GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning

We present GLM-4.1V-Thinking, a vision-language model (VLM) designed to advance general-purpose multimodal understanding and reasoning. In this report, we share our key findings in the development of the reasoning-centric training framework. We first develop a capable vision foundation model with significant potential through large-scale pre-training, which arguably sets the upper bound for the final performance. We then propose Reinforcement Learning with Curriculum Sampling (RLCS) to unlock the full potential of the model, leading to comprehensive capability enhancement across a diverse range of tasks, including STEM problem solving, video understanding, content recognition, coding, grounding, GUI-based agents, and long document understanding. We open-source GLM-4.1V-9B-Thinking, which achieves state-of-the-art performance among models of comparable size. In a comprehensive evaluation across 28 public benchmarks, our model outperforms Qwen2.5-VL-7B on nearly all tasks and achieves comparable or even superior performance on 18 benchmarks relative to the significantly larger Qwen2.5-VL-72B. Notably, GLM-4.1V-9B-Thinking also demonstrates competitive or superior performance compared to closed-source models such as GPT-4o on challenging tasks including long document understanding and STEM reasoning, further underscoring its strong capabilities. Code, models and more information are released at https://github.com/THUDM/GLM-4.1V-Thinking.

Urban1960SatSeg: Unsupervised Semantic Segmentation of Mid-20$^{th}$ century Urban Landscapes with Satellite Imageries

Historical satellite imagery, such as mid-20$^{th}$ century Keyhole data, offers rare insights into understanding early urban development and long-term transformation. However, severe quality degradation (e.g., distortion, misalignment, and spectral scarcity) and annotation absence have long hindered semantic segmentation on such historical RS imagery. To bridge this gap and enhance understanding of urban development, we introduce $\textbf{Urban1960SatBench}$, an annotated segmentation dataset based on historical satellite imagery with the earliest observation time among all existing segmentation datasets, along with a benchmark framework for unsupervised segmentation tasks, $\textbf{Urban1960SatUSM}$. First, $\textbf{Urban1960SatBench}$ serves as a novel, expertly annotated semantic segmentation dataset built on mid-20$^{th}$ century Keyhole imagery, covering 1,240 km$^2$ and key urban classes (buildings, roads, farmland, water). As the earliest segmentation dataset of its kind, it provides a pioneering benchmark for historical urban understanding. Second, $\textbf{Urban1960SatUSM}$(Unsupervised Segmentation Model) is a novel unsupervised semantic segmentation framework for historical RS imagery. It employs a confidence-aware alignment mechanism and focal-confidence loss based on a self-supervised learning architecture, which generates robust pseudo-labels and adaptively prioritizes prediction difficulty and label reliability to improve unsupervised segmentation on noisy historical data without manual supervision. Experiments show Urban1960SatUSM significantly outperforms existing unsupervised segmentation methods on Urban1960SatSeg for segmenting historical urban scenes, promising in paving the way for quantitative studies of long-term urban change using modern computer vision. Our benchmark and supplementary material are available at https://github.com/Tianxiang-Hao/Urban1960SatSeg.

DiscoVLA: Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval

The parameter-efficient adaptation of the image-text pretraining model CLIP for video-text retrieval is a prominent area of research. While CLIP is focused on image-level vision-language matching, video-text retrieval demands comprehensive understanding at the video level. Three key discrepancies emerge in the transfer from image-level to video-level: vision, language, and alignment. However, existing methods mainly focus on vision while neglecting language and alignment. In this paper, we propose Discrepancy Reduction in Vision, Language, and Alignment (DiscoVLA), which simultaneously mitigates all three discrepancies. Specifically, we introduce Image-Video Features Fusion to integrate image-level and video-level features, effectively tackling both vision and language discrepancies. Additionally, we generate pseudo image captions to learn fine-grained image-level alignment. To mitigate alignment discrepancies, we propose Image-to-Video Alignment Distillation, which leverages image-level alignment knowledge to enhance video-level alignment. Extensive experiments demonstrate the superiority of our DiscoVLA. In particular, on MSRVTT with CLIP (ViT-B/16), DiscoVLA outperforms previous methods by 1.5% in R@1, reaching a final score of 50.5% R@1. The code is available at https://github.com/LunarShen/DsicoVLA.

SpecOffload: Unlocking Latent GPU Capacity for LLM Inference on Resource-Constrained Devices

Efficient LLM inference on resource-constrained devices presents significant challenges in compute and memory utilization. Due to limited GPU memory, existing systems offload model weights to CPU memory, incurring substantial I/O overhead between the CPU and GPU. This leads to two major inefficiencies: (1) GPU cores are underutilized, often remaining idle while waiting for data to be loaded; and (2) GPU memory has low impact on performance, as reducing its capacity has minimal effect on overall throughput.In this paper, we propose SpecOffload, a high-throughput inference engine that embeds speculative decoding into offloading. Our key idea is to unlock latent GPU resources for storing and executing a draft model used for speculative decoding, thus accelerating inference at near-zero additional cost. To support this, we carefully orchestrate the interleaved execution of target and draft models in speculative decoding within the offloading pipeline, and propose a planner to manage tensor placement and select optimal parameters. Compared to the best baseline, SpecOffload improves GPU core utilization by 4.49x and boosts inference throughput by 2.54x. Our code is available at https://github.com/MobiSense/SpecOffload .

TempMe: Video Temporal Token Merging for Efficient Text-Video Retrieval

Most text-video retrieval methods utilize the text-image pre-trained CLIP as a backbone, incorporating complex modules that result in high computational overhead. As a result, many studies focus on efficient fine-tuning. The primary challenge in efficient adaption arises from the inherent differences between image and video modalities. Each sampled video frame must be processed by the image encoder independently, which increases complexity and complicates practical deployment. Although existing efficient methods fine-tune with small trainable parameters, they still incur high inference costs due to the large token number. In this work, we argue that temporal redundancy significantly contributes to the model's high complexity due to the repeated information in consecutive frames. Existing token compression methods for image models fail to solve the unique challenges, as they overlook temporal redundancy across frames. To tackle these problems, we propose Temporal Token Merging (TempMe) to reduce temporal redundancy. Specifically, we introduce a progressive multi-granularity framework. By gradually combining neighboring clips, we merge temporal tokens across different frames and learn video-level features, leading to lower complexity and better performance. Extensive experiments validate the superiority of our TempMe. Compared to previous efficient text-video retrieval methods, TempMe significantly reduces output tokens by 95% and GFLOPs by 51%, while achieving a 1.8X speedup and a 4.4% R-Sum improvement. Additionally, TempMe exhibits robust generalization capabilities by integrating effectively with both efficient and full fine-tuning methods. With full fine-tuning, TempMe achieves a significant 7.9% R-Sum improvement, trains 1.57X faster, and utilizes 75.2% GPU memory usage. Our code will be released.

Learn from the Learnt: Source-Free Active Domain Adaptation via Contrastive Sampling and Visual Persistence

Domain Adaptation (DA) facilitates knowledge transfer from a source domain to a related target domain. This paper investigates a practical DA paradigm, namely Source data-Free Active Domain Adaptation (SFADA), where source data becomes inaccessible during adaptation, and a minimum amount of annotation budget is available in the target domain. Without referencing the source data, new challenges emerge in identifying the most informative target samples for labeling, establishing cross-domain alignment during adaptation, and ensuring continuous performance improvements through the iterative query-and-adaptation process. In response, we present learn from the learnt (LFTL), a novel paradigm for SFADA to leverage the learnt knowledge from the source pretrained model and actively iterated models without extra overhead. We propose Contrastive Active Sampling to learn from the hypotheses of the preceding model, thereby querying target samples that are both informative to the current model and persistently challenging throughout active learning. During adaptation, we learn from features of actively selected anchors obtained from previous intermediate models, so that the Visual Persistence-guided Adaptation can facilitate feature distribution alignment and active sample exploitation. Extensive experiments on three widely-used benchmarks show that our LFTL achieves state-of-the-art performance, superior computational efficiency and continuous improvements as the annotation budget increases. Our code is available at https://github.com/lyumengyao/lftl.

Quantized Prompt for Efficient Generalization of Vision-Language Models

In the past few years, large-scale pre-trained vision-language models like CLIP have achieved tremendous success in various fields. Naturally, how to transfer the rich knowledge in such huge pre-trained models to downstream tasks and datasets becomes a hot topic. During downstream adaptation, the most challenging problems are overfitting and catastrophic forgetting, which can cause the model to overly focus on the current data and lose more crucial domain-general knowledge. Existing works use classic regularization techniques to solve the problems. As solutions become increasingly complex, the ever-growing storage and inference costs are also a significant problem that urgently needs to be addressed. While in this paper, we start from an observation that proper random noise can suppress overfitting and catastrophic forgetting. Then we regard quantization error as a kind of noise, and explore quantization for regularizing vision-language model, which is quite efficiency and effective. Furthermore, to improve the model's generalization capability while maintaining its specialization capacity at minimal cost, we deeply analyze the characteristics of the weight distribution in prompts, conclude several principles for quantization module design and follow such principles to create several competitive baselines. The proposed method is significantly efficient due to its inherent lightweight nature, making it possible to adapt on extremely resource-limited devices. Our method can be fruitfully integrated into many existing approaches like MaPLe, enhancing accuracy while reducing storage overhead, making it more powerful yet versatile. Extensive experiments on 11 datasets shows great superiority of our method sufficiently. Code is available at https://github.com/beyondhtx/QPrompt.

PYRA: Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation

Recently, the scale of transformers has grown rapidly, which introduces considerable challenges in terms of training overhead and inference efficiency in the scope of task adaptation. Existing works, namely Parameter-Efficient Fine-Tuning (PEFT) and model compression, have separately investigated the challenges. However, PEFT cannot guarantee the inference efficiency of the original backbone, especially for large-scale models. Model compression requires significant training costs for structure searching and re-training. Consequently, a simple combination of them cannot guarantee accomplishing both training efficiency and inference efficiency with minimal costs. In this paper, we propose a novel Parallel Yielding Re-Activation (PYRA) method for such a challenge of training-inference efficient task adaptation. PYRA first utilizes parallel yielding adaptive weights to comprehensively perceive the data distribution in downstream tasks. A re-activation strategy for token modulation is then applied for tokens to be merged, leading to calibrated token features. Extensive experiments demonstrate that PYRA outperforms all competing methods under both low compression rate and high compression rate, demonstrating its effectiveness and superiority in maintaining both training efficiency and inference efficiency for large-scale foundation models. Our code will be released to the public.

Re-parameterized Low-rank Prompt: Generalize a Vision-Language Model within 0.5K Parameters

With the development of large pre-trained vision-language models, how to effectively transfer the knowledge of such foundational models to downstream tasks becomes a hot topic, especially in a data-deficient scenario. Recently, prompt tuning has become a popular solution. When adapting the vision-language models, researchers freeze the parameters in the backbone and only design and tune the prompts. On the one hand, the delicate design of prompt tuning exhibits strong performance. On the other hand, complicated structures and update rules largely increase the computation and storage cost. Motivated by the observation that the evolution pattern of the generalization capability in visual-language models aligns harmoniously with the trend of rank variations in the prompt matrix during adaptation, we design a new type of prompt, Re-parameterized Low-rank Prompt (RLP), for both efficient and effective adaptation. Our method could largely reduce the number of tunable parameters and storage space, which is quite beneficial in resource-limited scenarios. Extensive experiments further demonstrate the superiority of RLP. In particular, RLP shows comparable or even stronger performance than the latest state-of-the-art methods with an extremely small number of parameters. On a series of tasks over 11 datasets, RLP significantly increases the average downstream accuracy of classic prompt tuning by up to 5.25% using merely 0.5K parameters.

Consolidator: Mergeable Adapter with Grouped Connections for Visual Adaptation

Recently, transformers have shown strong ability as visual feature extractors, surpassing traditional convolution-based models in various scenarios. However, the success of vision transformers largely owes to their capacity to accommodate numerous parameters. As a result, new challenges for adapting large models to downstream tasks arise. On the one hand, classic fine-tuning tunes all parameters in a huge model for every task and thus easily falls into overfitting, leading to inferior performance. On the other hand, on resource-limited devices, fine-tuning stores a full copy of parameters and thus is usually impracticable for the shortage of storage space. However, few works have focused on how to efficiently and effectively transfer knowledge in a vision transformer. Existing methods did not dive into the properties of visual features, leading to inferior performance. Moreover, some of them bring heavy inference cost though benefiting storage. To tackle these problems, we propose consolidator to modify the pre-trained model with the addition of a small set of tunable parameters to temporarily store the task-specific knowledge while freezing the backbone model. Motivated by the success of group-wise convolution, we adopt grouped connections across the features extracted by fully connected layers to construct tunable parts in a consolidator. To further enhance the model's capacity to transfer knowledge under a constrained storage budget and keep inference efficient, we consolidate the parameters in two stages: 1. between adaptation and storage, and 2. between loading and inference. On a series of downstream visual tasks, our consolidator can reach up to 7.56 better accuracy than full fine-tuning with merely 0.35% parameters, and outperform state-of-the-art parameter-efficient tuning methods by a clear margin. Code is available at https://github.com/beyondhtx/Consolidator.