Research map / Large language models / Reasoning and chain of thought

R$^2$ec: Towards Large Recommender Models with Reasoning

NeurIPS 2025 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes R²ec, a large recommender with a dual-head (LM head for reasoning, recommendation head for item scoring) trained with RecPO, an RL framework using a fused discrete-ranking and softmax-similarity reward, achieving higher accuracy and lower latency than LLM-based recommenders on three Amazon datasets.

Problem

Existing LLM-based recommenders lack intrinsic reasoning, and reasoning-augmented pipelines use separate reasoning and recommendation modules that are costly and cannot be jointly optimized end-to-end, while annotated reasoning data for recommendation is scarce.

Approach

A single LLM with a language-modeling head that generates reasoning tokens and a recommendation head that scores items via inner product with an item embedding table derived from the same model; trained with GRPO/RLOO-style group advantages over sampled reasoning paths using a fused reward (NDCG plus softmax similarity), without reasoning annotations.

Key results

• On Instruments, R²ec (Gemma2-2B) reaches H@5 0.02635 vs 0.01750 for best traditional baseline SASRec and 0.01301 for LangPTune, with statistically significant improvements.
• On CDs and Vinyl, R²ec (Gemma2-2B) reports H@10 of 0.080357 and N@10 of 0.047154, with relative gains of up to about 71% over the best baseline on H@10.
• Average inference latency is 1.67 s on a single RTX 3090, versus 1.90 s for LangPTune, 4.62 s for D3, and 5.23 s for LLaRA; 0.0945 s with vLLM.

Measured effects reported in the paper

Surprising findings

• Gemma2-2B consistently outperforms the larger Qwen2.5-3B backbone in the reported experiments, by up to 2x on D3.
• Adding a continuous softmax-similarity reward alone (w/o Rd) performs worse than using only the discrete ranking reward (w/o Rc), suggesting the continuous term introduces noise when used alone.

Limitations

• Explicit reasoning generation increases inference latency due to autoregressive decoding.
• Only LoRA parameter-efficient tuning was used due to compute constraints, so full fine-tuning performance is unexplored.
• Reasoning-strategy analysis relies on 100 sampled outputs per dataset with human and GPT-4.1 annotation.

Experimental setup

Datasets: Amazon CDs and Vinyl Amazon Video Games Amazon Musical Instruments Electronics MovieLens GoodReads

Benchmarks: Amazon CDs and Vinyl Amazon Video Games Amazon Musical Instruments

Metrics: Hit Rate@K NDCG@K

Baselines: GRU4Rec Caser SASRec TIGER BigRec D^3 SDPO LLaRA SPRec LangPTune

Builds on: GRPO RLOO LLaRA LangPTune TIGER

Closest papers by meaning

• Reinforced Latent Reasoning for LLM-based Recommendation (ICLR 2026 · similarity 0.91)
• Think before Recommendation: Autonomous Reasoning-enhanced Recommender (NeurIPS 2025 · similarity 0.90)
• Re2LLM: Reflective Reinforcement Large Language Model for Session-based Recommendation (AAAI 2025 · similarity 0.89)
• LLMRG: Improving Recommendations through Large Language Model Reasoning Graphs (AAAI 2024 · similarity 0.88)
• ConvRec-R1: Training LLM-based Conversational Recommender Systems with Reinforcement Learning (ICLR 2026 · similarity 0.88)
• On the Unexpected Effectiveness of Reinforcement Learning for Sequential Recommendation (ICML 2024 · similarity 0.87)