From Lists to Decision Reports: The Deep Research Framework for Recommender Systems

Motivation

Typical recommender pipelines follow four steps: (1) model user interests from historical behavior, (2) retrieve candidates from a pool, (3) rank the candidates, and (4) present the results as a list. This "tool" paradigm leaves exploration, comparison, and information synthesis entirely to the user, incurring high decision‑making costs.

The authors propose a "Deep Research" paradigm that extends recommendation beyond list exposure to autonomous exploration and structured decision reporting.

RecPilot Framework

RecPilot consists of two cooperating agents.

User Trajectory Simulation Agent Captures the evolution of user intent by modeling action‑guided trajectories. An action‑guided aggregation strategy structures user behavior across interaction stages, enabling the model to learn the transition from broad browsing to final purchase. Reinforcement learning with a non‑model‑based reward function supplies three reward dimensions: result reward, semantic consistency, and logical constraints. This avoids over‑fitting to historical patterns and generates a high‑confidence candidate set by parallel exploration of multiple possible intent evolutions.

Self‑Evolving Report Generation Agent After obtaining the candidate set, the agent builds a dual‑channel Rubric–Experience model. Rubrics provide attribute‑based quantitative scores, while Experience extracts contextual signals from user text or behavior. The agent decomposes a complex purchase intent into multiple sub‑dimensions, scores items per dimension, and continuously updates preference weights from real feedback (e.g., final purchase) without retraining, achieving a closed‑loop self‑evolution.

The final report contains four modules: simulated exploration paths, user‑intent summary, a consolidated recommendation list, and multi‑dimensional item analysis.

Experiments

Evaluations were conducted on a real‑interaction dataset from TMALL.

Trajectory Simulation RecPilot significantly outperformed traditional sequential recommenders (SASRec, BERT4Rec) and advanced multi‑behavior/inference baselines (MBSTR, ReaRec). Ablation studies confirmed that high‑quality trajectory modeling is the primary driver of performance gains.

Report Generation A double‑blind test involving large language models and human judges measured six metrics: accuracy, coverage, information amount, clarity, consistency, and novelty. Compared with strong agent baselines such as Plan‑and‑Solve, RecPilot achieved a 77% win rate on the novelty metric, demonstrating the advantage of multi‑aspect interest decomposition.

Case Study: Buying a Refrigerator

In a traditional list mode, the system shows only images, titles, and prices, forcing the user to click each item to inspect parameters (e.g., number of doors, energy consumption). RecPilot’s deep‑report mode proceeds as follows:

Display the simulated exploration path, showing how the AI compared and filtered items.

Summarize the core intent (e.g., three‑door fridge with smart temperature control).

Present a top recommendation for rapid decision making.

List alternative recommendations aligned with different priorities (e.g., large capacity, high energy efficiency), each accompanied by rubric scores and experience cues.

This structured report dramatically reduces the user’s comparison burden.

Conclusion

RecPilot transforms recommender systems from passive exposure tools into active decision‑assistant agents. The framework is especially suitable for high‑cost decision domains, and a hybrid deployment that combines fast list‑based recommendations with deep analytical reports may offer a practical solution.

Repository:

https://github.com/RUCAIBox/RecPilot