Research map / Architectures and efficiency / Neural architecture search

Self-EvolveRec: Self-Evolving Recommender Systems with LLM-based Directional Feedback

NeurIPS 2026 · Architectures and efficiency › Neural architecture search · approach: Bi-level differentiable NAS · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Self-EvolveRec, an LLM-driven code-evolution framework for recommender pipelines that adds a persona-based User Simulator (qualitative critiques) and a Model Diagnosis Tool (quantitative probes) co-evolved with the codebase, outperforming AlphaEvolve, DeepEvolve and NAS baselines on CDs, Electronics, Office and MovieLens.

Problem

Automated recommender system design via NAS is limited to fixed operator spaces, and LLM-driven code evolution relies on scalar metrics (NDCG, HR) that give no diagnostic or directional guidance on why models fail.

Approach

Establish a directional feedback loop combining LLM user-persona critiques with quantitative diagnostic probes, and co-evolve the diagnosis tool with the codebase so its metrics track new components and translate simulator critiques into measurable signals.

Key results

• On CDs with SASRec seed, Self-EvolveRec reaches NDCG@5 0.3865 / HR@5 0.5274 versus DeepEvolve 0.3610 / 0.4870 and AlphaEvolve 0.3528 / 0.4623.
• Under agentic evaluators (Agent4Rec, PUB), Self-EvolveRec achieves the best View/Satisfy/Depth across datasets and seeds, with gains over DeepEvolve.
• Peak performance reached at iteration 8 (CDs) and 11 (Electronics) from a random seed, versus 13 to 19 iterations for baselines.

Surprising findings

• Removing feedback-aware planning/retrieval yields comparable NDCG/HR but lowers codebase quality scores for creativity and explicitness.
• AlphaEvolve and DeepEvolve sometimes underperform their own seed recommenders, indicating scalar-only evolution can degrade performance.

Limitations

• Relies on A/B-style simulated evaluation rather than real-world A/B testing, which the authors leave for future work.
• Per-iteration cost is high (about 19 minutes per iteration on CDs) compared with AlphaEvolve.

Experimental setup

Datasets: Amazon CDs Amazon Electronics Amazon Office MovieLens

Benchmarks: Amazon CDs Amazon Electronics Amazon Office MovieLens

Metrics: NDCG@5 HR@5 View Satisfy Depth Creativity Explicitness Insight Personalization

Baselines: AutoFIS NASRec AutoLossGen DNS-Rec AlphaEvolve DeepEvolve

Builds on: AlphaEvolve DeepEvolve OpenEvolve Agent4Rec NASRec

Closest papers by meaning

• AgentRecBench: Benchmarking LLM Agent-based Personalized Recommender Systems (NeurIPS 2025 · similarity 0.86)
• Re2LLM: Reflective Reinforcement Large Language Model for Session-based Recommendation (AAAI 2025 · similarity 0.85)
• Building Higher-Order Abstractions from the Components of Recommender Systems (AAAI 2024 · similarity 0.85)
• Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.85)
• Objective Shaping with Hard Negatives: Windowed Partial AUC Optimization for RL-based LLM Recommenders (NeurIPS 2026 · similarity 0.84)
• Influential Exemplar Replay for Incremental Learning in Recommender Systems (AAAI 2024 · similarity 0.84)