Reasoning-Driven Multimodal LLM for Domain Generalization

Conference: ICLR 2026
arXiv: 2602.23777
Unit: Xidian University / Microsoft Research Asia
Area: Domain Generalization / Multimodal Reasoning

TL;DR

Ours proposes RD-MLDG: the first framework to introduce Multimodal Large Language Model (MLLM) reasoning chains into Domain Generalization (DG). By constructing the DomainBed-Reasoning dataset, the study systematically analyzes two major challenges in reasoning supervision (optimization difficulty and reasoning mode mismatch). These are addressed through the synergy of Multi-Task Cross Training (MTCT) and Self-Aligned Reasoning Regularization (SARR). On four standard DG benchmarks, it achieves an average accuracy of 86.89%, significantly outperforming GPT-4o (83.46%) and all CLIP/ViT-based methods.

Background & Motivation

Existing DG methods (IRM, CORAL, MixStyle, SWAD, etc.) focus on feature-level invariance—improving generalization by aligning latent representations across different domains. CLIP-based methods introduce multimodal representations but remain limited to feature-level alignment. The problem is that feature-level invariance fails to capture higher-level cross-domain commonalities.

MLLMs exhibit powerful reasoning capabilities; reasoning chains can explicitly decompose the classification process into interpretable, domain-invariant steps (e.g., although "printer" images from different domains differ greatly visually, the category-related parts of their reasoning chains are highly consistent). However, directly using reasoning chains for supervision results in performance worse than direct label supervision—this contradiction motivates the deep analysis and design of the new framework.

Method

Overall Architecture

RD-MLDG aims to elevate DG from "feature-level invariance" to "reasoning process-level invariance": even if visual differences for the same category across domains are large, the category-related reasoning criteria remain highly consistent. Therefore, rather than aligning ambiguous features, it is better to let the model learn a set of domain-invariant reasoning processes. The pipeline is as follows: first, GPT-4o is used to create the DomainBed-Reasoning dataset for all samples in DomainBed. A diagnosis reveals that "direct supervision with reasoning chains is worse than label supervision" due to two issues: the optimization gap and reasoning mode mismatch. Subsequently, an InternVL3-8B is fine-tuned in two stages: MTCT uses the classification path as an anchor to pull the reasoning path across the optimization gap, and SARR replaces GPT-4o-style supervision with self-generated chains to eliminate mode mismatch, eventually yielding a reasoning-driven DG classifier.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["DomainBed 4 Benchmarks<br/>PACS / VLCS / OfficeHome / TerraInc"] --> B["DomainBed-Reasoning Dataset<br/>GPT-4o 5-stage Reasoning Chain<br/>No labels + REFLECTION + Rejection Sampling"]
    B --> C["Diagnosis: Direct Chain Supervision is Worse<br/>① Optimization Gap ② Reasoning Mode Mismatch"]
    C -->|"Challenge ①: Hard Convergence"| D["MTCT (Multi-Task Cross Training)<br/>CLS Path as Anchor + Joint Reasoning Optimization"]
    D -->|"Challenge ②: GPT-4o Style Mismatch"| E["SARR (Self-Aligned Reasoning Regularization)<br/>Self-Gen Chain Replacement + Keep Correct Conclusion Only<br/>Rerun MTCT, Iterative N=3"]
    E --> F["DG Classifier<br/>InternVL3-8B + LoRA"]

Key Designs

1. DomainBed-Reasoning Dataset: Upgrading Supervision from Labels to Interpretable Reasoning

The "reasoning process invariance" hypothesis requires data that can carry it. Since standard DG benchmarks only provide images and category labels, GPT-4o is used to generate a five-stage reasoning chain for each sample across PACS/VLCS/OfficeHome/TerraInc: SUMMARY → CAPTION → REASONING → REFLECTION → CONCLUSION. Three details ensure quality: no ground-truth labels are provided during generation to force reasoning to be based entirely on visual evidence and prevent answer leakage; an additional REFLECTION stage is inserted to reduce invalid generation and improve stability; finally, rejection sampling is performed to keep only chains that contain all components and reach a self-consistent conclusion. The category-related steps in these chains are naturally consistent across domains, serving as the carrier for process-level invariance and the supervision source for subsequent training.

2. MTCT (Multi-Task Cross Training): Using Classification Path as an Anchor to Bridge the Optimization Gap

Directly training with the chains is problematic—this is "Challenge I: Optimization Gap." On TerraInc, zero-shot with reasoning chains can increase ground-truth token probability by +43.28%p, but during SFT, reasoning supervision performs 0.93%p lower than direct label supervision because reasoning SFT only increases token probability by +1.88%p (compared to +43.38%p for direct label SFT). MTCT addresses this by constructing two prompts for the same image: a no-thinking classification path that directly predicts label \(y_i\), providing a stable signal, and a reasoning path that inputs the full chain \(\mathbf{r}_i\), providing rich semantics. The combined optimization is:

\[\mathcal{L}_{\text{MTCT}}=\frac{1}{B}\sum_{i=1}^{B}\Big[-\log p_{\theta}(y_i\mid\mathbf{x}_i,\mathbf{q}_{\text{cls}})-\frac{1}{T_i}\sum_{t=1}^{T_i}\log p_{\theta}(r_{i,t}\mid\mathbf{r}_{i,<t},\mathbf{x}_i,\mathbf{q}_{\text{reason}})\Big]\]

The reasoning chain loss is normalized by token length \(T_i\) to prevent long chains from dominating the gradient. The classification path acts as an "anchor," guiding the reasoning optimization and preventing overfitting on long sequences without learning high-confidence tokens.

3. SARR (Self-Aligned Reasoning Regularization): Gradually Replacing GPT-4o Style with the Model's Own Style

MTCT enables training, but "Challenge II: Reasoning Mode Mismatch" remains: GPT-4o chains are full of descriptive context (background, perspective), leading to only a +1.88%p token probability increase after SFT. Conversely, SFT with self-generated chains yields a +29.74%p increase but lacks informational depth. SARR allows the model to generate its own chains after MTCT, keeping only those where the <CONCLUSION> matches ground-truth labels \(\hat{\mathbf{r}}_i\) as refined supervision. The model is then re-tuned using the same \(\mathcal{L}_{\text{SARR}}\) target for \(N\) iterations (typically \(N=3\)). Each round replaces part of the hard-to-optimize GPT-4o descriptions with the model's own style, balancing semantic richness and optimizability.

Loss & Training

Both stages share the \(\mathcal{L}_{\text{MTCT}}\) form of visual-classification plus normalized reasoning chain loss. SARR rounds simply replace the supervision chain from GPT-4o's version to the self-generated chain \(\hat{\mathbf{r}}_i\). InternVL3-8B is used as the base with LoRA rank 8 applied to both vision encoders and language decoders. Each stage runs for 3 epochs with a batch size of 128, a learning rate of 5e-4, using the AdamW optimizer on 4× A100 80GB.

Key Experimental Results

Main Results: SOTA Comparison on DomainBed Benchmarks

Method	Backbone	PACS	VLCS	OfficeHome	TerraInc	Average
CORAL	ResNet-50	86.20	78.80	68.70	47.60	70.33
SMOS	ResNet-50	89.40	79.80	71.60	55.40	74.05
SWAD	ViT-B/16	91.30	79.40	76.90	45.40	73.25
CLIP	ViT-B/16	96.20	81.70	82.00	33.40	73.33
SIMPLE+	ViT-B/16	99.00	82.70	87.70	59.00	82.10
CLIP-LoRA	ViT-B/16	97.10	83.10	83.90	55.70	79.95
DGCLDTP	ViT-B/16	97.03	84.79	87.65	63.27	83.19
GPT-4o	MLLM	97.83	85.41	90.12	60.49	83.46
InternVL3-8B	MLLM	96.26	85.67	85.10	46.84	78.47
RD-MLDG	MLLM	98.13	87.03	91.73	70.65	86.89

RD-MLDG achieves an average accuracy of 86.89%, surpassing GPT-4o by +3.43%p and the strongest CLIP method DGCLDTP by +3.70%p. The improvement on TerraInc is particularly striking—rising from 46.84% (InternVL3-8B) to 70.65% (+23.81%p), even exceeding GPT-4o by +10.16%p.

Ablation Study (InternVL3-8B, OfficeHome / TerraInc)

Config	OfficeHome	\(\Delta\)	TerraInc	\(\Delta\)
Zero-shot	85.10	—	46.84	—
+ CLS only	89.39	—	66.69	—
+ Reasoning only (Baseline)	88.76	—	64.56	—
+ MTCT	90.58	+1.81	67.19	+2.63
+ SARR	90.91	+2.14	65.29	+0.73
+ MTCT + SARR (Full)	91.73	+2.97	70.65	+6.09

Key Findings: (1) Using reasoning chains alone (Reasoning only) is lower than direct classification (CLS only) by 0.63%p / 2.13%p, validating Challenge I; (2) MTCT alone provides significant gains; (3) The combined effect of MTCT + SARR far exceeds individual usage, especially on TerraInc with a +6.09%p boost.

Key Findings

• SARR Iteration Analysis: On TerraInc, accuracy reaches 70.06% at \(N=1\), 70.59% at \(N=2\) (significant at \(p<0.01\)), and 70.65% at \(N=3\) (not significantly different from \(N=2\)). Token probability distributions also converge within 2-3 rounds.
• MTCT Token Analysis: After MTCT, the proportion of high-confidence category tokens (>0.75) increases from 86.33% to 90.23%, while the low-confidence proportion (<0.25) drops from 7.59% to 3.19%. Although 19.33% of GPT-4o chain tokens remain in the low-probability zone (semantic details being hard to fit), critical classification tokens are significantly enhanced.

Highlights & Insights

Highlights: * A New Perspective on "Process-Level Invariance": Elevates DG from feature-level invariance to reasoning process-level invariance; category-related steps in reasoning chains are naturally consistent across domains. * Problem-Driven Design: The systematic analysis of the two challenges (optimization difficulty and mode mismatch) leads to targeted designs (MTCT and SARR), forming a rigorous logic. * Exceptional Performance on TerraInc: Improvements of +23.81%p over the base model and +10.16%p over GPT-4o suggest that reasoning chains are particularly effective for scenarios with dramatic domain shifts.

Limitations: * Dependence on GPT-4o for initial reasoning chains entails high data construction costs. * Verification is limited to classification tasks; it is unclear if reasoning-driven DG generalizes to detection or segmentation. * Training overhead (4× A100) presents a barrier for widespread reproduction.

Rating

• Novelty: ⭐⭐⭐⭐⭐ First reasoning-driven DG framework + DomainBed-Reasoning dataset.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ 4 DG benchmarks + dual-model ablation + token-level analysis + parameter sensitivity.
• Writing Quality: ⭐⭐⭐⭐⭐ Complete logical loop from challenge discovery to analysis to method and validation.
• Value: ⭐⭐⭐⭐⭐ Opens a new reasoning-driven paradigm for Domain Generalization.

Related Papers

• [ICLR 2026] Vid-LLM: A Compact Video-based 3D Multimodal LLM with Reconstruction–Reasoning Synergy
• [ICLR 2026] JUDO: A Juxtaposed Domain-Oriented Multimodal Reasoner for Industrial Anomaly QA
• [ICML 2026] LIMSSR: LLM-Driven Sequence-to-Score Reasoning under Training-Time Incomplete Multimodal Observations
• [ICLR 2026] VTool-R1: VLMs Learn to Think with Images via Reinforcement Learning on Multimodal Tool Use
• [ICLR 2026] VLM-SubtleBench: How Far Are VLMs from Human-Level Subtle Comparative Reasoning?