Research map / Learning theory and algorithms / Combinatorial and search optimization
LDC-MTL: Balancing Multi-Task Learning through Scalable Loss Discrepancy Control
Peiyao Xiao, Chaosheng Dong, Shaofeng Zou, Kaiyi Ji
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace linear scalarization with bilevel loss discrepancy control | LDC-MTL · CelebA (40-task) | Δm%: 4.15 -> -1.31 | Improved |
| replace linear scalarization with bilevel loss discrepancy control | LDC-MTL · QM9 (11-task) | Δm%: 177.6 -> 49.5 | Improved |
| use LDC-MTL instead of the strongest baseline FairGrad | LDC-MTL · Cityscapes (2-task) | Δm%: 5.18 -> -0.57 | Improved |
| use LDC-MTL instead of GO4Align | LDC-MTL · NYU-v2 (3-task) | Δm%: -6.08 -> -4.40 | Hurt |
| loss discrepancy control reduces spread of task losses | LDC-MTL vs LS · CelebA (40-task) | mean task loss: 0.287 -> 0.189 (std 0.195 -> 0.133) | Improved |
| add dynamic loss discrepancy control instead of weight-swept LS | LDC-MTL · Cityscapes (2-task) | Δm%: 5.13 -> -0.57 | Improved |
| increase penalty constant lambda from 0.02 to 0.1 (ablation) | LDC-MTL · Cityscapes (2-task) | Δm%: 1.96 -> -0.57 | Improved |
| vary lambda among 0.005, 0.008, 0.01, 0.02 (ablation) | LDC-MTL · CelebA (40-task) | Δm%: -1.27, -1.16, -1.31, -0.96 | No clear effect |
| randomly reorder task losses before computing the upper-level gap (ablation) | LDC-MTL · CelebA (40-task) | Δm%: -1.31 -> -1.30, -1.00, -1.62 | No clear effect |
| use single-loop first-order update (drop the inner loop) for O(1) cost | LDC-MTL · training time | time vs LS: at most 1.11x increase vs LS | No clear effect |
| use O(1) LDC-MTL instead of O(K) gradient manipulation | LDC-MTL vs Nash-MTL · CelebA (40-task) | training time: approximately 12x less training time | Improved |
| use dynamic task weights whose final values are near equal rather than fixed LS | LDC-MTL · QM9 (11-task) | Δm%: 177.6 -> 49.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GO4Align — GO4Align: Group Optimization for Multi-Task Alignment (NeurIPS 2024) · better on 3 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.88)
- Revisiting Fairness in Multitask Learning: A Performance-Driven Approach for Variance Reduction (CVPR 2025 · similarity 0.87)
- MTLoRA: Low-Rank Adaptation Approach for Efficient Multi-Task Learning (CVPR 2024 · similarity 0.86)
- Delve into the Applicability of Advanced Optimizers for Multi-Task Learning (NeurIPS 2026 · similarity 0.86)
- NTKMTL: Mitigating Task Imbalance in Multi-Task Learning from Neural Tangent Kernel Perspective (NeurIPS 2025 · similarity 0.86)
- Robust Multi-Task Learning with Excess Risks (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
