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..
FlexAC: Towards Flexible Control of Associative Reasoning in Multimodal Large Language Models
Authors: shengming yuan, Xinyu Lyu, Shuailong Wang, Beitao Chen, Jingkuan Song, Lianli Gao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate the effectiveness of Flex AC in controlling associative behavior, we conduct experiments across three fronts: hallucination mitigation (CHAIR [14] and POPE [15] for low-association tasks), creativity enhancement (VDAT and Creation-MMBench [16] for high-association tasks), and general-purpose evaluation (MME [17], MMMU [18], and MMStar [19]). Results show that Flex AC enables flexible modulation of associative reasoning capability, achieving state-of-the-art performance on both lowand high-association tasks while enhancing general capabilities. |
| Researcher Affiliation | Academia | 1University of Electronic Science and Technology of China 2Southwestern University of Finance and Economics, Chengdu, China 3Tongji University |
| Pseudocode | No | The paper describes methods and phases like "Phase I: Offline Control Vector Construction" and "Phase II: Inference-Time Control" with steps, but these are presented as descriptive text and mathematical equations, not as structured pseudocode blocks or algorithms. |
| Open Source Code | Yes | Our code is available at https://github.com/ylhz/Flex AC. |
| Open Datasets | Yes | To evaluate the effectiveness of Flex AC, we conduct experiments on three benchmark types: (1) hallucination, using CHAIR [14] and POPE [15] for low-association tasks), (2) creativity, using our proposed VDAT for associative reasoning and Creation MMBench [21] for open-ended image-grounded generation; and (3) general-purpose capability, using MME [17], MMMU [18] and MMStar [19]. From the COCO2014 [25] dataset, we randomly selected 2000 images and then applied Instance Selection to choose 50 images for generating the general association vector. |
| Dataset Splits | No | The paper mentions selecting 2000 images from COCO2014 and then 50 images for general association vector generation, but it does not specify explicit training, validation, or test splits for models like LLaVA-1.5, Qwen-VL, and Deepseek-VL2, nor how the other benchmarks (CHAIR, POPE, Creation-MMBench, MME, MMMU, MMStar) are split, only that they are used for evaluation. |
| Hardware Specification | Yes | All experiments were conducted on 8 RTX 4090 GPUs. |
| Software Dependencies | No | The paper mentions evaluating on LLaVA-1.5, Qwen VL, and Deepseek-VL, and using the VLMEval Kit, but it does not provide specific version numbers for these software components or any other libraries/frameworks used. |
| Experiment Setup | Yes | For the layer intervention, we manipulated the following layers based on each model s associative strength: Qwen-VL (layers 15, 16, 17), LLa VA-1.5 (layers 11, 12, 13), and Deepseek-VL (layers 4, 5, 6). For Flex AC-P (faithfulness-enhanced) and Flex ACC (creativity-enhanced), the control coefficient α is set to -1 and 1, respectively. |