DiBS-MTL Introduces Transformation-Invariant Multitask Learning

DiBS‑MTL stays robust to arbitrary loss rescaling and converges to Pareto‑stationary points even with nonconvex losses. The paper was submitted in September 2025. Read more: getnews.me/dibs-mtl-introduces-tran... #multitasklearning #dibsmtl

Study Shows Gradient-Based Fix for Multi-Task Learning Imbalance

Weighting each task’s loss by the inverse of its gradient norm matches exhaustive grid‑search performance without costly hyper‑parameter sweeps. Read more: getnews.me/study-shows-gradient-bas... #multitasklearning #gradientnorms

LDC-MTL: Scalable Loss Control for Balanced Multi-Task Learning

LDC‑MTL, a new loss discrepancy control method for multi‑task learning, maintains constant O(1) computational cost while balancing many tasks. The preprint appeared Sep 2025. getnews.me/ldc-mtl-scalable-loss-co... #ldcmtl #multitasklearning #ai

New Scheduler Reduces Gradient Interference in Multitask Learning

On 21 Sept 2025 researchers unveiled a scheduler that builds an interference graph and uses graph‑coloring to group tasks, yielding faster convergence on six benchmark datasets. getnews.me/new-scheduler-reduces-gr... #multitasklearning

Multi-Task Dense Prediction Fine-Tuning with Mixture of Fine-Grained
Experts
Duo Su, Xi Ye et al.
Paper
Details
#MultiTaskLearning #DensePrediction #FineTuning

Meta GenAI Boosts AI Learning with CGPO, Tackling Reward Hacking and Improving Multi-Task Performance Researchers at Meta GenAI introduced CGPO, a new post-training method for reinforcement learning that outperforms existing techniques by addressing reward hacking and optimizing multi-task learning. C...

🚀📊🤖 Meta GenAI Boosts AI Learning with CGPO, Tackling Reward Hacking and Improving Multi-Task Performance www.azoai.com/news/2024100... #AI #ReinforcementLearning #CGPO #MetaGenAI #RewardHacking #MultiTaskLearning #STEM #Coding #Optimization #LLM @arxiv-stat-ml.bsky.social