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..
ACT as Human: Multimodal Large Language Model Data Annotation with Critical Thinking
Authors: Lequan Lin, Dai Shi, Andi Han, Feng Chen, Qiuzheng Chen, Jiawen Li, Zhaoyang Li, Jiyuan Zhang, Zhenbang Sun, Junbin Gao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through empirical studies, we derive 7 insights on how to enhance annotation quality while efficiently reducing the human cost, and then translate these findings into user-friendly guidelines. (3) We theoretically analyze how to modify the loss function so that models trained on ACT data achieve similar performance to those trained on fully human-annotated data. Our experiments show that the performance gap can be reduced to less than 2% on most benchmark datasets while saving up to 90% of human costs. |
| Researcher Affiliation | Collaboration | 1 University of Sydney, Australia 2 University of Cambridge, United Kingdom 3 Riken AIP, Japan 4 University of Adelaide, Australia 5 Byte Dance, Australia |
| Pseudocode | No | The paper describes mathematical formulations and processes (e.g., ACT Data Pipeline steps, sampling rules, loss functions) but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures. |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code for the methodology described, nor does it provide a direct link to a code repository. Mentions of URLs (e.g., for datasets or existing implementations like ResNet18 on Kaggle) do not constitute code release for this paper's work. |
| Open Datasets | Yes | Table 5: Dataset sources and licenses retrieved from https://paperswithcode.com/datasets. Dataset Source License CIFAR10 https://www.cs.toronto.edu/~kriz/cifar.html N/A Fashion-MNIST https://github.com/zalandoresearch/fashion-mnist MIT Stanford Cars https://paperswithcode.com/dataset/stanford-cars Custom (non-commercial) Tweet Eval-Emotion https://github.com/cardiffnlp/tweeteval N/A Tweet Eval-Irony https://github.com/cardiffnlp/tweeteval N/A VQA-RAD https://paperswithcode.com/dataset/vqa-rad CC0 1.0 Universal |
| Dataset Splits | Yes | Table 4: Dataset details. Dataset Type Description #Classes Size Train Size Test CIFAR10 CV Image classification of basic image categories. 10 50,000 10,000 Fashion-MNIST CV Image classification of cloth items. 10 60,000 10,000 Stanford Cars CV Image classification of car models. 196 8,143 8,040 Tweet Eval-Emotion NLP Text emotion classification. 4 3,257 1,421 Tweet Eval-Irony NLP Text irony detection. 2 2,862 784 VQA-RAD (Close-end) VQA Question-answer pairs on radiology images. 2 940 262 |
| Hardware Specification | Yes | A.2 Device & Random Seed All experiments are conducted with 1 to 8 NVIDIA SXM5 H100 GPUs with 80GB memories. |
| Software Dependencies | No | The paper mentions several models (ResNet18, RoBERTa-base, BLIP-VQA) and optimizers (Adam, SGD, AdamW) and libraries (XGBoost) but does not provide specific version numbers for any of these software components. |
| Experiment Setup | Yes | A.3 Downstream Training Sampling For the thresholding sampling rule, the threshold τ is determined by the quartile corresponding to the human budget proportion... Res Net18 For all datasets, we fine-tune the Res Net18 model initialized with Image Net-pretrained weights for 10 epochs. The batch size is set to 4096 for CIFAR-10 and Fashion-MNIST, and 32 for the Stanford Cars dataset... The key hyperparameters include the learning rate, with a search space of [1e-2, 1e-3, 5e-4, 1e-4], and the power-tuning parameter for ACT losses, with values selected from [0.6, 0.7, 0.8, 0.9, 1.0]... Ro BERTa For text classification tasks, we fine-tune the Ro BERTa-base model for 5 epochs. We set the batch size to 32 and use the Adam W optimizer. The key hyperparameters include the learning rate, with a search space of [1e-4, 5e-5, 2e-5, 1e-5], and the power-tuning parameter for ACT losses, with values selected from [0.6, 0.7, 0.8, 0.9, 1.0]... BLIP-VQA For VQA-RAD, we fine-tune the BLIP-VQA model initialized with the blip-vqa-base for 10 epochs... We use a batch size of 32 and optimize the model using the Adam W optimizer. The key hyperparameters include the learning rate, searched over [2e-4, 1e-4, 5e-5, 2e-5], and the power-tuning parameter for ACT losses, selected from [0.6, 0.7, 0.8, 0.9, 1.0]. |