SenseNova-MARS: Empowering Multimodal Agentic Reasoning and Search via Reinforcement Learning

TL;DR

SenseNova-MARS tackles the gap in vision-language models' ability to interleave dynamic tool use with reasoning in knowledge-intensive visual tasks. It proposes a two-stage training pipeline (cold-start SFT followed by reinforcement learning) and a novel BN-GSPO algorithm to stabilize multi-tool RL trajectories. The framework combines image search, text search, and image crop as integral tools within a unified agentive reasoning process and introduces the HR-MMSearch benchmark for high-resolution, knowledge-rich evaluation. Experiments show state-of-the-art performance on open-source and perception benchmarks, demonstrating robust tool invocation and fine-grained visual understanding with strong practical impact for multimodal AI systems.

Abstract

While Vision-Language Models (VLMs) can solve complex tasks through agentic reasoning, their capabilities remain largely constrained to text-oriented chain-of-thought or isolated tool invocation. They fail to exhibit the human-like proficiency required to seamlessly interleave dynamic tool manipulation with continuous reasoning, particularly in knowledge-intensive and visually complex scenarios that demand coordinated external tools such as search and image cropping. In this work, we introduce SenseNova-MARS, a novel Multimodal Agentic Reasoning and Search framework that empowers VLMs with interleaved visual reasoning and tool-use capabilities via reinforcement learning (RL). Specifically, SenseNova-MARS dynamically integrates the image search, text search, and image crop tools to tackle fine-grained and knowledge-intensive visual understanding challenges. In the RL stage, we propose the Batch-Normalized Group Sequence Policy Optimization (BN-GSPO) algorithm to improve the training stability and advance the model's ability to invoke tools and reason effectively. To comprehensively evaluate the agentic VLMs on complex visual tasks, we introduce the HR-MMSearch benchmark, the first search-oriented benchmark composed of high-resolution images with knowledge-intensive and search-driven questions. Experiments demonstrate that SenseNova-MARS achieves state-of-the-art performance on open-source search and fine-grained image understanding benchmarks. Specifically, on search-oriented benchmarks, SenseNova-MARS-8B scores 67.84 on MMSearch and 41.64 on HR-MMSearch, surpassing proprietary models such as Gemini-3-Flash and GPT-5. SenseNova-MARS represents a promising step toward agentic VLMs by providing effective and robust tool-use capabilities. To facilitate further research in this field, we will release all code, models, and datasets.

Figures (19)

• Figure 1: Overall performance of SenseNova-MARS-8B compares to other models across six benchmarks. All other models are evaluated under agentic workflow. SenseNova-MARS-8B demonstrates exceptional performance on the search-oriented benchmarks such as MMSearch jiang2024mmsearch, HR-MMSearch and FVQA wu2025mmsearch, surpassing leading proprietary models such as Gemini-3-Flash gemini3flash and GPT-5 openai2025gpt5. For the high-resolution perception benchmark such as V* Bench wu2024v and HR-Bench wang2025divide, SenseNova-MARS-8B also outperforms existing open-source models, including DeepEyesV2 hong2025deepeyesv2 and Mini o3 lai2025mini.
• Figure 2: Reasoning trajectory of SenseNova-MARS. SenseNova-MARS tackles the challenging visual task by leveraging an integrated suite of text search, image search, and image crop tools within the reasoning process.
• Figure 3: The illustration of SenseNova-MARS RL training pipeline. SenseNova-MARS adaptively invokes the image search, text search and image crop tools in the multi-turn reasoning process to obtain the final answer. The policy VLM is optimized by the BN-GSPO algorithm, driven by the format reward and answer reward.
• Figure 4: Cold-start data generation pipeline. It consists of data mining, trajectory synthesis and quality verification.
• Figure 5: Statistics of our proposed HR-MMSearch benchmark. HR-MMSearch is characterized by the high-resolution images and knowledge-intensive question, covering areas such as Sports, Leisure&Culture, Science&Technology, Business&Finance, Games, and Academic Research.