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..
VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set
Authors: Shufan Shen, Junshu Sun, Qingming Huang, Shuhui Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments across multiple VLMs (e.g., CLIP, LLa VA) demonstrate the superior capability of VL-SAE in interpreting and enhancing the vision-language alignment. For interpretation, the alignment between vision and language representations can be understood by comparing their semantics with concepts. For enhancement, the alignment can be strengthened by aligning vision-language representations at the concept level, contributing to performance improvements in downstream tasks, including zero-shot image classification and hallucination elimination. |
| Researcher Affiliation | Academia | 1State Key Lab. of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences 2University of Chinese Academy of Sciences EMAIL EMAIL |
| Pseudocode | No | The paper describes methods using mathematical equations (e.g., Equation 1-9) and textual descriptions of the architecture and training process, but it does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured steps in a code-like format. |
| Open Source Code | Yes | Codes are available at https://github.com/ssfgunner/VL-SAE. |
| Open Datasets | Yes | We fed the CC3M dataset [45] containing 3 million image-text pairs into the VLM to extract the corresponding vision-language representations. ... We enhance the alignment mechanism of multiple CVLMs and conduct evaluations across various zero-shot image classification datasets [59, 25, 41, 6, 18, 36, 3, 19, 10, 39, 51, 7, 61]. ... Table 2 shows the performance comparisons on the POPE [30] benchmark. |
| Dataset Splits | Yes | These representations are randomly divided into training and test sets at a ratio of 4:1. |
| Hardware Specification | Yes | For the resource requirements, the VL-SAE models for different VLMs [43, 34, 1] are all trained using a single NVIDIA Ge Force RTX 4090 GPU. |
| Software Dependencies | No | The paper mentions the use of existing VLMs like Open CLIP, LLa VA, and Qwen-VL and discusses training precisions (full-precision, half-precision FP16), but it does not provide specific version numbers for any software libraries, frameworks, or programming languages (e.g., Python, PyTorch versions). |
| Experiment Setup | Yes | For LVLMs [34, 1], we first train the auxiliary autoencoder for 50 epochs to perform the explicit representation alignment. The training process is configured with a batch size of 2048, a weight decay of 0.01, and a learning rate of 5e-5. Then, the VL-SAE is trained for 10 epochs based on the intermediate representations of the autoencoder with a batch size of 512 and a learning rate of 1e-4. For CVLMs [43], we adopt the same strategy as the LVLMs but omit the training of the auxiliary autoencoder because CVLM representations are naturally aligned through cosine similarity. ... In practice, we set αcd and βcd to 0.6 and 0.8, respectively. Additionally, the values of α and β in Equation 11 are determined to be 0.7 and 0.9, respectively. |