Vision-Language Model (VLM)-based image quality assessment (IQA) has been significantly advanced by incorporating Chain-of-Thought (CoT) reasoning. Recent work has refined image quality reasoning by applying reinforcement learning (RL) and leveraging active visual tools. However, such strategies are typically language-centric, with visual information being treated as static preconditions. Quality-related visual cues often cannot be abstracted into text in extenso due to the gap between discrete textual tokens and quality perception space, which in turn restricts the reasoning effectiveness for visually intensive IQA tasks. In this paper, we revisit this by asking the question, "Is natural language the ideal space for quality reasoning?" and, as a consequence, we propose Q-Tacit, a new paradigm that elicits VLMs to reason beyond natural language in the latent quality space. Our approach follows a synergistic two-stage process: (i) injecting structural visual quality priors into the latent space, and (ii) calibrating latent reasoning trajectories to improve quality assessment ability. Extensive experiments demonstrate that Q-Tacit can effectively perform quality reasoning with significantly fewer tokens than previous reasoning-based methods, while achieving strong overall performance. This paper validates the proposition that language is not the only compact representation suitable for visual quality, opening possibilities for further exploration of effective latent reasoning paradigms for IQA. Source code will be released to support future research.
Figures & tables
Figure 1: Motivation and overview of Q-Tacit. We compare Q-Tacit and Q-Insight [ 25 ] , which perform quality reasoning in a latent space and a text space, respectively. <|lvr_start|> and <|lvr_end|> encapsulate the latent reasoning process, and <|lvr|> is a placeholder for a latent slot token that allocates one latent reasoning step. Q-Tacit excels at image quality scoring on out-of-distribution datasets (Right) while requiring only 10% visible token counts compared to Q-Insight (Left).
Figure 2: Overview of Q-Tacit. We introduce a compact latent quality-reasoning segment wrapped by special tokens <|lvr_start|> and <|lvr_end|> , where the model propagates latent embeddings (i.e., internal hidden states) as a compact quality space instead of generating textual rationales. Then directly outputs the final quality score in <answer> and </answer> when a stopping criterion is met.
Figure 3: Example training data for Q-Tacit. The left part shows the mixture for constructing the latent space in Stage I. The Bboxes field specifies the ROI to supervise latent reconstruction. <lvr> functions as a placeholder to indicate where the latent segment should be inserted, which will expand to a series of latent span <|lvr_start|><|lvr|>...<|lvr|><|lvr_end|> in practice. The right part is quality-aligned calibration for latent quality reasoning in Stage II.
Methods
KonIQ-10K
SPAQ
KADID
PIPAL
LIVEW
AGIQA
CSIQ
AVG.
Handcrafted
NIQE [ 35 ]
0.533
0.679
0.468
0.195
0.493
0.560
0.718
0.521
(SPL 2012)
/ 0.530
/ 0.664
/ 0.405
/ 0.161
/ 0.449
/ 0.533
/ 0.628
/ 0.481
BRISQUE [ 34 ]
0.225
0.490
0.429
0.267
0.361
0.541
0.740
0.436
(TIP 2012)
/ 0.226
/ 0.406
/ 0.356
/ 0.232
/ 0.313
/ 0.497
/ 0.556
/ 0.369
Non-VLM Deep-learning
Table 1: PLCC / SRCC results of NR-IQA models trained on KonIQ-10K . Exceptions marked with † use a multi-task training set in [ 25 ] for GRPO calibration. Top two results are highlighted in bold and underlined , respectively.
Figure 4: Token-space coupling and attention weight distribution during quality scoring: Q-Insight vs. Q-Tacit. Left/Right: t-SNE visualization of visual and reasoning tokens for Q-Insight/Q-Tacit. Middle: normalized attention weights over the image and reasoning tokens after the quality score is generated. Q-Tacit obtained similar attention allocated to reasoning with only 11 tokens (vs. 112 for Q-Insight).
Figure 5: Sensitivity to localized distortions. We compare Q-Insight, VisualQuality-R1, and Q-Tacit by reporting predicted quality scores for each image (w/o distorted patch, top) and its counterpart with a localized corruption (w/ distorted patch, bottom). Blur, noise, and compression are injected within the red box (left to right). Q-Tacit consistently lowers its predicted scores in response to the local distortion.
Table 2: Ablation on the effect of latent reasoning pipeline (PLCC/SRCC). We compare Q-Tacit with two variants: Stage I-only and Stage I+II w/o LVR . Top two results are highlighted in bold and underlined , respectively.
Table 3: Ablation on the latent space composition (PLCC/SRCC). We vary the Stage I training data on three subsets ( ✓ indicates included data), while keeping Stage II identical across rows. Top two results are highlighted in bold and underlined , respectively.
Table 4: Ablation on reward design (PLCC/SRCC). Q-Insight and R2RL are two different reward design, and τ is the threshold in our dense quality scoring reward. Top two results are highlighted in bold and underlined , respectively.
Table 5: Ablation on latent reasoning budget T (PLCC/SRCC). We vary the latent-segment length T∈{4,16} and compare with the default setting T=8 . Top two results are highlighted in bold and underlined , respectively.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A: Qualitative score prediction examples from KADID-10k [ 27 ] . For each example, the normalized MOS as ground-truth score (GT) is shown above the image, and the bar plot compares predicted scores from Q-Tacit, Q-Insight [ 25 ] , and VisualQuality-R1 [ 58 ] . The dashed line indicates GT for easier visual comparison.
Method
#Parameters
Inference Time ↓
Inference Memory
#TFLOPs ↓
Q-Insight [ 25 ]
8.29 B
2.22 s
15.6 G
8.13
VisualQuality-R1 [ 58 ]
8.29 B
2.04 s
15.6 G
7.49
Q-Tacit(Ours)
8.29 B
0.83 s
15.6 G
6.62
Appendix
Table A: Averaged model complexity comparison using 512×384×3 images as input tested on single GH200 GPU.
Image Quality Assessment (IQA) in open-world environments remains challenging due to limited generalization and interpretability. Recent approaches based on multimodal large language models (MLLMs) introduce textual reasoning for quality prediction, yet their judgments rely heavily on semantically biased internal representations, making them insensitive to low-level perceptual degradations. We propose IQA-T1, a tool-based visual evidence reasoning framework that augments MLLM reasoning with explicit perceptual observations. During inference, the model autonomously invokes specialized analysis tools to generate structured visual evidence, such as noise residual maps, gradient statistics, and frequency spectra, which are progressively integrated into the reasoning process. To support this paradigm, we construct Q-Tool, a dataset containing 11k multimodal reasoning chains grounded in tool-generated evidence. Extensive experiments on seven IQA benchmarks show that IQA-T1 achieves the best overall performance across datasets while producing interpretable and evidence-grounded quality assessments. Code and dataset are available at https://github.com/zibuyu-02/IQA-T1.
Jinjian Wu, Jiaqi Tang, Wei Wei +5
School of Computer Science, Northwestern Polytechnical University, Xi’an, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China
Reasoning-induced vision-language models (VLMs) advance image quality assessment (IQA) with textual reasoning, yet their scalar scores often lack sensitivity and collapse to a few values, so-called discrete collapse. We introduce ME-IQA, a plug-and-play, test-time memory-enhanced re-ranking framework. It (i) builds a memory bank and retrieves semantically and perceptually aligned neighbors using reasoning summaries, (ii) reframes the VLM as a probabilistic comparator to obtain pairwise preference probabilities and fuse this ordinal evidence with the initial score under Thurstone's Case V model, and (iii) performs gated reflection and consolidates memory to improve future decisions. This yields denser, distortion-sensitive predictions and mitigates discrete collapse. Experiments across multiple IQA benchmarks show consistent gains over strong reasoning-induced VLM baselines, existing non-reasoning IQA methods, and test-time scaling alternatives.
Kanglong Fan, Tianhe Wu, Wen Wen +6
City University of Hong Kong, Hong Kong SAR, China · ByteDance Inc., China
Image Quality Assessment (IQA) is a long-standing problem in computer vision. Previous methods typically focus on predicting numerical scores without explanation or providing low-level descriptions lacking precise scores. Recent reasoning-based vision language models (VLMs) have shown strong potential for IQA by jointly generating quality descriptions and scores. However, existing VLM-based IQA methods often suffer from unreliable reasoning due to their limited capability of integrating visual and textual cues. In this work, we introduce Zoom-IQA, a VLM-based IQA model to explicitly emulate key cognitive behaviors: uncertainty awareness, region reasoning, and iterative refinement. Specifically, we present a two-stage training pipeline: 1) supervised fine-tuning (SFT) on our Grounded-Rationale-IQA (GR-IQA) dataset to teach the model to ground its assessments in key regions, and 2) reinforcement learning (RL) for dynamic policy exploration, stabilized by our KL-Coverage regularizer to prevent reasoning and scoring diversity collapse, with a Progressive Re-sampling Strategy for mitigating annotation bias. Extensive experiments show that Zoom-IQA achieves improved robustness, explainability, and generalization. The application to downstream tasks, such as image restoration, further demonstrates the effectiveness of Zoom-IQA.
Guoqiang Liang, Jianyi Wang, Zhonghua Wu +2
National University of Singapore · S-Lab, Nanyang Technological University, Singapore · SenseTime Research