Tag
The paper proposes World Snapshot Models (WSMs), a unified framework reformulating nearly all visual tasks as generation tasks, validating a 'Unified Law for Visual Tasks' across more than 1,000 visual tasks akin to how LLMs unified NLP.
Iris-Z is a unified generative backbone that jointly models image appearance and depth, surpassing specialized monocular depth models at scale while retaining general image generation abilities, with two models released to the community.
The paper shows that sparsity-inducing regularization causes functional modularity to emerge in multitask neural networks, identifying compact task-selective subnetworks ('functional tokens') that can be recombined across pathways without retraining.
This paper evaluates small recurrent-depth language models (Ouro-1.4B/2.6B-Thinking, HRM-Text-1B) fine-tuned jointly across 13 clinical NLP datasets, finding they outperform or match similarly sized Qwen3 transformers on most tasks, especially inference and question answering.
The paper presents DPTM-DT, a dual-pretrained Transformer framework for multitask drug-target prediction that combines molecular graph embeddings, protein language-model embeddings, and physicochemical descriptors, achieving state-of-the-art performance on benchmark datasets.
This paper introduces the first benchmark for automatic lyric transcription in Greek songs, demonstrating that adapting Whisper through multitask learning and two-stage training achieves a 27.2% Word Error Rate, significantly improving over zero-shot baselines.
This paper introduces PopFS, a method for learning a shared feature set that is robust across heterogeneous populations, allowing each population to train its own model. Experiments on diverse tabular datasets and a 43-state COVID-19 nowcasting study show strong average and worst-population performance.
This paper argues that representation learning, not model-based planning, is the key to scalable multitask deep reinforcement learning. It introduces MR.Q, a simple model-free algorithm with auxiliary predictive objectives that outperforms prior world-model-based methods across diverse continuous control tasks.
TabPFN-MT extends PFNs to multitask in-context learning for tabular data, achieving state-of-the-art on small-to-medium datasets while reducing inference cost from O(T) to O(1) forward passes.