Retrv-R1: A Reasoning-Driven MLLM Framework for Universal and Efficient Multimodal Retrieval

TL;DR

Retrv-R1 addresses the challenge of applying reinforcement-learning-based reasoning to universal multimodal retrieval by introducing an information compression module with a details-inspection mechanism, a synthetic CoT-activation stage, and a curriculum-based RL reward. It combines a two-stage coarse-to-fine retrieval architecture with a specially designed ICM that compresses candidates into two tokens per item, preserving critical content and relationships, and a two-phase training pipeline (SFT on synthetic CoT data followed by RL with a token-efficiency curriculum). Empirical results on M-BEIR and out-of-domain benchmarks show state-of-the-art retrieval performance and strong generalization, along with markedly improved efficiency and practical applicability to RAG tasks. The work demonstrates that explicit, task-tailored reasoning and structured information retention can endow multimodal retrieval systems with robust accuracy, efficiency, and adaptability across diverse data modalities and tasks.

Abstract

The success of DeepSeek-R1 demonstrates the immense potential of using reinforcement learning (RL) to enhance LLMs' reasoning capabilities. This paper introduces Retrv-R1, the first R1-style MLLM specifically designed for multimodal universal retrieval, achieving higher performance by employing step-by-step reasoning to produce more accurate retrieval results. We find that directly applying the methods of DeepSeek-R1 to retrieval tasks is not feasible, mainly due to (1) the high computational cost caused by the large token consumption required for multiple candidates with reasoning processes, and (2) the instability and suboptimal results when directly applying RL to train for retrieval tasks. To address these issues, Retrv-R1 introduces an information compression module with a details inspection mechanism, which enhances computational efficiency by reducing the number of tokens while ensuring that critical information for challenging candidates is preserved. Furthermore, a new training paradigm is proposed, including an activation stage using a retrieval-tailored synthetic CoT dataset for more effective optimization, followed by RL with a novel curriculum reward to improve both performance and efficiency. Incorporating these novel designs, Retrv-R1 achieves SOTA performance, high efficiency, and strong generalization ability, as demonstrated by experiments across multiple benchmarks and tasks.

Figures (7)

• Figure 1: Illustration of (A) overall pipeline; (B) structure of information compression module (ICM); and (C) overall process of model training with three stages.
• Figure 2: Analysis results of RL fine-tuning.
• Figure 3: A qualitative example of the synthesized retrieval CoT data for SFT.
• Figure 4: A qualitative example of the retrieval result generated from Retrv-R1.
• Figure 5: A qualitative example of the retrieval result generated from Retrv-R1.