multi-task-learning

The paper investigates fine-tuning strategies for customer support LLMs, comparing multi-task training, sequential updates, and model merging across multiple model families. It concludes that multi-task full fine-tuning is the most robust default, while specialist models degrade off-task and require reliable routing.

This paper introduces the Lightweight Ranking Heads framework to accelerate multi-task experimentation in production recommender systems by enabling dynamic task injection without retraining backbone models, reducing iteration cycles from weeks to days at YouTube scale.

MT-ProtBERT is a multi-task learning model for classifying intrinsically disordered proteins under data scarcity, integrating self-supervised and biochemistry-informed tasks to outperform existing methods like PARROT.

The paper derives exact task-information–state frontiers for resource-sufficient learning, showing how limited advance task information reduces retained state through partitions with low-rank operators, with applications to attention and digital twin models.

This paper introduces AALT, a composition-aware active imitation learning method that organizes demonstrations into latent topologies to maximize start-goal connectivity, improving task success in structured multi-task robotic domains with minimal expert queries.

This paper presents the first comprehensive empirical study on multi-task learning for predictive process monitoring, showing improvements over single-task learning in next-activity prediction and class imbalance mitigation.

The paper introduces TAPe+MLv3, a compact computer vision system using structured representation for multi-task tasks, achieving competitive performance on benchmarks like COCO with fewer than 100,000 parameters.

This paper introduces a semantic-aware completeness estimation method for robust multimodal sentiment analysis with incomplete modalities, aiming to improve accuracy when sentiment-related cues are missing.

This paper presents a framework for funder name disambiguation in scientific publications using multi-functional embedding models trained with multi-task learning, achieving over 0.90 accuracy and outperforming general-purpose LLMs like GPT-5.2.

SHELF is a synthetic harness for evaluating LLM fitness on multi-task bibliographic benchmarks, addressing gaps in existing benchmarks like MTEB by providing controlled data and evaluation tasks for libraries and archives.

The paper introduces TPGC, a dual-prior prompt initialization method for multi-task graph pre-training that combines task and structural priors to improve alignment and transferability, achieving superior performance in few-shot scenarios.

The study demonstrates that different representation learning objectives recover distinct latent structures from psychometric data, with contrastive objectives enhancing teacher-child retrieval but PCA-based methods better preserving behavioral phenotype organization.

This paper proposes a multi-task learning approach that leverages turn-taking dynamics to enhance intent recognition in multi-party conversations, outperforming existing methods that ignore interaction patterns.

This paper introduces PromptKWS, a novel prompt-guided open-vocabulary keyword spotting framework that uses prompt embeddings and cross-attention to improve accuracy, achieving over 10% improvement in wakeup rate and over 15% in accuracy compared to baseline systems.

The paper investigates the geometric challenges in merging differentially private task models and introduces DP-Merging, a framework to enhance mergeability while maintaining privacy guarantees.

This paper proposes a unified framework for decentralized learning under communication constraints, introducing the DMFL-SQ algorithm that combines graph-based personalization, agnostic fairness, and compressed communication to achieve reduced communication while maintaining predictive performance and improving fairness across clients.

DeMMO is an interpretable framework for longitudinal, multi-disease, and multi-outcome learning of digital mobility outcomes using multi-task learning, achieving improved prediction performance over baselines on the Mobilise-D dataset.

MoTE introduces task-specific expert routing to replace dense decoder feed-forward networks in multi-task video understanding, improving accuracy and efficiency with interpretable, sparse computation.

This paper introduces SLAaaT, a method that enables AI agents to dynamically switch between specialized LoRA adapters during trajectories, improving composition of capabilities and outperforming baselines on synthetic coding tasks.

EMAN is an optimization-driven framework for multi-task learning that adaptively grows network capacity by monitoring optimization evidence, achieving improved performance on benchmarks like PASCAL-Context and NYUv2.