Research map / Vision-language and multimodal / Multimodal LLMs

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

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 1 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.

Robust-R1 trains Qwen2.5-VL-3B with SFT on an 11K degradation-annotated A-OKVQA-derived reasoning dataset, then GRPO with a degradation-parameter reward and a length reward, to improve MLLM robustness to real-world visual degradations.

Problem

MLLMs degrade substantially under real-world visual corruptions, and existing robust MLLM methods rely on implicit encoder-level adaptation with limited interpretability and isolated optimization.

Approach

Makes degradation handling explicit via structured reasoning chains (degradation type and intensity, influence, pristine reasoning, conclusion), trained by SFT and then GRPO with a degradation-parameter reward and an intensity-scaled reasoning-length reward.

Key results

• On R-Bench the SFT+RL model reaches overall 0.5017 versus 0.4845 for Qwen2.5-VL-3B and 0.3718 for the best robust baseline (Robust CLIP).
• On MMMB at 100% intensity, SFT+RL scores 75.35 versus 74.50 for Qwen2.5-VL-3B.
• On MMStar at 100% intensity, SFT+RL and Qwen2.5-VL-3B are reported at 49.53 and 48.66; SFT alone at 49.53.

Surprising findings

• The reasoning-free fine-tuned variant (w/o Reasoning) underperforms the base Qwen2.5-VL-3B on R-Bench overall (0.4471 vs 0.4845), with a sharp drop at high intensity on MCQ.

Limitations

• Robust baselines (TeCoA, Robust CLIP, Robust LLaVA) are evaluated as reported in the paper's setup without detail on their reproduction.
• Training data relies on GPT-4o-generated reasoning annotations, with no human verification described in the main text.
• Compute and hardware details are not stated in the main text.

Experimental setup

Datasets: A-OKVQA R-Bench MMMB MMStar RealWorldQA

Benchmarks: R-Bench MMMB MMStar RealWorldQA

Metrics: accuracy captioning score

Baselines: Qwen2.5-VL-3B Gemma3-4B InternVL-4B TeCoA Robust CLIP Robust LLaVA

Builds on: GRPO Qwen2.5-VL A-OKVQA R-Bench

Closest papers by meaning

• Reinforcing Multimodal Reasoning Against Visual Degradation (NeurIPS 2026 · similarity 0.92)
• Robust-U1: Can MLLMs Self-Recover Corrupted Visual Content for Robust Understanding? (ICML 2026 · similarity 0.91)
• Predictive Regularization Against Visual Representation Degradation in Multimodal Large Language Models (CVPR 2026 · similarity 0.90)
• Scaling Test-Time Robustness of Vision-Language Models via Self-Critical Inference Framework (CVPR 2026 · similarity 0.89)
• On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs (ICML 2026 · similarity 0.89)
• Remember Me: Bridging the Long-Range Gap in LVLMs with Three-Step Inference-Only Decay Resilience Strategies (AAAI 2026 · similarity 0.89)