acceptodds
Withdrawn from ICLR 2027

EvoRobust: Self-Improving MLLMs Robustness via Claim-Level Self-Verification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal large language models lose accuracy and make more unsupported statements when images are blurred, noisy, or heavily compressed. The robustness gap is hard to close by self-improvement, because the model being improved is also the only verifier and it reads exactly the same corrupted pixels. We introduce EvoRobust, which raises MLLM robustness via claim-level self-verification: it decomposes a free-form description into atomic claims, turns each claim into a targeted probe, re-answers every probe on the same corrupted image several times, and aggregates the resulting agreement into a claim-structured reward for Group Relative Policy Optimization. We are explicit about what this signal is: it estimates the agreement rate of a frozen verifier rather than a truth probability, repeated probing reduces decoding variance rather than recovering lost visual evidence, and policy optimization only needs the reward to rank candidates correctly. Trained on unlabeled degraded images, EvoRobust raises the R-Bench Overall score of Qwen3-VL-4B from to , improves on averaged holistic self-scoring () under an identical optimizer, and is first or tied-first in seven of nine corrupted settings across MMMB, MMStar, and RealWorldQA without reducing clean accuracy. We will release the code and configurations shortly.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Self-Evolving, Robustness, Visual Understanding

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_DvKiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a framework, called EvoRobust, for improving MLLM robustness to visual corruption through claim-level self-verification. The framework consists of four roles: Fact Parser, Query Generator, Voting Critic, and Credit Assigner, then uses GRPO for policy optimization. The authors evaluate EvoRobust on R-Bench and three additional synthetic-corruption benchmarks, with reward ablations examining its effectiveness and the contributions of individual components.

Critical strengths

S1. The method requires only unlabeled degraded images for training, without clean counterparts, task labels, or an external teacher, and introduces no additional verification overhead at inference time.

S2. The problem setting is practically relevant: the method constructs a training signal through claim-level self-verification for visual corruption scenarios where clean counterparts and task labels are unavailable.

Critical weaknesses

W1. The performance gains are modest, and their statistical reliability remains unclear. On R-Bench, the Overall score improves by 0.0111 over the backbone and by 0.0067 over averaged holistic self-scoring. However, each configuration is evaluated in only a single run, without multiple seeds, confidence intervals, or significance tests, making it difficult to determine whether these small differences are consistent.

W2. Averaging over extracted claims can reward the addition of claims with high agreement, while omissions are not penalized. The practical impact of this composition bias is not evaluated. It therefore remains unclear to what extent the gains reflect improved visual understanding or a shift toward more conservative descriptions that are easier to verify.

Questions for authors

Please see the questions in Weaknesses

Reviewer_KWkkOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the robustness of multimodal large language models (MLLMs) on degraded images and proposes EvoRobust, a self-improvement method that does not rely on per-image human-provided answers, clean original images, or external teachers. The authors use frozen copies of the initial model to decompose generated captions into individual claims and generate corresponding verification questions. Repeated question answering and consistency judgments grounded in the image are then used to construct caption-level rewards for training the policy model with GRPO.

Critical strengths
  1. The research problem has practical value. Images in real-world applications often suffer from blur, noise, and compression. Improving MLLMs' image understanding under these conditions is a worthwhile research direction.
Critical weaknesses
  1. The reliability of the training supervision depends too heavily on the MLLM's own capabilities. The models responsible for generating and verifying captions originate from the same initial MLLM and see the same degraded image. If the model misunderstands the image, the verification process may repeat the same error and assign high rewards to incorrect captions. Although the paper discusses this risk, it does not use independent annotations to directly validate the reward's ability to distinguish correct from incorrect captions. The reliability of the method's central supervision signal therefore remains insufficiently validated.

  2. The training pipeline is overly complex and computationally expensive. The method requires decomposing captions, generating sub-questions, repeatedly answering each question, and judging consistency. The entire training process involves approximately 3.27 million logical calls and consumes 237.85 GPU-hours. In comparison, the R-Bench Overall score increases only from the base model's 0.6671 to 0.6782. Although this verification pipeline is used only during training, the reported gains do not sufficiently justify its complexity and computational cost.

  3. The method's core motivation lacks direct experimental support. Decomposing a caption into sub-questions and answering them repeatedly does not add visual information to the degraded image itself. The method needs evidence that this procedure produces more accurate judgments than the original caption and thereby provides effective training signals. Although the paper presents theoretical analysis and ablation experiments, it does not directly verify that judgments become more accurate after decomposition. It also does not adequately distinguish whether the gains come from improved visual understanding or from changes in caption content, such as including fewer details that are difficult to verify. The key link between sub-question verification and improved visual robustness therefore remains insufficiently validated.

  4. The method has limited novelty. In essence, it uses different prompts to make the same pretrained MLLM serve multiple frozen roles, forming a reward system for RL training. If the contribution amounts only to a way of providing rewards, it does not seem to meet the bar for acceptance at ICLR.

Questions for authors

See weaknesses.

Reviewer_wHcPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EVOROBUST, a self-improvement framework based on claim-level self-verification to enhance MLLM robustness under visual corruption. The method decomposes generated image descriptions into atomic claims, repeatedly verifies each claim, and uses the resulting agreement rates as rewards for GRPO training. This enables training without ground-truth labels or clean reference images. Experiments on R-Bench, MMMB, MMStar, and RealWorldQA demonstrate improvements over the original backbone and several baselines, while ablation studies evaluate the contributions of key components, including claim decomposition and repeated probing.

Critical strengths

Novel Reward Design: The paper proposes a claim-level self-verification reward that enables MLLMs to improve robustness without ground-truth labels or clean reference images. Decomposing descriptions into atomic claims provides more fine-grained feedback than holistic self-scoring.

Comprehensive Experimental Evaluation: The paper evaluates the proposed method across multiple benchmarks and corruption severities. The ablation studies further demonstrate the contributions of individual components, including claim decomposition and repeated probing.

Critical weaknesses

Does the proposed training actually improve the factual correctness of generated descriptions? An analysis of claim-level correctness and information coverage before and after training would help clarify whether the improvements reflect better visual understanding rather than the omission of difficult-to-verify claims.

Comparison with Robust-U1: A comparison with Robust-U1[1], which also addresses robustness under visual corruption, or a discussion of the differences between the two approaches would be valuable.

Evaluation on Larger Models: The current experiments are limited to 4B-scale models. It would be valuable to investigate whether the improvements generalize to larger models.

[1] Tang, Jiaqi, et al. "Robust-U1: Can MLLMs Self-Recover Corrupted Visual Content for Robust Understanding?." arXiv preprint arXiv:2606.08063 (2026).

Questions for authors

Although Appendix G provides an illustrative example of the verification pipeline, the paper lacks qualitative comparisons between the original and adapted policies. Could the authors provide examples showing how EVOROBUST changes generated descriptions, particularly whether it reduces hallucinations while preserving informative visual details?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.