Deep-research agents are capable of executing multi-step web exploration, targeted retrieval, and sophisticated question answering. Despite their powerful capabilities, deep-research agents face two critical bottlenecks: (1) the lack of large-scale, challenging datasets with real-world difficulty, and (2) the absence of accessible, open-source frameworks for data synthesis and agent training. To bridge these gaps, we first construct DeepResearch-9K, a large-scale challenging dataset specifically designed for deep-research scenarios built from open-source multi-hop question-answering (QA) datasets via a low-cost autonomous pipeline. Notably, it consists of (1) 9000 questions spanning three difficulty levels from L1 to L3 (2) high-quality search trajectories with reasoning chains from Tongyi-DeepResearch-30B-A3B, a state-of-the-art deep-research agent, and (3) verifiable answers. Furthermore, we develop an open-source training framework DeepResearch-R1 that supports (1) multi-turn web interactions, (2) different reinforcement learning (RL) approaches, and (3) different reward models such as rule-based outcome reward and LLM-as-Judge feedback. Finally, empirical results demonstrate that agents trained on DeepResearch-9K under our DeepResearch-R1 achieve state-of-the-art results on challenging deep-research benchmarks. We release the DeepResearch-9K dataset on https://huggingface.co/datasets/artillerywu/DeepResearch-9K and the code of DeepResearch-R1 on https://github.com/Applied-Machine-Learning-Lab/SIGIR2026_DeepResearch-R1.

Sequential recommendation aims to capture user preferences by modeling sequential patterns in user-item interactions. However, these models are often influenced by noise such as accidental interactions, leading to suboptimal performance. Therefore, to reduce the effect of noise, some works propose explicitly identifying and removing noisy items. However, we find that simply relying on collaborative information may result in an over-denoising problem, especially for cold items. To overcome these limitations, we propose a novel framework: Interest Alignment for Denoising Sequential Recommendation (IADSR) which integrates both collaborative and semantic information. Specifically, IADSR is comprised of two stages: in the first stage, we obtain the collaborative and semantic embeddings of each item from a traditional sequential recommendation model and an LLM, respectively. In the second stage, we align the collaborative and semantic embeddings and then identify noise in the interaction sequence based on long-term and short-term interests captured in the collaborative and semantic modalities. Our extensive experiments on four public datasets validate the effectiveness of the proposed framework and its compatibility with different sequential recommendation systems. The code and data are released for reproducibility: https://github.com/Applied-Machine-Learning-Lab/IADSR.

High-frequency leaf area index (LAI) dataset is essential for vegetation dynamic monitoring and crop yield estimation. However, due to the negative impacts of land surface heterogeneity, current hectometric-resolution LAI products cannot satisfy the uncertainty requirement of LAI dataset in practice. Here, we proposed a method named Utilization of Sub-Pixel Information (USPI) that leverages fine-scale remote sensing data to improve the accuracy of hectometric-resolution LAI retrieval. Specifically, based on machine learning (ML) models trained by representative samples, we retrieved the USPI LAI from MODIS reflectance by incorporating the sub-pixel information from Sentinel-2 LAI estimates. The USPI LAI was comprehensively evaluated using 30-m LAI reference maps in three aspects: the performance of different ML models, the comparison with MODIS LAI products, and the potential correction of USPI LAI for clumping effect. Results showed that Gaussian Process Regression (GPR) model outperformed other ML models for deriving LAI estimates. Furthermore, USPI LAI exhibited better performance than MODIS LAI product, with bias, root mean square error (RMSE), and R2of -0.308, 0.593, and 0.826, respectively, especially for pixels contaminated by atmospheric conditions. Nevertheless, the underestimation of USPI LAI should be noted because the effective LAI provided by Sentinel-2 was involved in the GPR training process. Thus, it is necessary to introduce the accurate clumping index dataset for further improvement of USPI LAI retrievals. Our study indicates that incorporating the sub-pixel information from decametric-resolution data can effectively reduce the uncertainty of hectometric-resolution LAI retrieval, which is promising for generating the high-accuracy LAI time series dataset.

Leaf area index (LAI) is a pivotal biophysical parameter for characterizing canopy structure and monitoring vegetation growth. Although the look-up table (LUT) method has been widely employed for LAI retrieval, the optimization of key retrieval processes remains to be explored. Here, we proposed a generic optimization strategy for LUT-based inversion based on Landsat -8 imagery and global ground LAI measurements. Specifically, based on the LUT generated by the PROSAIL model, LAI inversion was optimized by introducing several functions, including band selection, artificial noise addition, cost function (CF) substitution, and multiple solutions. Furthermore, the optimized LUT-based inversion method was compared to the Simplified Level 2 Product Prototype Processor (SL2P) method and the ground-measurement-derived (GMD) regression method to comprehensively evaluate its performance over various vegetation types. Results showed that the combination of Red, near-infrared (NIR), and shortwave infrared-1 (SWIR1) bands was well suited to capture LAI dynamics. In terms of accuracy and efficiency, the best performance was achieved by the optimal band combination and retrieval parameter settings (i.e., root-mean-square error (RMSE) as CF, noise level of 20%, and multiple solutions of 5%), with the RMSE and${R} ^{2}$of 0.817 and 0.740, respectively. In addition, the optimized LUT-based inversion was superior to SL2P method in accuracy and to GMD regression method in efficiency. Overall, the optimized LUT-based inversion strategy can be applied for estimating decametric-resolution LAI with high accuracy over different regions and observation dates at a global scale, exhibiting high adaptability and generalization capability, especially for crops, and requiring no ground LAI measurements.

Leaf area index (LAI) is an important indicator for monitoring vegetation growth and estimating crop yields. The empirical-based model using vegetation indices (VIs) is an effective method for LAI estimation at the regional scale. However, due to the complexity of canopy radiation interaction processes, the leaf chlorophyll content (Cab) and saturation effects on canopy reflectance restrict the accuracy of VI-based LAI retrieval. To address these limitations, we propose a novel chlorophyll-insensitive vegetation index (CIVI) using red, red-edge and near-infrared bands to improve regional LAI mapping. The CIVI was developed based on the sensitivity analysis of red-edge band reflectance to LAI andCabusing the simulation dataset from the PROSAIL model. Then, the performance of CIVI was carefully evaluated from two aspects: the sensitivity of VI to LAI and other parameters, and the accuracy of LAI estimates using different VIs over homogeneous (cropland and grassland) and non-homogeneous (forest) biome canopies. The results suggested that CIVI can capture LAI variations well while remaining insensitive toCabvariations. Additionally, the sensitivity of CIVI to other vegetation biochemical and biophysical parameters did not increase significantly compared to that of other VIs. Furthermore, CIVI exhibited the best performance of LAI retrievals over both homogeneous (R2=0.938, RMSE=0.447 and rRMSE=21.3%) and non-homogenous (R2=0.635, RMSE=0.693 and rRMSE=14.0%) canopies among all selected VIs, especially for the high LAI. Our results indicated that the developed CIVI incorporating red-edge bands with a suitable formula can effectively reduce theCaband saturation effects, which is promising for improving VI-based LAI estimation.