Research map / Multi-agent systems and games / LLM agents

VRAgent-R1: Boosting Video Recommendation with MLLM-based Agents via Reinforcement Learning

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 1 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 VRAgent-R1, pairing an MLLM-based Item Perception agent (key-frame retrieval, collaborative perception, recommendation-relevant summarization) with a GRPO-trained User Simulation agent that reasons over user history before predicting preferences and next videos.

Problem

Frozen-LLM prompt-based user simulators give poor item modeling and user preference fidelity for video recommendation, and SFT yields binary answers without reasoning.

Approach

Use an MLLM to progressively distill multimodal video content into concise textual captions for recommendation, and fine-tune a Qwen2.5-7B user simulator with GRPO and rewards tied to real user behavior (preference judgment and next-video selection), with CoT reasoning about user status.

Key results

• IP Agent with SASRec improves HR@10 by 4.3% and NDCG@10 by 6.0% on MicroLens-100K (0.0994 / 0.0548).
• US Agent reaches 0.715 accuracy on preference judgment vs 0.535 for GPT-4o and 0.585 for MLLM-MSR (SFT).
• Next-video selection accuracy with m=3 is 0.641 vs 0.442 for MLLM-MSR (SFT), about 45% relative gain.

Surprising findings

• Qwen2.5-7B-based prompt-only simulation performs near random chance, and DeepSeek-R1 and GPT-4o are only slightly better.
• Removing CoT reasoning from RFT drops accuracy sharply (0.580 vs 0.715 accuracy with CoT).

Limitations

• User action space limited to watch and comment behaviors in the dataset.
• Code and data not released; release deferred pending commercial approval.
• Evaluation on a single video dataset with a limited number of cold-start users (1000) and simulation tasks.

Experimental setup

Datasets: MicroLens-100K MovieLens-1M

Benchmarks: MicroLens-100K MovieLens-1M

Metrics: HR@10 NDCG@10 HR@20 NDCG@20 Accuracy F1 Recall Precision selection accuracy

Baselines: GPT-4o DeepSeek-R1 Qwen2.5-7b LLM_Simulator Agent4Rec TALLREC MLLM-MSR RecAgent DSSM LightGCN DeepFM NextItNet

Builds on: GRPO DeepSeek-R1 Visual-RFT SASRec MLLM-MSR

Closest papers by meaning

• Contrastive Representation for Interactive Recommendation (AAAI 2025 · similarity 0.86)
• AgentRecBench: Benchmarking LLM Agent-based Personalized Recommender Systems (NeurIPS 2025 · similarity 0.86)
• LLM-Powered User Simulator for Recommender System (AAAI 2025 · similarity 0.86)
• ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding (NeurIPS 2025 · similarity 0.85)
• UNEX-RL: Reinforcing Long-Term Rewards in Multi-Stage Recommender Systems with UNidirectional EXecution (AAAI 2024 · similarity 0.84)
• AgentSelect: Benchmark for Narrative Query-to-Agent Recommendation (ICML 2026 · similarity 0.84)