ThinkRec: Thinking-based recommendation via LLM

Recent advances in large language models (LLMs) have enabled more semantic-aware recommendations through natural language generation. Existing LLM for recommendation (LLM4Rec) methods mostly operate in a System 1-like manner, relying on superficial features to match similar items based on click history, rather than reasoning through deeper behavioral logic. This often leads to superficial and erroneous recommendations. Motivated by this, we propose ThinkRec, a thinking-based framework that shifts LLM4Rec from System 1 to System 2 (rational system). Technically, ThinkRec introduces a thinking activation mechanism that augments item metadata with keyword summarization and injects synthetic reasoning traces, guiding the model to form interpretable reasoning chains that consist of analyzing interaction histories, identifying user preferences, and making decisions based on target items. On top of this, we propose an instance-wise expert fusion mechanism to reduce the reasoning difficulty. By dynamically assigning weights to expert models based on users’ latent features, ThinkRec adapts its reasoning path to individual users, thereby enhancing precision and personalization. Extensive experiments on real-world datasets demonstrate that ThinkRec significantly improves the accuracy and interpretability of recommendations. Our implementations are available at https://github.com/Yu-Qi-hang/ThinkRec.

💡 Research Summary

The paper “ThinkRec: Thinking‑based Recommendation via LLM” addresses a fundamental shortcoming of current large‑language‑model (LLM) based recommender systems: they operate largely as System 1 (fast, intuitive) models, matching items based on superficial click patterns without engaging deeper behavioral reasoning. To shift LLM‑based recommendation toward a System 2 (slow, rational) paradigm, the authors propose ThinkRec, a framework that (1) activates explicit “thinking” in the LLM through data augmentation and synthetic reasoning traces, and (2) adapts the reasoning process to each user via an instance‑wise expert fusion mechanism built on Low‑Rank Adaptation (LoRA) modules.

Thinking activation. The authors first enrich each item’s textual representation. Raw titles and descriptions are passed through a pre‑trained summarization model to extract up to ten salient keywords, which are concatenated with the original text. This keyword‑augmented metadata supplies concise, high‑signal semantic cues for the LLM. Because public recommendation datasets lack explicit reasoning steps, the authors synthesize a subset of training instances using a strong reasoning model (e.g., Qwen‑7B). For each sampled user‑item pair, they generate a chain‑of‑thought style explanation that outlines (a) analysis of the user’s historical interactions, (b) identification of preferences, (c) evaluation of the target item, and (d) a natural‑language justification. These synthetic reasoning examples are mixed with ordinary binary interaction samples (click / no‑click) during training. The LLM is thus jointly optimized on a language‑modeling objective that includes both the answer token and the reasoning text, encouraging it to produce not only a recommendation decision but also a human‑readable rationale.

Instance‑wise expert fusion. Recognizing that user behavior is highly heterogeneous, ThinkRec introduces a personalized expert system. Traditional collaborative‑filtering models (MF, LightGCN) generate latent user vectors. These vectors are clustered to define user groups, each associated with a distinct LoRA adapter that is fine‑tuned on the group’s data. At inference time, a gating network computes similarity between a target user’s latent vector and each LoRA’s representation, producing a weighted combination of adapters. This dynamic fusion yields a user‑specific reasoning path, reducing the cognitive load on any single model and improving both accuracy and interpretability. Because the reasoning capability is distilled into the LoRA parameters, the final ranking stage can be performed with a single forward pass that outputs a single token (e.g., “Yes”/“No”), making the system efficient for large‑scale deployment.

Experiments. The authors evaluate ThinkRec on three real‑world datasets (including Amazon Books and MovieLens variants). Compared with state‑of‑the‑art LLM‑based recommenders (e.g., P5, InstrucRec) and strong collaborative baselines, ThinkRec achieves an average AUC gain of 7.96 % and a METEOR improvement of 56.54 % for generated explanations. Ablation studies confirm that each component—keyword augmentation, synthetic reasoning data, and LoRA‑based expert fusion—contributes positively to performance, with the expert fusion providing the largest boost in personalized accuracy. Qualitative examples demonstrate that ThinkRec can avoid superficial genre‑matching errors (e.g., recommending “Hyperion” merely because it shares the “science‑fiction” tag) by reasoning about nuanced user dislikes (philosophical themes) that are captured in the synthetic chain‑of‑thought.

Limitations and future work. The approach relies on the quality of the synthetic reasoning traces; noisy or biased explanations could misguide the LLM. Moreover, the number of LoRA experts grows with the granularity of user clustering, potentially increasing memory and compute demands. The authors suggest future research on automated verification of generated reasoning, more scalable expert management (e.g., hierarchical adapters), and extension to multimodal domains such as video or music recommendation.

Contributions. In summary, ThinkRec introduces (1) a novel thinking activation pipeline that bridges recommendation objectives with language‑model reasoning, (2) an instance‑wise expert fusion mechanism that personalizes the reasoning process, and (3) extensive empirical evidence that these innovations jointly improve both recommendation accuracy and the interpretability of LLM‑generated justifications. The work positions LLMs not merely as powerful scorers but as genuine reasoning agents within recommender systems, opening a new research direction toward rational, explainable recommendation.