MAREval: A Multi-Agent Framework for Evaluating Natural Language Recommendation Explanations

LLM-as-a-Judge, which leverages a large language model to evaluate natural language generated by another AI application or model, has become a standard, scalable approach for accelerating and extending costly human evaluation. Yet most work treats a judge as a static artifact, evaluating it once at construction or against a fixed benchmark. We argue instead that an LLM judge operating in a deployed system is better understood as having a lifecycle. It must be built, trained, deployed, and continuously maintained as the surrounding data evolves, and each phase poses distinct technical and operational challenges. We present such a lifecycle for the LLM judges that evaluate recommendation explanations at Netflix. Everything we report comes out of a series of controlled online member-facing experiments, in which our pipeline generated and the judges assessed hundreds of thousands of distinct show-level explanations per week across a changing catalog. Our framework has four phases. (I) Birth defines the evaluation criteria and builds curated benchmark datasets with human labels and rationales. (II) Training refines the judges' rubrics via Reasoning-Aligned Rubric Tuning (RART), which uses a meta-judge over reasoning output as the learning signal. (III) Deployment puts one judge in two online roles, quality gating and reflective generation. (IV) Monitoring runs a continuous Human-in-the-Loop (HITL) alignment process that detects drift and triggers re-tuning behind a human review gate. We report results from a five-week online A/B test over tens of millions of members on the Netflix mobile app, in which judge-aligned explanations shifted member viewing toward novel content (previously unwatched) and increased successful browse-to-play sessions relative to a no-explanation control, with no quality-related escalations.

LLM recommendation agents increasingly produce structured recommendation reports: sets of items accompanied by natural-language justifications. Yet existing evaluations often reduce this setting to reranking small shortlisted candidate sets or judge reports mainly by semantic plausibility. We introduce Recommendation Atlas (Agentic Tool-Level Assessment for Shopping), or RecoAtlas, a benchmark and toolkit for evaluating shopping agents with behavior-grounded metrics. RecoAtlas complements held-out interaction metrics with learned utility proxies for relevance, complementarity, and diversity derived from interaction data, while separately measuring semantic coherence and explanation quality. Its controlled tool environment exposes agents to either semantic, behavior-aligned, or faulty tools, enabling diagnosis of whether performance gains arise from stronger reasoning, better signals, or more effective tool-use policies. Across controlled experiments, we show that RecoAtlas exhibits key properties of a meaningful benchmark for agentic systems: performance scales with model capacity and test-time compute, improves with stronger and better-aligned tools, degrades under noisy or misaligned signals, and reveals that semantic plausibility does not necessarily capture behavior-grounded utility. RecoAtlas provides a foundation for developing and evaluating shopping assistants that optimize not only for plausible recommendations, but also for coherent, behaviorally grounded recommendation sets.

Large language models are increasingly used as interactive recommender assistants. Their evaluation should therefore go beyond plausible item recommendation and test whether they can recognize flawed recommendation requests. Existing recommender benchmarks mainly assess ranking, generation, or preference satisfaction, while existing error-detection benchmarks are usually not grounded in recommendation-specific user and candidate evidence. To address this gap, we introduce RPCBench, a benchmark for evaluating Recommender-Premise Critique: the ability to detect, diagnose, and properly handle faulty premises in natural-language recommendation requests. RPCBench contains evidence-grounded test instances from five recommendation domains and covers ten types of premise failures. Each instance provides a visible recommendation context and a corrupted user query. We further design a fine-grained evaluation framework that measures proactive detection, error localization, post-detection handling strategy, and evidence faithfulness. Through a systematic evaluation of 11 LLMs, we find that proactive detection is the main bottleneck in Recommender-Premise Critique, and models perform worst on underspecified-premise errors. We also observe that target-critical information density matters more than redundant evidence, and that longer reasoning does not monotonically improve critique quality: performance peaks at intermediate reasoning length, while overly long reasoning is accompanied by an overthinking penalty. The code is available at https://github.com/ZhongruChen/RPCBench.