From Passive Exposure to Active Decision Assistant: Deep Research Framework for Recommenders

Motivation

Current recommender pipelines model user interest, retrieve candidates, rank them, and present a list, leaving exploration, comparison and synthesis to the user. This “tool” paradigm burdens users with decision cost.

Inspired by the “Deep Research” paradigm in information retrieval, the authors propose a new framework that generates a structured decision report instead of a plain list.

Deep Research Paradigm and RecPilot Framework

The Deep Research paradigm treats the recommender as an autonomous agent that explores the product space and produces a report. RecPilot implements this with two cooperating agents:

User Trajectory Simulation Agent : captures intent evolution, uses action‑guided aggregation to model the transition from broad browsing to final purchase, and employs reinforcement learning with multi‑dimensional rewards (result, semantic consistency, path constraints) to generate diverse, high‑confidence candidate sets.

Self‑evolving Report Generation Agent : builds a Rubric‑Experience dual‑channel model—Rubrics provide attribute‑based scores, Experience extracts contextual signals from user behavior—and decomposes multi‑aspect interests into sub‑dimensions. It then produces a structured report containing exploration paths, intent summary, a consolidated recommendation list, and per‑dimension item analysis. The agent updates preference weights from real feedback without retraining.

Experimental Evaluation

RecPilot was evaluated on a real‑world TMALL interaction dataset. In trajectory‑simulation tests it outperformed sequence baselines (SASRec, BERT4Rec) and advanced multi‑behavior baselines (MBSTR, ReaRec). Ablation studies confirmed that high‑quality trajectory modeling is the key driver.

Report‑generation quality was measured with six metrics (accuracy, coverage, information, clarity, consistency, novelty) using a double‑blind test with large language models and human judges. Compared with strong agent baselines such as Plan‑and‑Solve, RecPilot achieved overall superiority, especially a 77 % win rate on the novelty metric.

Case Study: Buying a Refrigerator

In the traditional list mode the system shows only images, titles and prices, forcing the user to inspect each item manually. RecPilot’s deep‑report mode first displays the simulated exploration trajectory, then summarizes the user’s core intent (e.g., three‑door fridge, smart temperature control), provides a top recommendation, and lists multi‑aspect analyses (capacity‑focused, energy‑saving options), dramatically reducing comparison effort.

Conclusion

RecPilot demonstrates that a recommender can evolve from a passive exposure tool to an active decision‑assistant, delivering structured, explainable reports that lower user decision cost. The authors suggest combining fast list‑based recommendation with deep‑analysis reports for high‑cost decision scenarios, and anticipate future systems that answer not only “what to recommend” but also “why” and “how to decide”.