[SIGIR 2025] RecThinker: Evolution from Passive Assistant to Proactive Investigator in Recommender Systems

TL;DR

RecThinker is a novel agentic framework that treats recommendation as an autonomous investigation. By analyzing what it doesn't know (information gaps), it proactively utilizes specialized tools to fetch user, item, and collaborative data. Through a combination of SFT and GRPO reinforcement learning, it sets a new SOTA for reasoning-heavy recommendation tasks.

1. The Death of the Passive Recommender

Most current LLM-based recommenders act as passively. You give them a user history, they give you a rank. Even recent "agentic" models typically follow a fixed loop: Read -> Think -> Rank.

The problem? Real-world data is messy. User profiles are often fragmented, and item descriptions are sparse. A passive agent tries to "hallucinate" its way through these gaps. RecThinker changes the objective: it asks, "Do I have enough information to make this decision?" If not, it stops to search for more evidence before committing to a ranking.

2. Methodology: The "Analyze-Plan-Act" Paradigm

RecThinker shifts the architecture toward an Investigator role. The core logic handles two types of evidence:

• User-Centric ($K_u$): Long-term preferences and inferred intent.
• Item-Centric ($K_{c_i}$): Semantic properties and context.

The Engine Under the Hood

The agent computes an Information Gap (${\Delta }_t$). If the evidence is insufficient, it invokes the Tool-Augmented Reasoning loop.

Figure 1: The RecThinker workflow, highlighting the Analyze-Plan-Act cycle and the transition from raw inputs to tool-enriched reasoning trajectories.

Specialized Recommender Tools

Unlike generic agents that just use "Google Search," RecThinker uses 5 domain-specific tools:

1. User Profile/History Search: For incremental retrieval of past behaviors.
2. Item Info Search: Navigates an Item Relation Graph (e.g., "also-bought" links).
3. Collaborative Tools: Uses Similar User Search (hybrid sparse/dense similarity) and Knowledge Graph Search (2/3-hop traversals) to find high-order signals.

3. Training: SFT Meets GRPO

The authors realized that standard fine-tuning isn't enough for complex tool-use. They implemented a two-stage strategy:

• Stage 1: Self-Augmented SFT: The model watches "perfect" trajectories that resulted in correct rankings and valid formats, internalizing the basic "logic" of when to call a tool.
• Stage 2: Policy Refinement via GRPO: Using Group Relative Policy Optimization, the model is rewarded not just for accuracy, but for Tool Efficiency. It gets penalized for being "too lazy" (no tools) or "too recursive" (infinite tool calls).

4. Experimental Showdown

RecThinker was tested against classic CF models (BPR, SASRec) and advanced agents (AgentCF, PersonaX).

Performance Highlights:

• Across the board SOTA: Outperformed baselines by ~11% on both Amazon and MovieLens.
• Data Density Resilience: Significant gains in "Sparse" scenarios where tools like the Knowledge Graph provided the missing links that other models couldn't find.

Table 1: Quantifying the superiority of RecThinker across diverse datasets. Note the massive jump compared to standard LLMRank.

Critical Insights from Ablation:

The History Tool and Item Tool are the heavy lifters. Without them, performance plummets. While the Knowledge Graph Tool is used less frequently (~11-15%), it acts as the "clutch" player for solving the hardest, most ambiguous recommendation cases.

5. Summary & Future Outlook

RecThinker proves that active investigation is the future of Large Recommender Models (LRMs). By explicitly modeling the "Information Gap," we can reduce hallucinations and improve matching precision.

Limitations: The multi-step reasoning comes at a test-time latency cost. Future work should look into "distilling" these complex reasoning chains back into smaller, faster student models while retaining the investigative spirit.

Takeaway for Practitioners: If your recommendation data is sparse, don't just prompt your LLM harder—give it the tools to find the missing pieces of the puzzle.