Self-Evolving Vision-Language Models for Image Quality Assessment via Voting and Ranking¶

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

ICLR 2026; self-supervised voting and ranking evolves VLMs for image quality assessment

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

EvoQuality is a fully self-supervised post-training framework that enables a VLM (Qwen2.5-VL-7B) to improve its image quality assessment (IQA) capabilities without any human-annotated labels. It adapts the self-consistency principle to the ranking-based nature of IQA via pairwise majority voting to generate pseudo-labels, then uses those labels as fidelity rewards in GRPO-based policy optimization. The resulting model surpasses state-of-the-art supervised VLM-based IQA models on 5 of 7 benchmarks.

Problem¶

Supervised IQA model training requires costly large-scale subjective studies to obtain mean opinion scores (MOS), limiting dataset scale and diversity. Existing self-supervised self-consistency methods (e.g., majority-vote-based TTRL) are designed for discrete reasoning tasks and cannot be directly applied to IQA, which is inherently continuous, subjective, and where no absolute "correct" answer exists — only relative quality comparisons are reliable.

Method¶

EvoQuality operates as a two-stage iterative loop applied to an unlabeled image corpus:

Offline stage — Pairwise Pseudo-Label Generation: For $N$ images, random pairs $(x_i, x_j)$ are formed. The current VLM policy $\pi_\theta$ is queried $K=32$ times per pair using a binary comparison prompt, and majority voting produces pseudo-preference labels: $$p^*(x_i, x_j) = \begin{cases} 1 & \text{if } K_x > K_y \\ 0.5 & \text{if } K_x = K_y \\ 0 & \text{if } K_x < K_y \end{cases}$$ Positional bias is mitigated by randomly permuting image order.

Online stage — Policy Evolution via Fidelity Rewards: The VLM is prompted to predict absolute quality scores $q_k(x_i)$ from $K$ sampled reasoning trajectories. A predicted comparative probability is derived via the Thurstone model: $$p_k(x_i, x_j) = \Phi\!\left(\frac{q_k(x_i) - \mu(\mathbf{q}(x_j))}{\sqrt{\sigma^2(\mathbf{q}(x_i)) + \sigma^2(\mathbf{q}(x_j)) + \gamma}}\right)$$ The fidelity reward for response $k$ is: $$r_k(x_i) = \frac{1}{|P_i|}\sum_{j \in P_i}\!\left(\sqrt{p^*(x_i,x_j)\,p_k(x_i,x_j)} + \sqrt{(1-p^*)(1-p_k)}\right)$$ Relative advantages $a_k = (r_k - \mu(\mathbf{r}))/\sigma(\mathbf{r})$ are computed per group, and the policy is updated with clipped GRPO plus KL regularization against a frozen reference policy $\pi_\text{ref}$. Two iterations are run: Round 1 uses authentic KONIQ pairs; Round 2 adds synthetically distorted variants.

Key Contributions¶

Self-supervised IQA post-training: first framework to apply self-consistency/voting to the perceptual, continuous-valued domain of IQA without any MOS labels.
Voting-to-ranking pipeline: adapts majority voting to produce pairwise pseudo-labels that feed directly into a fidelity loss, bridging discrete voting and continuous ranking.
GRPO-based perceptual policy optimization: formulates fidelity rewards as RL signals for online VLM fine-tuning.
Iterative self-evolution: two-round loop that progressively bootstraps quality perception, with Round 2 leveraging synthetic distortions for broader coverage.
Composability: EvoQuality can be stacked on top of a pre-trained supervised IQA model (VisualQuality-R1+EvoQuality) for further generalization gains.

Results¶

EvoQuality@Round2 improves the Qwen2.5-VL-7B baseline by +31.8% PLCC and +33.7% SRCC on a weighted average across 8 benchmarks.
Per-dataset gains at Round 2 vs. baseline: KADID +28.9% PLCC, PIPAL +46.2%, TID2013 +77.4%, CSIQ +20.6%, LIVEW +18.0%.
Versus supervised VLM models (Table 3, single-source training on KONIQ):
Outperforms Q-Insight on weighted-average PLCC (0.762 vs. 0.704) and SRCC (0.719 vs. 0.667).
Surpasses all supervised models on TID2013 PLCC (0.674), exceeding NIQE (next best) by >0.15.
Competitive or superior on 5 of 7 OOD benchmarks despite using zero ground-truth labels.
Multi-source setting (Table 4, co-trained on KONIQ+SPAQ+KADID): VQ-R1+EvoQuality achieves WAVG PLCC 0.768 / SRCC 0.716, outperforming standalone DeQA-Score (0.748 / 0.700) and Q-Align (0.714 / 0.688).

Limitations¶

Training is restricted to KONIQ (authentic) + synthetic distortions; generalization to highly domain-specific content (e.g., medical images, satellite imagery) is untested.
Pseudo-label quality is bounded by the base VLM's initial pairwise discrimination ability — if the base model is poor at comparisons, the bootstrapping loop may converge to a wrong consensus.
Two rounds of GRPO training on 8× A100 GPUs require ~24 hours total; scaling to more iterations or larger VLMs is not evaluated.
No ablation on sensitivity to the voting budget $K$; $K=32$ is fixed without a cost-accuracy trade-off analysis reported.
Performance on in-distribution KONIQ is lower than supervised counterparts (PLCC 0.835 vs. 0.953 for DeQA-Score), reflecting the supervised advantage when labels are available.

Relevance to Vision-Language Models¶

EvoQuality demonstrates a practically important self-training recipe for VLMs in perceptual domains where discrete verifiable rewards do not exist, extending the GRPO/self-consistency paradigm (TTRL, MM-UPT) beyond reasoning tasks. The fidelity-reward construction via the Thurstone model is a transferable design pattern for any continuous subjective evaluation task (e.g., aesthetic scoring, text-image alignment rating) where ranking is more reliable than absolute scoring. The work establishes that VLMs carry sufficient latent perceptual knowledge to bootstrap their own quality judgment through internal consensus, without any external supervision — a finding relevant to low-annotation adaptation of general-purpose VLMs. It also offers a blueprint for composing self-supervised post-training with supervised pre-trained VLM heads, raising generalization above either approach alone.