Resolving Token-Space Gradient Conflicts: Token Space Manipulation for Transformer-Based Multi-Task Learning

2025-07-10

AI Summary

To address gradient conflicts and negative transfer arising from divergent task objectives in Transformer-based multi-task learning (MTL), this paper proposes the Dynamic Token Modulation and Expansion framework for MTL (DTME-MTL). DTME-MTL is the first method to explicitly detect inter-task gradient conflicts in token space. It enables parameter-efficient, collaborative optimization across tasks within a shared encoder—without replicating parameters or architectural components—via lightweight, task-adaptive token modulation and sparse expansion mechanisms. Crucially, it requires no modification to the backbone architecture and incurs negligible computational overhead. Extensive experiments on standard MTL benchmarks—including GLUE and MTL-Benchmark—demonstrate consistent and significant improvements in overall performance and per-task robustness. These results validate DTME-MTL’s effectiveness, generalizability, and scalability in diverse multi-task settings.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Natural Language Processing: Language Grounding & Multi-modal NLP
• Multiagent Systems: Multiagent Learning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Multi-Task Learning (MTL) enables multiple tasks to be learned within a shared network, but differences in objectives across tasks can cause negative transfer, where the learning of one task degrades another task's performance. While pre-trained transformers significantly improve MTL performance, their fixed network capacity and rigid structure limit adaptability. Previous dynamic network architectures attempt to address this but are inefficient as they directly convert shared parameters into task-specific ones. We propose Dynamic Token Modulation and Expansion (DTME-MTL), a framework applicable to any transformer-based MTL architecture. DTME-MTL enhances adaptability and reduces overfitting by identifying gradient conflicts in token space and applying adaptive solutions based on conflict type. Unlike prior methods that mitigate negative transfer by duplicating network parameters, DTME-MTL operates entirely in token space, enabling efficient adaptation without excessive parameter growth. Extensive experiments demonstrate that DTME-MTL consistently improves multi-task performance with minimal computational overhead, offering a scalable and effective solution for enhancing transformer-based MTL models.

Problem

• Resolving gradient conflicts in token space for MTL
• Improving transformer adaptability without parameter duplication
• Enhancing multi-task performance with minimal computational overhead

Innovation

• Dynamic Token Modulation for gradient conflict resolution
• Token space adaptation without parameter duplication
• Scalable transformer-based multi-task learning enhancement