Multi-Task Learning: Unifying Diverse AI Challenges from Clinic to City

Latest 50 papers on multi-task learning: Nov. 2, 2025

Multi-task learning (MTL) is rapidly evolving, moving beyond simple efficiency gains to become a cornerstone for building more robust, generalizable, and intelligent AI systems. By enabling models to learn multiple related tasks simultaneously, MTL fosters shared knowledge, reduces the need for vast labeled datasets, and inherently improves performance on individual tasks. Recent research showcases significant breakthroughs, pushing the boundaries of what MTL can achieve across a diverse array of applications, from medical diagnostics to urban planning and recommender systems.

The Big Idea(s) & Core Innovations

The overarching theme in recent MTL advancements is the pursuit of synergy and robustness in complex, real-world scenarios. Researchers are tackling challenges like data imbalance, task interference, and generalization limitations, often by drawing inspiration from human learning and biological systems.

One significant trend is the use of dynamic adaptation and uncertainty awareness. For instance, AW-EL-PINNs: A Multi-Task Learning Physics-Informed Neural Network for Euler-Lagrange Systems in Optimal Control Problems from Southwest University introduces adaptive loss weighting to dynamically balance state, control, and adjoint losses, improving accuracy and stability in optimal control problems. Similarly, EvidMTL: Evidential Multi-Task Learning for Uncertainty-Aware Semantic Surface Mapping from Monocular RGB Images integrates uncertainty estimation using evidential loss functions, enhancing the reliability of autonomous navigation systems from a University of Technology, Research Institute for Robotics, and AI Lab Inc. collaboration. Meanwhile, A Weighted Predict-and-Optimize Framework for Power System Operation Considering Varying Impacts of Uncertainty by Yingrui Z. et al. proposes a framework that adaptively weighs different sources of uncertainty in power system operations, increasing robustness.

Another critical innovation lies in mitigating task conflicts and fostering synergy. Direct Routing Gradient (DRGrad): A Personalized Information Surgery for Multi-Task Learning (MTL) Recommendations by Yuguang Liu, Yiyun Miao, and Luyao Xia from Whisper Bond Technologies Inc., Independent Researcher, and Tongji University introduces personalized gradient routing to dynamically adjust task stakes, effectively resolving the ‘seesaw problem’ in recommendation systems. Complementing this, NTKMTL: Mitigating Task Imbalance in Multi-Task Learning from Neural Tangent Kernel Perspective from Shanghai Jiao Tong University and Shanghai Innovation Institute uses Neural Tangent Kernel (NTK) theory to balance convergence speeds across tasks, addressing task imbalance. Shifting the paradigm, SyMerge: From Non-Interference to Synergistic Merging via Single-Layer Adaptation by Jung et al. from Sungkyunkwan University and NAVER AI Lab redefines model merging to foster active synergy rather than just avoiding interference, using a single adaptable layer and expert-guided self-labeling. This goal of synergy is further explored by AIM: Adaptive Intervention for Deep Multi-task Learning of Molecular Properties from ETH Zürich, which dynamically mediates gradient conflicts to improve data efficiency in molecular property prediction.

In the realm of data efficiency and generalization, several papers offer novel strategies. MeTA-LoRA: Data-Efficient Multi-Task Fine-Tuning for Large Language Models from Jilin University proposes a two-stage framework for LLMs that significantly reduces task-specific data requirements while promoting cross-task knowledge transfer. MTL-KD: Multi-Task Learning Via Knowledge Distillation for Generalizable Neural Vehicle Routing Solver by Zheng et al. from Shenzhen University and Southern University of Science and Technology, leverages knowledge distillation for label-free training of heavy decoder models, enhancing generalization for Vehicle Routing Problems. Addressing imbalanced data, Hurdle-IMDL: An Imbalanced Learning Framework for Infrared Rainfall Retrieval from the Chinese Academy of Meteorological Sciences and others, transforms the learning objective to unbiased ideals, tackling the long-tail issue in rainfall retrieval. Bias-Corrected Data Synthesis for Imbalanced Learning by Lyu et al. from Duke and Rutgers Universities introduces a bias-correction procedure to improve accuracy in imbalanced learning, even extending to multi-task scenarios.

Foundational models and advanced architectures are also making waves. Diffusion Transformers as Open-World Spatiotemporal Foundation Models by Yuan et al. from Tsinghua University and TsingRoc.ai introduces UrbanDiT, a pioneering open-world foundation model for urban spatio-temporal learning with zero-shot capabilities. Probabilistic Hyper-Graphs using Multiple Randomly Masked Autoencoders for Semi-supervised Multi-modal Multi-task Learning by Pˆırvu Mihai-Cristian and Marius Leordeanu unifies neural graphs and masked autoencoders to improve multi-modal, semi-supervised MTL, reducing training complexity by combining pre-training and fine-tuning. For industrial applications, LinkedIn Post Embeddings: Industrial Scale Embedding Generation and Usage across LinkedIn demonstrates how multi-task fine-tuning leads to superior post embeddings that power various LinkedIn products.

Under the Hood: Models, Datasets, & Benchmarks

The advancements in MTL are underpinned by innovative models, datasets, and rigorous benchmarks. Here’s a glimpse into the key resources driving this progress:

• Architectures & Frameworks:
  • M2H (https://github.com/UAV-Centre-ITC/M2H.git): A multi-task learning framework with efficient window-based cross-task attention for monocular spatial perception.
  • S4D (https://github.com/MSA-LMC/S4D): A dual-modal learning framework leveraging static facial expression data to improve dynamic facial expression recognition.
  • FALCON: A model combining AR-Bert, feature transfer, and multi-task learning for extracting spatio-temporal interactions from Wikipedia.
  • PatenTEB (patembed family) (https://github.com/iliass-y/patenteb): A model family for patent text embedding, developed by Iliass Ayaou and Denis Cavallucci from ICUBE Laboratory, INSA Strasbourg.
  • AortaDiff (https://github.com/yuxuanou623/AortaDiff.git): A unified multitask diffusion framework from the University of Oxford and others for contrast-free AAA imaging.
  • NMT-Net (https://github.com/MatZar01/Multi_Forecasting): A neuroplasticity-inspired dynamic ANN for multi-task demand forecasting by Mateusz Żarski and Sławomir Nowaczyk.
  • ControlAudio (https://control-audio.github.io/Control-Audio/): A progressive diffusion model for text-to-audio generation with precise timing and phoneme controls, from Tsinghua University and collaborators.
  • UrbanDiT (https://github.com/tsinghua-fib-lab/UrbanDiT): An open-world foundation model using diffusion transformers for urban spatio-temporal learning.
  • PHG-MAE (https://sites.google.com/view/dronescapes-dataset): Unifies neural graphs and masked autoencoders for multi-modal, multi-task learning, by Pˆırvu Mihai-Cristian and Marius Leordeanu from Politehnica University of Bucharest.
  • MTL-FSL (https://github.com/huatxxx/MTL-FSL): A framework by Huatxxx from UCSF that uses feature-similarity Laplacian graphs for predicting Alzheimer’s disease progression.
  • KERMT (accelerated version) (https://github.com/NVIDIA/BioNeMo): An accelerated implementation of a chemical pretrained graph neural network for drug property prediction, from Merck & Co., Inc. and NVIDIA.
  • M3ST-DTI (https://github.com/M3ST-DTI): A multi-task learning model for drug-target interactions based on multi-modal features and multi-stage alignment from the University of Science and Technology of China.
  • SpatialViLT & MaskedSpatialViLT: Enhanced vision-language models from Islam, A. et al. for visual spatial reasoning.
  • Fake-in-Facext (FiFa-MLLM) (https://github.com/lxq1000/Fake-in-Facext): The first model supporting artifact-grounding explanations and bounding box visual prompts for deepfake analysis, from Beijing University of Posts and Telecommunications.
  • AW-EL-PINNs: Developed by Chuandong Li and Runtian Zeng from Southwest University, this framework integrates Euler-Lagrange theorem with PINNs for optimal control problems.
  • ExpA & EARL (https://github.com/deepseek-ai/EARL): An expanded action space and RL algorithm for LLMs to reason beyond language, from Chalmers University of Technology and SAP.
  • GRADE: A personalized multi-task fusion framework via Group-relative Reinforcement Learning from Kuaishou Technology.
  • VAMO: A unified multi-task learning framework for generative auto-bidding with validation-aligned optimization by Lv et al. from Alibaba Group and Tsinghua University.
  • FW-Merging (https://github.com): A Frank-Wolfe optimization method for scalable data-free model merging from Imperial College London and Samsung AI Center.
  • RL-AUX (https://openreview.net/forum?id=vtVDI3w_BLL): Reinforcement Learning for Auxiliary Task Generation from Columbia University.
• Datasets & Benchmarks:
  • WikiInteraction dataset (https://anonymous.4open.science/r/FALCON-7EF9): A new dataset for extracting spatio-temporal human interactions from Wikipedia biographies, introduced by Zhongyang Liu et al. from ShanghaiTech University.
  • PatenTEB benchmark (https://github.com/iliass-y/patenteb): A 15-task benchmark for patent text embedding by Iliass Ayaou and Denis Cavallucci.
  • ETR-fr dataset: The first high-quality, paragraph-aligned dataset fully compliant with European ETR guidelines for easy-to-read text generation, from Université Caen Normandie and Koena SAS.
  • MSK-CHORD dataset: Utilized by OncoReason: Structuring Clinical Reasoning in LLMs for Robust and Interpretable Survival Prediction from New York University, Tandon School of Engineering, for cancer outcome prediction.
  • ADNI dataset: Used in Multi-Task Learning with Feature-Similarity Laplacian Graphs for Predicting Alzheimer’s Disease Progression to predict Alzheimer’s disease progression.
  • MODIS time-series dataset for Saskatchewan: Leveraged by Knowledge-Aware Mamba for Joint Change Detection and Classification from MODIS Times Series for land cover change detection.
  • MENLO dataset (https://huggingface.co/datasets/facebook/menlo): With 6,423 annotated prompt-response preference pairs across 47 language varieties, developed by Meta Superintelligence Labs for evaluating native-like LLM quality.
  • LIFULL HOME’S dataset: Utilized by Estimation of Fireproof Structure Class and Construction Year for Disaster Risk Assessment for facade image analysis and disaster risk assessment.
  • Dronescapes dataset (automated extension) (https://sites.google.com/view/dronescapes-dataset): Used in Probabilistic Hyper-Graphs using Multiple Randomly Masked Autoencoders for Semi-supervised Multi-modal Multi-task Learning.

Impact & The Road Ahead

The impact of these advancements resonates across numerous domains, promising more intelligent, adaptable, and reliable AI systems. In medical AI, models like OncoReason provide interpretable survival predictions by structuring clinical reasoning, while AortaDiff offers contrast-free AAA imaging, enhancing patient safety. Autonomous navigation in mechanical thrombectomy, as demonstrated by the world model-based approach from Kings College London, signifies a major step towards AI-driven robotic surgical precision.

For urban and environmental applications, UrbanDiT’s zero-shot spatio-temporal modeling and the probabilistic multi-task learning for wind-farm power prediction by Simon M. Brealy et al. from The University of Sheffield, along with Hurdle-IMDL for infrared rainfall retrieval, lay the groundwork for smarter cities and more efficient resource management. MTL for disaster risk assessment, as seen in the work on estimating fireproof structures from facade images, offers proactive solutions for urban planning.

In recommender systems and NLP, DRGrad’s personalized gradient routing tackles task conflicts at industrial scale, and LinkedIn’s post embeddings exemplify the real-world impact of multi-task fine-tuning for semantic understanding. MeTA-LoRA and ETR-fr’s contributions to data-efficient fine-tuning and easy-to-read text generation are crucial for democratizing access to powerful LLMs and enhancing cognitive accessibility. Furthermore, the ability of LLMs to reason beyond language with ExpA opens up new avenues for sophisticated AI agents interacting with real-world environments.

The future of multi-task learning is bright, characterized by a move towards more adaptive, synergistic, and explainable models. The continuous development of robust frameworks like FW-Merging for scaling model merging, and neuroplasticity-inspired dynamic ANNs in NMT-Net, points towards AI systems that can learn, adapt, and generalize with remarkable efficiency and insight. As researchers continue to refine methods for balancing task conflicts and fostering mutual enhancement, multi-task learning is set to unlock even greater potential, transforming how AI tackles the world’s most complex challenges. The insights from these papers suggest a future where AI systems are not just specialists but versatile generalists, capable of learning from and contributing to a rich tapestry of tasks simultaneously.