MED-DSLC: Multi-Expert-Domain Classification via Domain Supervision and Logit Calibration

2026-07-13 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, medium

Summary

MED-DSLC, a novel approach for Multi-Expert-Domain (MED) classification, addresses the challenge of logit miscalibration in Vision-Language Models (VLMs) like CLIP. While fine-tuning VLMs with techniques such as LoRA improves in-domain accuracy, it often degrades out-of-domain performance, creating a fragmented model ecosystem. The core issue arises when merging independently trained domain-specific LoRAs, as their logits lose global comparability, causing cross-domain interference and inflated confidence for out-of-domain classes. MED-DSLC tackles this by integrating domain supervised training with domain-wise logit scaling, explicitly restoring global logit comparability. This lightweight solution effectively preserves within-domain discrimination and minimizes cross-domain logit interference using minimal data. Extensive experiments on various fine-grained benchmarks demonstrate MED-DSLC's significant improvements, including a +15% increase in mean accuracy, enhanced cross-domain robustness, and better scalability for MED classification problems. The research highlights the critical role of output-level calibration for achieving truly zero-shot VLMs in multi-domain, data-imbalanced scenarios.

Key takeaway

For Machine Learning Engineers building multi-domain Vision-Language Model systems, you should consider integrating MED-DSLC to overcome logit miscalibration issues. This approach allows you to merge independently fine-tuned LoRAs without degrading out-of-domain accuracy, significantly improving mean accuracy by +15% and enhancing cross-domain robustness. Implementing domain supervised training and logit scaling will ensure global logit comparability, enabling more scalable and truly zero-shot VLM applications, especially in data-imbalanced environments.

Key insights

Logit miscalibration in merged VLM experts degrades multi-domain performance; MED-DSLC restores global comparability through domain supervision and logit scaling.

Principles

• Independent VLM expert training miscalibrates cross-domain logits.
• Global logit comparability is vital for multi-domain zero-shot recognition.
• Output-level calibration enables true zero-shot VLM specialization.

Method

MED-DSLC combines domain supervised training and domain-wise logit scaling. This explicitly restores global logit comparability across merged VLM experts, reducing cross-domain interference and improving accuracy with minimal data.

In practice

• Apply MED-DSLC to merge fine-tuned VLM LoRAs.
• Implement domain-wise logit scaling for cross-domain robustness.
• Use domain supervision to enhance multi-domain VLM accuracy.

Topics

• MED-DSLC
• Multi-Expert-Domain Classification
• Vision-Language Models
• Logit Calibration
• Domain Adaptation
• Zero-Shot Learning

Code references

• YyzHarry/multi-domain-imbalance
• sejong-rcv/Multi-Modal-Guided-Multi-Source-Domain-Adaptation-for-Object-Detection

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Harmonizing Multi-Objective LLM Unlearning via Unified Domain Representation and Bidirectional Logit Distillation — Multi-objective LLM unlearning requires harmonizing data representation and employing bidirectional distillation for balanced efficacy and robustness.
• A Systematic Benchmark of Intensity Normalisation Methods for 3D Knee MRI Segmentation and Cross-Domain Generalisability — Intensity normalization improves MRI segmentation generalizability, but domain shift remains a larger challenge for clinical deployment.
• Polyp-D2ATL: Deep Domain-Adaptive Transfer Learning for Colorectal Polyp Classification under Label Distribution Shift — Polyp-D2ATL uses deep domain-adaptive transfer learning to classify colorectal polyps, effectively mitigating label distribution shift and data imbalance.
• Exploring Dualistic Meta-Learning to Enhance Domain Generalization in Open Set Scenarios — MEDIC employs dualistic meta-learning and gradient matching to balance domain and class boundaries, enhancing open set domain generalization.
• Domain-Aware Hierarchical Contrastive Learning for Semi-Supervised Generalization Fault Diagnosis — DAHCL improves semi-supervised fault diagnosis by correcting pseudo-label bias and leveraging uncertain unlabeled data through domain-aware and hierarchical contrastive learning.
• PRiSM: Prototype Regularization for Few-Shot VLMs — Prototype bias from class imbalance severely degrades few-shot VLM performance; PRiSM rectifies this with novel regularization.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.