Adrian Bulat

reViT explores whether a single recurrent Transformer block can replace a full-depth vision encoder while preserving accuracy at comparable inference FLOPs. It shares attention across recurrent steps and constructs each step's feed-forward network by mixing a small shared bank of experts in weight space. A continuous normalized-depth coordinate controls the mixture, allowing different transformations at different depths. The paper evaluates supervised ImageNet-1k training and distillation from a DINOv2 teacher. A model with eight experts, trained using only the teacher's final output features, retains nearly all of the teacher's ImageNet linear-probe accuracy and transfers to classification, segmentation and depth prediction. Elastic-depth training enables one checkpoint to run at multiple tested depths. For deployment at a fixed depth, the recurrent block can also be expanded into a conventional dense graph, removing online routing and weight merging while increasing deployment storage.

FLaRe studies how language models can reason with continuous latent states and generate only their final answers. It combines a compact representation of symbolic reasoning, question-conditioned flow matching, and training on verified model-generated thoughts. The work evaluates whether these thoughts improve answers, support varied reasoning paths, admit faithful explanations, benefit from additional computation, and reduce inference cost. Experiments on arithmetic tasks examine the training choices needed to make latent reasoning effective.

UltraViT addresses the cost of visual encoding when deploying large vision-language models on edge devices. Its pyramidal encoder combines different spatial mixing operations, with architectural choices guided by measured device latency rather than computational proxies alone. A two-stage pre-training procedure first transfers detailed spatial representations through dense distillation and then applies generative supervision from a frozen language model trained with mixed capacity. This procedure develops the semantic grounding needed for subsequent multimodal alignment and is reported to outperform contrastive and self-supervised alternatives. Experiments show that the combination of device-aware architecture design and generative training improves the efficiency and performance of vision encoding for large vision-language models, exceeding encoder-focused baselines while achieving approximately 1.7 times their on-device speed.

This work introduces HIT, an approach to image super-resolution that adapts visual autoregressive modeling to produce several output resolutions in one forward pass. Conventional residual quantization does not ensure that intermediate token scales correspond to image scales, limiting earlier methods to a fixed output resolution. HIT instead builds a hierarchy of image tokens with overlap between scales, encouraging consistent representations at each resolution. Training also incorporates a direct preference optimization objective that favors the high-resolution target over its low-resolution input, using only the paired training images. Together, these changes support a 300-million-parameter model, compared with the billion-parameter VARSR baseline, without requiring external training data. The resulting system reports state-of-the-art super-resolution performance while reducing model size and providing outputs at multiple scales.