommender architectures as typed skill genomes. Test AUC beats the best manual baseline on six CTR/MTL/MDL sets, max +0.0085.
The gains that actually moved recommender systems often came from architectural inductive bias: task priors written into the network, not extra layers. CTR models cross sparse categorical fields; DeepFM and DCN put that into the graph. Target-aware rankers let each candidate light up a different slice of user history, which is what DIN and DIEN do with attention. Multi-task models share representation without washing out task-specific signal, which is the job of MMoE and PLE routing. Those modules are hand-built. Move to a new task, feature mix, or business setting, and the knowledge inside one family almost never transfers as a reusable part.
NAS (neural architecture search) can pick operators, wiring, or losses inside a grammar a human already wrote. NASRec and AutoLossGen live in that box. They rarely invent a module that sits outside it. LLM code evolution (FunSearch, AlphaEvolve, Self-EvolveRec) opens the space. Unconstrained edits then flood the queue with invalid or low-value programs, and a successful idea stays trapped in a one-off diff. The next run starts from a codebase and a transcript again. A self-evolving recommender should keep what worked.
EvoSkillRec, from City University of Hong Kong, Kuaishou, and Xi'an Jiaotong University, compiles existing recommenders into atomic executable skills and represents a model as a typed skill genome. The internship work was done at Kuaishou.
A skill is not a prompt. It ships PyTorch code, named tensor I/O, an inductive-bias tag, task applicability, failure signatures, and a validation suite, and it talks through a blackboard tensor context called SkillContext. Two skills compose when the producer's output keys cover the consumer's inputs. The card's goodfor / badfor notes are retrieval copy.
The seed library is compiled from Torch-RecHub and UniRank implementations of DeepFM, Wide & Deep, DCN, DIN, DIEN, SASRec, MMoE, PLE, and related models. Isolate modules, write cards, rewrite them as BaseSkill files, then validate. The hard check rebuilds each seed model from its extracted skills and asks for end-to-end match. Tier-1 holds 94 skills in 11 categories and stays fixed within a run. Tier-2 admits a code-space invention only after the same pipeline and a measured lift inside a host genome.
Each generation proposes 8 candidates by default, 6 from skill space and 2 from code space.
Skill space treats the model as a genome graph and applies add, replace, hybridize, and specialize. Retrieval hard-filters on task, interface, and loadable code, then scores task fit, tensor keys, diagnosis text, matching inductive bias, and historical gain.
Code space writes a new module when the library cannot address the diagnosed failure. The LLM never sees the training data. It gets a design brief: parent genome, diagnosis, recent memory, exposed tensor keys, and constraints. A planner emits structural sketches; a synthesizer turns them into nn.Module code. Sketches must be macro-level (replace a node, insert, branch into fusion, or replace fusion) and may read only tensors the genome already exposes. Static analysis drops file, network, subprocess, or dynamic-eval capabilities. Only modules with valid forward and backward passes go to training. Skills carried by survivors sediment into Tier-2 and become ordinary retrieval targets later, so they no longer spend code-space budget.
If the best validated score plateaus for k generations, the budget shifts toward code space, then relaxes once a new skill is promoted. An AutoResearch controller runs diagnosis, proposal, training, survivor selection, and sedimentation as one loop.
Code-space proposals in the main tables come from GPT-5.5. Selection uses validation; reported numbers are test AUC.
| Task | Best manual | NASRec | OpenEvolve | EvoSkillRec | vs best baseline |
| CTR-MovieLens | 0.7487 Wide & Deep | 0.7361 | 0.7521 | 0.7522 | +0.0001 |
| CTR-Amazon Books | 0.6587 Wide & Deep | 0.6487 | 0.6654 | 0.6708 | +0.0054 |
| CTR-Amazon Beauty | 0.6666 AutoInt | 0.6632 | 0.6760 | 0.6845 | +0.0085 |
| MTL-Census | 0.9726 AITM | 0.9711 | 0.9722 | 0.9748 | +0.0022 |
| MDL-Amazon | 0.7012 SAR-Net | 0.6870 | 0.6903 | 0.7033 | +0.0021 |
| MDL-MovieLens | 0.7924 SAR-Net | 0.7619 | 0.7730 | 0.7932 | +0.0008 |
NASRec sits below the best manual model on all six sets. OpenEvolve beats the manual designs on the three CTR sets and falls behind on MTL and MDL, by 0.0194 AUC on MDL-MovieLens. EvoSkillRec is the only method in the comparison that beats the strongest hand-designed baseline on every set. The MovieLens CTR edge over OpenEvolve is +0.0001, a tie for any practical purpose.
Ablating a space: the full model still leads all six, but the margin over the stronger single-space variant is often +0.0002 (MovieLens CTR) or +0.0005 (Beauty). Code space is the stronger single-space run on CTR; skill space leads on MTL-Census and MDL-Amazon; MDL-MovieLens is 0.7924 vs 0.7926. Promoted code skills range from 16 on Beauty to 45 on each multi-domain set. Skill-space survivor yield on the two MDL sets is 3.1% and 6.2%, because many recombinations collapse to already-evaluated architectures and are dropped before training.
On industrial QK-Video, evolution starts from RankMixer and runs 30 rounds under a joint AUC and training-MFU objective (achieved training FLOP/s over the GPU peak). A candidate is eligible only if training MFU stays at least 95% of the seed and validation AUC stays within 10^{-3} of the incumbent. Of five Pareto points, only P1 from round 19 keeps both gains on test: 0.93598 vs 0.93580 AUC, 1.477% vs 1.433% MFU. The raw validation-AUC max is 0.9301 and the raw MFU max is 1.605%. All five front models scale RankMixer from L=3, E=2 to L=4, E=3 and insert a Hadamard cross mixer after DIN attention. P1 adds a low-rank task router at the tower and posts the lowest MFU of the five.
Swap the LLM: Claude Opus 4.8 leads CTR-Books at 0.6738 test AUC. GLM-5.2 leads MTL at 0.9739 and MDL at 0.7935, essentially tied with the GPT-5.5 run at 0.7932. DeepSeek-V4-Pro is weaker on all three, with a lower rate of valid proposals.
A growing, typed library is what NAS never had and what raw LLM diffs throw away. A team that already owns CTR, multi-task, and multi-domain modules can let a validated invention re-enter search under the same interface checks.
Most of the AUC table is incremental. The largest public CTR gain is +0.0085 on Beauty. MovieLens and MDL-MovieLens sit near run noise. The piece worth copying is the promotion-and-reuse loop.
The main numbers are tied to GPT-5.5. RQ5 shows a different backbone discovers different architectures. The framework and the model that drives it do not separate.
The paper puts the statistics problem first. Some reported edges are 10^{-4} AUC, and every cell in Table 1 is a single evolution run. Multi-seed repeats were too expensive because each cell is a full search, not a single fit. Small orderings are observations.
Each code-space proposal costs two LLM calls plus a full train, and many die before training. The reported experiments cost about $7,000 across backbones and tasks. Those figures describe EvoSkillRec plus the model that proposes code.
Cross-task transfer is untested; each library grows inside one task family. The only industrial contact is QK-Video, and the library still cold-starts from public academic architectures. P1's test AUC gain is 1.7×10^{-4}, the same order as the tiny public-benchmark edges.
Replacing an experienced ranking-architecture engineer is not in reach. Backbones differ sharply in how often they emit viable architectures, and budget, promotion, and evaluation still need a human in the loop.