Research map / Vision-language and multimodal / Multimodal LLMs

Reasoning-Driven Multimodal LLM for Domain Generalization

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds DomainBed-Reasoning (GPT-4o reasoning chains for DomainBed images) and proposes RD-MLDG, which fine-tunes InternVL3-8B with a joint classification+reasoning loss (MTCT) plus iterative self-labeled reasoning regularization (SARR), reaching 86.89% average DomainBed accuracy.

Problem

Domain generalization methods rely on feature-level invariance; naively distilling reasoning chains into MLLMs for classification is hard to optimize and suffers from mismatch between large-model and self-generated reasoning patterns under domain shift.

Approach

Jointly train a direct classification prompt alongside reasoning-chain supervision (MTCT) so the classification loss anchors reasoning optimization, then iteratively replace GPT-4o chains with the model's own correct-conclusion chains (SARR) to reduce pattern mismatch.

Key results

• RD-MLDG on InternVL3-8B reaches 86.89% average DomainBed accuracy, above GPT-4o (83.46%) and DGCLDTP (83.19%).
• On TerraInc, MTCT plus SARR lifts InternVL3-8B from 46.84% to 70.65%, with the reasoning-only baseline at 64.56% under the ablation setting.
• On the FGVC-Aircraft base-to-new split, harmonic mean rises from 19.01% to 34.63% for InternVL3-8B.

Surprising findings

• Reasoning-chain SFT yields slightly lower accuracy than direct label SFT on TerraInc (-0.93 points), despite reasoning helping in zero-shot.
• Self-generated reasoning chains are easier to fit than GPT-4o chains, yet GPT-4o-chain SFT retains richer context.

Limitations

• Results depend on GPT-4o-generated reasoning chains, and DomainBed-Reasoning release is not stated.
• Baseline comparisons mix backbones (ResNet-50, ViT-B/16, MLLMs), making the comparison not controlled.
• Gains on PACS are small and PACS results are near ceiling.

Experimental setup

Datasets: PACS VLCS OfficeHome TerraInc VOLDOGER FGVC-Aircraft

Benchmarks: PACS VLCS OfficeHome TerraInc VOLDOGER VQA VOLDOGER VE FGVC-Aircraft base-to-new

Metrics: accuracy average accuracy across domains harmonic mean

Baselines: CORAL MLDG MixStyle DomainDrop SMOS RES GGA SWAD CLIP CoOp MaPLe SIMPLE+

Builds on: LLaVA-CoT DomainBed

Closest papers by meaning

• Multimodal Unsupervised Domain Generalization by Retrieving Across the Modality Gap (ICLR 2025 · similarity 0.90)
• Towards Generalizing to Unseen Domains with Few Labels (CVPR 2024 · similarity 0.89)
• PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization (CVPR 2024 · similarity 0.87)
• DreamPRM: Domain-reweighted Process Reward Model for Multimodal Reasoning (NeurIPS 2025 · similarity 0.87)
• Is Large-scale Pretraining the Secret to Good Domain Generalization? (ICLR 2025 · similarity 0.87)
• Activating Visual Context and Commonsense Reasoning Through Masked Prediction in VLMs (AAAI 2026 · similarity 0.87)