Huanjin Yao

Huanjin Yao, Wenhao Wu, Zhiheng Li · amazon-science

Large pre-trained vision models achieve impressive success in computer vision. However, fully fine-tuning large models for downstream tasks, particularly in video understanding, can be prohibitively computationally expensive. Recent studies turn their focus towards efficient image-to-video transfer learning. Nevertheless, existing efficient fine-tuning methods lack attention to training memory usage and exploration of transferring a larger model to the video domain. In this paper, we present a novel Spatial-Temporal Side Network for memory-efficient fine-tuning large image models to video understanding, named Side4Video. Specifically, we introduce a lightweight spatial-temporal side network attached to the frozen vision model, which avoids the backpropagation through the heavy pre-trained model and utilizes multi-level spatial features from the original image model. Extremely memory-efficient architecture enables our method to reduce 75% memory usage than previous adapter-based methods. In this way, we can transfer a huge ViT-E (4.4B) for video understanding tasks which is 14x larger than ViT-L (304M). Our approach achieves remarkable performance on various video datasets across unimodal and cross-modal tasks (i.e., action recognition and text-video retrieval), especially in Something-Something V1&V2 (67.3% & 74.6%), Kinetics-400 (88.6%), MSR-VTT (52.3%), MSVD (56.1%) and VATEX (68.8%). We release our code at https://github.com/HJYao00/Side4Video.

Huanjin Yao, Qixiang Yin, Jingyi Zhang et al.

In this work, we aim to incentivize the reasoning ability of Multimodal Large Language Models (MLLMs) via reinforcement learning (RL) and develop an effective approach that mitigates the sparse reward and advantage vanishing issues during RL. To this end, we propose Share-GRPO, a novel RL approach that tackle these issues by exploring and sharing diverse reasoning trajectories over expanded question space. Specifically, Share-GRPO first expands the question space for a given question via data transformation techniques, and then encourages MLLM to effectively explore diverse reasoning trajectories over the expanded question space and shares the discovered reasoning trajectories across the expanded questions during RL. In addition, Share-GRPO also shares reward information during advantage computation, which estimates solution advantages hierarchically across and within question variants, allowing more accurate estimation of relative advantages and improving the stability of policy training. Extensive evaluations over six widely-used reasoning benchmarks showcase the superior performance of our method. Code will be available at https://github.com/HJYao00/R1-ShareVL.

Qixiang Yin, Huanjin Yao, Jianghao Chen et al.

Although Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities across diverse tasks, they encounter challenges in terms of reasoning efficiency, large model size and overthinking. However, existing lightweight MLLMs lack the capability to balance high efficiency and performance at a small scale. To this end, we propose Tiny-R1V, a novel lightweight 3B model that achieves faster inference and higher accuracy via a two-stage optimization, while unifying multimodal reasoning across multiple tasks with fewer inference tokens. In the first stage, Tiny-R1V introduces Length-Informed Relative Policy Optimization (LIPO), a new reinforcement learning method, to train each reasoning model, including mathematical reasoning, chart reasoning, and OCR capability. The LIPO dynamically adjusts the advantages of responses within groups by prioritizing concise yet high-quality responses to encourage the generation of shorter and more accurate responses. In the second stage, we propose Adaptive Model Merging (AMM), a training-free model merging method that merges multiple specialist models into a unified architecture. Specifically, AMM adaptively adjusts the weights of task vectors via a novel gradient projection regularization loss function, thus mitigating redundant conflicts between them. Extensive evaluations on ten widely-used reasoning benchmarks covering mathematics, structured data (charts, tables, documents), OCR, and general capabilities showcase the superior performance of Tiny-R1V, enabling lightweight models to excel in diverse multimodal reasoning tasks. Code will be available at \href{https://github.com/buptyqx/Tiny-R1V}{https://github.com/buptyqx/Tiny-R1V}

Jingyi Zhang, Jiaxing Huang, Huanjin Yao et al.

Recent studies generally enhance MLLMs' reasoning capabilities via supervised fine-tuning on high-quality chain-of-thought reasoning data, which often leads models to merely imitate successful reasoning paths without understanding what the wrong reasoning paths are. In this work, we aim to enhance the MLLMs' reasoning ability beyond passively imitating positive reasoning paths. To this end, we design Step-wise Group Relative Policy Optimization (StepGRPO), a new online reinforcement learning framework that enables MLLMs to self-improve reasoning ability via simple, effective and dense step-wise rewarding. Specifically, StepGRPO introduces two novel rule-based reasoning rewards: Step-wise Reasoning Accuracy Reward (StepRAR) and Step-wise Reasoning Validity Reward (StepRVR). StepRAR rewards the reasoning paths that contain necessary intermediate reasoning steps via a soft key-step matching technique, while StepRAR rewards reasoning paths that follow a well-structured and logically consistent reasoning process through a reasoning completeness and logic evaluation strategy. With the proposed StepGRPO, we introduce R1-VL, a series of MLLMs with outstanding capabilities in step-by-step reasoning. Extensive experiments over 8 benchmarks demonstrate the superiority of our methods.