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..
Conditional Representation Learning for Customized Tasks
Authors: Honglin Liu, Chao Sun, Peng Hu, Yunfan Li, Xi Peng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on classification and retrieval tasks demonstrate the superiority and generality of the proposed CRL. |
| Researcher Affiliation | Academia | Honglin Liu1, Chao Sun2, Peng Hu1, Yunfan Li1 , Xi Peng1,3 1College of Computer Science, Sichuan University, Chengdu, China 2Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China 3National Key Laboratory of Fundamental Algorithms and Models for Engineering Numerical Simulation, Sichuan University, Chengdu, China EMAIL {sunchao}@aircas.ac.cn |
| Pseudocode | Yes | The complete process of our CRL is outlined in Algorithm 1. To deliver a more intuitive understanding of CRL s working mechanism and underlying rationale, we provide an example about learning a color-conditioned representation as illustrated in Fig. 2. Algorithm 1 Conditional Representation Learning (CRL) Input: Criterion C, LLM Prompt P1, VLM Prompt P2, Images X Output: Transformed Conditional Representation R 1: Query an LLM to generate the descriptive texts W related to the user-specified criterion C via Eq.(1). 2: Compute the text basis T via Eq.(2). 3: Compute the original universal image representation I via Eq.(3). 4: Transform I into conditional representation R via Eq.(4), which could be then utilized for various customized tasks. |
| Open Source Code | Yes | Extensive experiments on classification and retrieval tasks demonstrate the superiority and generality of the proposed CRL. The code is available at XLearning-SCU/2025-NeurIPS-CRL. |
| Open Datasets | Yes | For this task, we utilize Clevr4-10k [56] and Cards [67] as benchmark datasets. ... We adopt Gene CIS[55] as the benchmark for this task... Following previous works [39], we use the category and attribute prediction benchmark of Deep Fashion [36] as the evaluation dataset for this task |
| Dataset Splits | Yes | Dataset. Following previous works [39], we use the category and attribute prediction benchmark of Deep Fashion [36] as the evaluation dataset for this task, which consists of 221k / 27k / 27k images for training / validating / testing. ... For fair comparisons, we adopt the logistic regression function from the scikit-learn package [45] to perform few-shot learning, under the number of shots 1, 5, 10 per class, respectively. To alleviate the influence of randomness, we stochastically select the training data 20 times for each shot and report the mean result. |
| Hardware Specification | Yes | Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: All experiments were conducted on a single Nvidia RTX 3090 GPU. |
| Software Dependencies | No | For fair comparisons, we adopt the logistic regression function from the scikit-learn package [45] to perform few-shot learning, under the number of shots 1, 5, 10 per class, respectively. |
| Experiment Setup | Yes | The training process consists of two stages. In the first stage, we only train the MLP and freeze the CLIP model for 1000 epochs, with an initial learning rate of 1e-4. In the second stage, we freeze the MLP and slightly fine-tune the CLIP model for 100 epochs, with a smaller initial learning rate of 1e-6. The optimizer, the decaying rate, the decaying step size and the triplet margin are set to Adam, 0.9, 3 and 0.3, respectively. |