arXiv:2508. 00955v3 Announce Type: replace-cross Abstract: Adapting generative Multimodal Large Language Models (MLLMs) into universal embedding models typically demands resource-intensive contrastive pre-training, while traditional hard negative mining methods suffer from severe false negative contamination.
arXiv:2607. 00784v1 Announce Type: cross Abstract: Vision-language pretraining remains dominated by contrastive objectives, whereas vision-only self-supervised learning has largely adopted non-contrastive methods.
Pretrained vision-language models such as CLIP excel at zero-shot recognition but often fail at compositionality, particularly attribute-object and relational structures. Recent studies mitigate this...
MLLMCLIP introduces a heterogeneous distillation framework that transfers multimodal knowledge from a generative Multimodal Large Language Model (MLLM) teacher directly into a discriminative CLIP student, eliminating the need for synthetic hard negatives. The method uses an attention-based per-layer token selection and a CKA-based distillation loss to bridge architectural differences between the two models. As a result, MLLMCLIP achieves state‑of‑the‑art compositional accuracy and improves zero‑shot classification and image‑text retrieval performance.
arXiv:2609.05730v1 Announce Type: cross Abstract: Contrastive Language-Image Pretraining (CLIP) is a building block of many machine learning applications. Scaling laws have guided resource allocation...
The paper introduces a post‑training approach that enables a single inference process to transition from text reasoning to image synthesis, eliminating the need for explicit modality switching. Using the 14B BAGEL model, the authors demonstrate that targeted post‑training data and reward‑weighted training improve multimodal image generation across four independent T2I benchmarks. The study highlights the benefits of joint text‑image generation and strategic data selection for enhancing T2I performance.
arXiv:2609.37631v1 Announce Type: new Abstract: Transformers are typically trained from random initialization, requiring all their capabilities to emerge from large-scale optimization. Recent work sh...
ViTAMINS is a method that incorporates synthetic hard negatives into unsupervised vision transformer pretraining to enhance representation quality. The approach is evaluated on ImageNet and a range of downstream tasks—including transfer learning, image retrieval, copy detection, and image/video segmentation—showing significant performance gains. The synthetic negatives also lead to emergent properties, such as representations that encode explicit semantic information and act as strong classifiers, improving over baselines by up to 11.3%.
The paper introduces a method to retrofit depth recurrence into pretrained vision‑language models, reducing the language backbone’s parameter count by pruning layers and looping over the remaining ones with linear adapters. A three‑stage recovery training—language‑only pretraining, vision‑language projector alignment, and full fine‑tuning—helps recover performance, with distillation from the non‑recurrent model further improving results. Experiments on TinyLLaVA‑1.4B show the retrofitted recurrent model outperforms the unpruned baseline while saving 23.3% of language parameters and performs better on reasoning‑oriented benchmarks.
The paper introduces V‑GIFT, a lightweight method that enhances visual instruction tuning for multimodal large language models by incorporating a small set of visually grounded self‑supervised tasks expressed as natural language instructions. These tasks—such as rotation prediction, color matching, and cross‑view correspondence—are reformulated into image‑instruction‑response triplets, providing supervision that cannot be solved without visual evidence. The approach requires no human annotations, architectural changes, or extra training stages, and consistently improves vision‑centric performance across various models, regimes, and benchmarks when only 3‑10% of the training data is augmented with these instructions.
arXiv:2610.26863v2 Announce Type: replace Abstract: Modern text-to-image (T2I) systems use internet-scale language pretraining for text-conditioned image generation. We ask whether language pretraini...
The paper introduces SCION, a single-model framework that jointly learns generative and self‑supervised representation tasks without relying on labels or pretrained encoders. SCION uses a pixel‑space encoder conditioned on a flow timestep and a shared embedding, training contrastive and flow‑matching objectives simultaneously with gradient‑norm balancing and stop‑gradient techniques. On ImageNet 256×256, SCION achieves state‑of‑the‑art FID scores (8.92 with JiT‑B and 5.89/3.47 with JiT‑L) surpassing prior class‑unconditional and pretrained‑conditioned methods.