Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Adaptive Gradient Masking for Balancing ID and MLLM-based Representations in Recommendation
Authors: Yidong Wu, Siyuan Chen, Binrui Wu, Fan Li, Jiechao Gao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on offline and online A/B tests validate the effectiveness of our approach in mitigating convergence inconsistency and improving performance. |
| Researcher Affiliation | Collaboration | Yidong Wu Imperial College London London, UK Siyuan Chen University of Bristol Bristol, UK Binrui Wu Fudan University Shanghai, China Fan Li Kuai Shou Technology Beijing, China Jiechao Gao Stanford University Stanford, USA |
| Pseudocode | No | The paper describes the methodology using textual explanations and mathematical equations, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code and model are available at: AGM. |
| Open Datasets | Yes | We conduct offline experiments on four open-source datasets from diverse recommendation domains. First, we choose the Microlens dataset [60], which features user-item interactions, video introductions, and video cover images. In addition, we adopt three categories from the Amazon dataset Baby, Sports, and Electronics [61, 62] which contain user-item interactions, product descriptions, and images. |
| Dataset Splits | No | All raw datasets are preprocessed with a 5-core setting on both items and users, as described in [12, 10]. The paper mentions preprocessing, but does not provide specific train/test/validation splits with percentages or sample counts for the main experiments. |
| Hardware Specification | Yes | In the fine-tuning phase, we adopted Qwen2vl-2b[51] as the backbone model. The model was fine-tuned for 5 epochs across four distinct datasets, utilizing a batch size of 128 on 4 A100 GPUs. For AGM, offline evaluations were conducted using TensorFlow 2.15.0 on a single RTX 4090 GPU, selecting Adam as the optimizer. |
| Software Dependencies | Yes | For AGM, offline evaluations were conducted using Tensor Flow 2.15.0 on a single RTX 4090 GPU, selecting Adam as the optimizer. |
| Experiment Setup | Yes | The model was fine-tuned for 5 epochs across four distinct datasets, utilizing a batch size of 128 on 4 A100 GPUs. Hyperparameters, including batch size and learning rate, were systematically tuned across candidate sets of {256, 512, 1024, 2048} and {1e 3, 1e 4, 1e 5}, respectively. The best model was selected based on the minimum validation loss, and early stopping was applied with a patience of 5 to prevent over-fitting. |