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..
Compress & Cache: Vision token compression for efficient generation and retrieval
Authors: Adrian Bulat, Yassine Ouali, Georgios Tzimiropoulos
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | An in-depth ablation study confirms the efficacy of our approach. For generative tasks, we achieve a 2 higher compression rate without compromising capabilities, setting a new state-of-the-art. For discriminative tasks, we establish new state-of-the-art results on image retrieval and compositionality benchmarks. |
| Researcher Affiliation | Collaboration | Adrian Bulat 1,2 Yassine Ouali 1 Georgios Tzimiropoulos1,3 1Samsung AI Cambridge 2Technical University of Iasi 3QMUL |
| Pseudocode | No | The paper describes the 'Double forward bottleneck algorithm' in Section 3.2 using descriptive text and mathematical equations, but does not present it as a clearly labeled pseudocode or algorithm block. |
| Open Source Code | No | No code is provided alongside the submission, however all details are provided to allow for full reproducibility. |
| Open Datasets | Yes | At a given iteration, depending on the loss (i.e. autoregressive or discriminative), we sample a batch either from LLa VA-665K [34] (for generative) or from CC3M [48] (for discriminative). ... We evaluate our approach on a diverse collection of datasets, mainly: GQA [18], MMB [37], MME [32], POPE [31], SQA [39], Text VQA [50], Vis Wiz [11] and VQAv2 [10]. ... Discriminative benchmarks: We evaluate our model on a diverse set of retrieval benchmarks: Flicr30k [58], MS-COCO [33], No Caps [1] and Sugar Crepe [13]... We report results on all datasets from [60] (i.e. Arxiv QA [28], Chart QA [40], Doc VQA [54], Info VQA [41], Plot QA [42], Slide VQA [53]). |
| Dataset Splits | Yes | For the generative objective, we use the same collection of vision-language datasets encompassing 3.2M samples introduced in [23] and previously used to train the LLa VA-One Vision model that we start from. For the discriminative objective, to allow for fair comparisons, we use the same data as in [60]. Additionally, "To ensure fairness, in all cases, we fully align the test-time settings and processing with [34]." |
| Hardware Specification | Yes | The training runs were performed on 24 AMD MI300X GPUs using pytorch [43] and deepspeed [45]. Also, for latency measurements, "we benchmark the LLa VA-1.5 7B model on a RTX 4090 GPU." |
| Software Dependencies | No | The training runs were performed on 24 AMD MI300X GPUs using pytorch [43] and deepspeed [45]. The paper mentions software tools (pytorch, deepspeed) but does not provide specific version numbers for them. |
| Experiment Setup | Yes | Unless otherwise stated, the models are trained for 10,000 iterations, using a batch size of 1024, Adam W [38] with no weight decay and a learning rate of 2e 4 decayed to 0 using a cosine scheduler. All other layers remain frozen except for the Lo RA adapters (rank = 64, α = 128). ... The sampling ratio between the two is 1:1. |