PaperlayerView exact revision on arXiv

Source-linked AI summary

Robust-R1: Degradation-Aware Reasoning for Robust Visual Understanding

Jiaqi Tang, Jianmin Chen, Wei Wei, Xiaogang Xu, Runtao Liu, Xiangyu Wu, Qipeng Xie, Jiafei Wu, Lei Zhang, Qifeng Chen

arXiv:2512.17532v1cs.CVcs.AI

TL;DR

MLLMs lose performance under real-world visual degradations, while existing robustness approaches are limited in interpretability and joint optimization. Robust-R1 explicitly reasons about degradation and its semantic effects, and reports state-of-the-art R-Bench performance alongside smaller degradation-related drops on three general visual benchmarks.

  • Problem

    Real-world visual degradations substantially reduce MLLM performance, while existing robustness approaches lack explicit degradation-impact diagnosis and neglect degradation propagation between the visual encoder and language model.

    Source: §1 · pp. 1–2

  • Method

    Robust-R1 explicitly perceives degradation type and intensity, analyzes semantic impact, and reconstructs distortion-free interpretations through structured reasoning chains.

    Source: §1 · p. 2

  • Results

    Robust-R1 achieves state-of-the-art performance across degradation intensities on R-Bench and smaller performance drops than all baselines under multi-level degradations on MMMB, MMStar, and RealWorldQA.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Robust-R1 is presented as a route toward more robust, interpretable, and efficient multimodal systems in visually challenging environments.

    Source: §6 · p. 7

Abstract

from arXiv · show

Multimodal Large Language Models struggle to maintain reliable performance under extreme real-world visual degradations, which impede their practical robustness. Existing robust MLLMs predominantly rely on implicit training/adaptation that focuses solely on visual encoder generalization, suffering from limited interpretability and isolated optimization. To overcome these limitations, we propose Robust-R1, a novel framework that explicitly models visual degradations through structured reasoning chains. Our approach integrates: (i) supervised fine-tuning for degradation-aware reasoning foundations, (ii) reward-driven alignment for accurately perceiving degradation parameters, and (iii) dynamic reasoning depth scaling adapted to degradation intensity. To facilitate this approach, we introduce a specialized 11K dataset featuring realistic degradations synthesized across four critical real-world visual processing stages, each annotated with structured chains connecting degradation parameters, perceptual influence, pristine semantic reasoning chain, and conclusion. Comprehensive evaluations demonstrate state-of-the-art robustness: Robust-R1 outperforms all general and robust baselines on the real-world degradation benchmark R-Bench, while maintaining superior anti-degradation performance under multi-intensity adversarial degradations on MMMB, MMStar, and RealWorldQA.

1 Introduction

Robust-R1 addresses the limited interpretability and isolated optimization of implicit robustness methods by explicitly reasoning about visual degradations. It combines degradation-aware training and reasoning with a dataset and evaluations spanning real-world and adversarial degradations.

1 Introduction

  • Existing implicit robustness approaches focus on visual-encoder distortions but lack explicit degradation-impact diagnosis and neglect degradation propagation between the encoder and language model..Source: §1 · pp. 1–2
  • Robust-R1 perceives degradation type and intensity, analyzes semantic impact, and reconstructs distortion-free interpretations to produce results with interpretable reasoning traces.Its implementation uses supervised fine-tuning for foundational reasoning, a reward for accurate degradation parameters, and a complementary reward to scale chain length with severity.Source: §1 · p. 2
  • The 11K dataset contains realistic degradations synthesized across acquisition, transmission, environment, and postprocessing, with structured chains linking degradation parameters and influence to pristine semantics and conclusions.It includes 10K training and 1K validation samples, with reasoning-chain complexity scaled to degradation intensity.Source: §1 · p. 2
  • Robust-R1 achieves state-of-the-art performance across low, medium, and high degradation intensities on R-Bench and smaller performance drops than all baselines under multi-level degradations on MMMB, MMStar, and RealWorldQA.The adversarial intensity levels are 25%, 50%, and 100%.Source: §1 · p. 2

2 Related Work

Robustness work has moved from encoder-focused adversarial training toward large-scale adversarial pre-training, while multimodal reasoning research has broadened beyond mathematical tasks. Robust-R1 applies intrinsic MLLM reasoning to visual distortions as an explicit route to robust understanding.

Robust Visual Understanding

  • Encoder-focused methods such as TeCoA, Sim-CLIP, and Robust CLIP optimize resilience to localized distortions, but reliance on limited adversarial datasets can compromise generalization.Robust LLaVA uses large-scale adversarial pre-training, though its computational and annotation costs limit scalability.Source: §2 · p. 2
  • Unlike implicit adaptation paradigms, Robust-R1 uses degradation-aware reasoning to improve interpretability alongside robustness..Source: §2 · p. 2

Multimodal Reasoning

  • Visual CoT and V* extend multimodal reasoning beyond mathematical visual tasks by parsing scene elements and their relations.Robust-R1 extends this reasoning line to explicitly reason about visual distortions.Source: §2 · p. 2

3 Methodology

Robust-R1 explicitly perceives degradation type, intensity, and impact, then reconstructs pristine semantics to guide robust outputs. Its workflow combines supervised fine-tuning, reward-based parameter alignment, and reasoning-length adaptation.

3 Methodology

  • Robust-R1 perceives degradation parameters and their impact, reconstructs pristine semantic representations, and conditions robust outputs on degradation-aware reasoning.Its workflow uses supervised fine-tuning, reward-based parameter alignment, and intensity-adapted reasoning length.Source: §3 · p. 3

3.1 Acquiring Basic Reasoning Ability

  • Structured tokens order degradation type, influence, pristine reasoning, and conclusion; supervised fine-tuning teaches the model to generate this chain sequentially.The task-specific answer token is optional during training and evaluation.Source: §3.1 · p. 3

3.2 Aligning Accurate Degradation Parameters

  • Because SFT alone leaves degradation-parameter perception inaccurate, the method rewards matching degradation types and matching intensity values.The reward penalizes type mismatches and rewards type matches in proportion to intensity accuracy.Source: §3.2 · pp. 3–4

3.3 Scaling to Suitable Reasoning Length

  • Higher degradation levels call for longer reasoning chains, while the length reward favors ground-truth optimal lengths and GRPO combines it with parameter alignment.The length reward decreases linearly with relative length discrepancy; the combined objective targets robustness and efficiency.Source: §3.3 · p. 4

4 Data Construction

The section constructs an 11K A-OKVQA-based dataset with realistic degradations and structured annotations. Its generation pipeline links degradation effects to pristine semantic reasoning and a conclusion, scaling chain length with degradation intensity.

  • The dataset contains 10K training and 1K validation samples, addressing the lack of explicit degradation and pristine-reasoning annotations in existing visual-understanding datasets.It is built on a subset of A-OKVQA.Source: §4 · p. 4
  • Degradations are synthesized across acquisition, transmission, environment, and postprocessing stages, with examples including lens blur, compression, noise, and watermark damage.Source: §4 · p. 4
  • The pipeline generates a degradation-influence description, recovers pristine semantic reasoning, and conditions the final conclusion on that reasoning and the ground-truth answer.The influence description provides supervision for the perception module, while pristine reasoning supports training the reconstruction module.Source: §4 · p. 5
  • Reasoning-chain length is dynamically scaled with total degradation intensity using intensity-calibrated prompt templates.The dataset-generation description links this scaling to reasoning efficiency and the length reward.Source: §4 · p. 5
  • The authors report that an inverse image-quality/degradation-intensity relation supports a real-world-like corruption distribution, and that lexical diversity supports modeling complex logical relationships.The resulting dataset supports SFT and subsequent GRPO optimization.Source: §4 · p. 5

5 Experiments

Robust-R1 is evaluated on real-world degradation and adversarial anti-degradation benchmarks, with ablations testing explicit reasoning and two reward components. The reported results show gains over baselines and identify how degradation-parameter accuracy and reasoning length relate to performance and efficiency.

5 Experiments

  • Robust-R1 uses Qwen2.5-VL-3B, with 25% of training data for SFT and 75% for RL; the vision encoder and visual projection layers remain frozen.The experiments assess R-Bench and adversarial degradation on MMMB, MMStar, and RealWorldQA.Source: §5 · pp. 5–6

5.1 Performance on R-Bench

  • On R-Bench, Robust-R1 with SFT and RL surpasses existing general and robust MLLM baselines in overall performance.Table 1 covers MCQ, VQA, and captioning across low, medium, and high degradation strengths.Source: Table 1 · p. 6, §5.1 · pp. 6–7

5.2 Anti-Degradation Performance

  • Across MMMB, MMStar, and RealWorldQA, Robust-R1 achieves state-of-the-art performance at all tested degradation levels.The evaluation applies degradations at 25%, 50%, and 100% intensity.Source: Table 2 · p. 6, §5.2 · p. 7
  • Figure 5 comparisons report fewer hallucinations and errors than the original baseline, while preference optimization balances inference efficiency with accurate degradation-parameter perception.The qualitative results concern visual understanding under degradation.Source: Figure 5 · p. 6, §5.2 · p. 7

5.3 Ablation Study

  • Removing degradation reasoning yields only marginal gains in some intensity ranges and fails at high intensity, whereas explicit reasoning improves robustness over adaptation-only and baseline models.The ablation compares reasoning-trained models with fine-tuning for adaptation alone.Source: §5.3 · p. 7
  • Adding rdeg improves R-Bench performance and reduces degradation-type misclassification and intensity-estimation bias on an out-of-domain test set.The reported analysis links rdeg with more precise degradation-parameter identification.Source: §5.3 · p. 7, Table 3 · p. 7, Figure 6 · p. 7
  • Adding rlen reduces average reasoning-chain length while maintaining performance, assigning longer chains to severe degradation and shorter ones to simpler cases.The passage reports that this adaptive allocation improves computational efficiency and overall performance.Source: §5.3 · p. 7, Figure 6 · p. 7, Table 3 · p. 7

6 Conclusion

The conclusion presents Robust-R1 as a framework using explicit degradation reasoning chains to improve multimodal understanding robustness. It identifies interpretability and efficiency as goals for multimodal systems in visually challenging environments.

6 Conclusion

  • Robust-R1 incorporates explicit degradation reasoning chains to enhance multimodal understanding robustness, with the aim of supporting more interpretable and efficient systems in visually challenging environments.The conclusion frames these as capabilities of systems the work seeks to build.Source: §6 · p. 7
Loading 2512.17532v1…