AgenticIQA: Adaptive & Interpretable IQA Framework

• The paper introduces a novel framework that integrates VLMs with conventional IQA tools to provide both precise scores and structured explanations.
• It employs a three-agent architecture—Planner, Executor, and Summarizer—to decompose image quality assessment into detection, analysis, and tool-driven scoring.
• Empirical results show high correlation scores and improved interpretability, outperforming baseline models on standard IQA datasets while trading off higher computational cost.

AgenticIQA is an agentic framework for adaptive and interpretable image quality assessment that integrates vision-LLMs (VLMs) with traditional image quality assessment (IQA) tools in a dynamic, query-aware pipeline. It is designed for settings in which conventional full-reference (FR-IQA) and no-reference (NR-IQA) models are too rigid because they primarily output scalar scores, while recent VLM-based approaches improve interpretability but may sacrifice scoring precision and still separate explanation from scoring. AgenticIQA addresses this by decomposing IQA into distortion detection, distortion analysis, tool selection, and tool execution, coordinated by a planner, executor, and summarizer, and by coupling those modules with the AgenticIQA-200K instruction corpus and the AgenticIQA-Eval benchmark (Zhu et al., 30 Sep 2025).

1. Problem formulation and design rationale

AgenticIQA is motivated by two limitations in prior IQA practice. First, conventional IQA systems, whether FR-IQA or NR-IQA, typically produce scalar scores through static pipelines with limited adaptability and interpretability. Second, VLM-based IQA methods can provide richer descriptions but often treat scoring and interpretation as independent processes. The framework explicitly rejects that separation: interpretation identifies perceptual degradations, while scoring abstracts them into a compact metric. In the formulation adopted by AgenticIQA, these two operations are interdependent rather than modularly isolated.

The framework therefore combines the precision of traditional IQA tools with the adaptability and explainability of VLMs. Its operating assumptions are that user queries may be object-specific or global, that distortions may be explicit or implicit, and that assessment may require FR or NR tools depending on the case. This makes the system query-aware, multimodal, and tool-augmented rather than a single-pass scorer. The paper characterizes this as a modular incorporation of heterogeneous components—VLMs plus IQA tools—so that the system can return both scores and structured, human-aligned explanations (Zhu et al., 30 Sep 2025).

2. Three-agent architecture

The architecture is organized around three agents: a Planner, an Executor, and a Summarizer. All three modules run on a VLM backbone, with examples including Qwen2.5-VL and GPT-4o. In this design, the VLM provides reasoning, interpretation, inter-module communication, and language grounding, while traditional IQA tools function as specialist perceptual detectors.

The Planner interprets the user query and generates a structured multimodal assessment plan. Its outputs specify subtask configurations such as query type and scope, whether the request concerns a specific object or the global image, whether distortions are explicit in the query or must be inferred from the image, whether the setting is FR or NR, and whether IQA tool selection and execution are required.

The Executor sequentially performs the operational subtasks specified by the plan. Those subtasks are distortion detection, distortion analysis, tool selection, and tool execution. The Planner therefore does not itself score quality; it instantiates a task graph that determines which perceptual evidence must be collected.

The Summarizer integrates the intermediate evidence into a final answer. Its responsibilities include generating detailed explanations that combine distortion analysis with tool scores, fusing multiple tool outputs with a perceptually motivated weighting scheme, and self-reflecting on whether the evidence is sufficient. If not, it can trigger re-planning. This closes a plan–execute–summarize loop rather than a one-shot inference path (Zhu et al., 30 Sep 2025).

3. Subtasks, tool augmentation, and score fusion

The first operational stage is distortion detection. When distortions are not explicit in the query, a dedicated module identifies candidate distortion types using VLM-based reasoning, potentially at region or object level. The output is a candidate set of distortions $D$.

The second stage is distortion analysis. For each detected or user-specified distortion and target image or object, severity is quantified using the labels none, slight, moderate, severe, and extreme, and a textual rationale is generated. The paper writes the structured output as

$$A_i = \left\{ \left(d_i, l_i, r_i\right) \mid d_i \in D_k \right\} = \mathcal{E}_{da}(x, t_{da}, D_k, O_k).$$

This stage turns raw distortion hypotheses into structured perceptual evidence.

The third stage is tool selection. For every identified distortion, the system selects the best matching IQA tool or model from a library through VLM interpretation over tool metadata: $T_i = \mathcal{E}_{ts}(d_i, t_{ts}, \mathcal{T}).$ This selection governs whether FR or NR tools are used and which tool is most appropriate for each distortion class.

The fourth stage is tool execution. The selected IQA tool is invoked and its raw scalar score is normalized through a monotonic logistic transform: $$\hat{q}_i = \beta_1 \left( \frac{1}{2} - \frac{1}{\exp(\beta_2(\tilde{q}_i - \beta_3))} \right) + \beta_4 \tilde{q}_i + \beta_5.$$ Here $\tilde{q}_i$ denotes the raw score from tool $i$, and the $\beta$ parameters are pre-estimated. Normalization is necessary because the framework may combine heterogeneous FR-IQA and NR-IQA tools with incompatible native scales.

Final quality synthesis is explicitly HVS-inspired. First, the mean tool score is computed as

$\bar{q} = \frac{1}{n}\sum_{i=1}^{n}\hat{q}_i.$

Then a Gaussian-like weighting is applied around $\bar{q}$ for each candidate quality level $c \in \{1,2,3,4,5\}$, ranging from bad to excellent: $$A_i = \left\{ \left(d_i, l_i, r_i\right) \mid d_i \in D_k \right\} = \mathcal{E}_{da}(x, t_{da}, D_k, O_k).$$0 with $$A_i = \left\{ \left(d_i, l_i, r_i\right) \mid d_i \in D_k \right\} = \mathcal{E}_{da}(x, t_{da}, D_k, O_k).$$1 set to 1. The VLM supplies a predicted probability $$A_i = \left\{ \left(d_i, l_i, r_i\right) \mid d_i \in D_k \right\} = \mathcal{E}_{da}(x, t_{da}, D_k, O_k).$$2 for each quality level, and the final fused score is given as

$$A_i = \left\{ \left(d_i, l_i, r_i\right) \mid d_i \in D_k \right\} = \mathcal{E}_{da}(x, t_{da}, D_k, O_k).$$3

The resulting score is thus neither a pure tool average nor a pure language-model judgment; it is a fusion of normalized tool evidence and VLM semantic priors (Zhu et al., 30 Sep 2025).

4. Instruction corpus and benchmark design

To train and align the agent roles, the framework introduces AgenticIQA-200K, a large-scale instruction dataset tailored for IQA agents. The dataset contains 50K planning instructions, 100K execution instructions, and 50K summarization instructions. Its construction uses detailed annotations and QA templates from Q-Pathway and DQ-495K, with GPT-4o converting these into fine-grained instruction–response pairs targeted at the three agent roles and their logical transitions. Coverage includes synthetic, authentic, and generative images, together with query types such as what, how, why, tool-specific, and object-specific.

For evaluation, the paper introduces AgenticIQA-Eval, described as the first benchmark for assessing the planning, execution, and summarization capabilities of VLM-based IQA agents. The benchmark contains 1000 images, comprising 750 NR-IQA images and 250 FR-IQA images, sampled from MICBench and Waterloo with additional synthetic distortions. It is divided into three multiple-choice tracks: Planner with 250 questions, Executor with 500 questions, and Summarizer with 250 questions. Each multiple-choice item has one human-verified ground-truth answer, and accuracy is reported per track (Zhu et al., 30 Sep 2025).

This benchmark design is significant because it evaluates subskills rather than only end-point scalar agreement. The assessment target is therefore execution quality within the agentic workflow, not only final score correlation.

5. Empirical results

On AgenticIQA-Eval, AgenticIQA with GPT-4o reports 80.4% on Planner, 64.4% on Executor–Distortion, 74.4% on Executor–Tool, 86.0% on Summarizer, and 76.3% overall. AgenticIQA with Qwen2.5 reports 76.8%, 63.2%, 76.4%, 85.2%, and 75.4% respectively. The next-best open-source baseline is reported as no more than 74.4% on Planner, no more than 55.6% on Executor–Distortion, no more than 78.0% on Executor–Tool, no more than 84.8% on Summarizer, and no more than 73.2% overall. The reported takeaway is that AgenticIQA outperforms open-source VLM baselines and matches or exceeds strong proprietary models on stepwise IQA reasoning (Zhu et al., 30 Sep 2025).

On standard IQA datasets, the paper reports Spearman’s rank correlation coefficient values of 0.9165 on TID2013, 0.8889 on BID, and 0.7937 on AGIQA-3K for AgenticIQA with GPT-4o. In the same comparison, the top baseline on TID2013 is TopIQ at 0.9075, Q-Align reports 0.8967 on BID, and Q-SiT reports 0.7901 on AGIQA-3K. These results indicate that the framework is competitive as a scorer while also supporting structured explanation.

For explanation quality, LLVisionQA results show 77.95% overall accuracy for AgenticIQA with GPT-4o, compared with 77.88% for plain GPT-4o and 76.32% for the best open-source baseline. The paper attributes the gain to structured, multi-stage reasoning. Ablation evidence further states that modular, agentic reasoning yields substantial accuracy gains, with TID2013 improving by 16% SRCC relative to a single-pass VLM configuration, and that explanation quality and local distortion understanding improve when planning, execution, and explanation stages are separated.

The paper also notes a cost trade-off. The agent incurs higher tokens and runtime because of flexibility and verification overhead, and improving sample efficiency is left as an open problem.

6. Position within the wider agentic evaluation literature

In the narrow sense established by the title paper, AgenticIQA denotes an image quality assessment framework. In adjacent literature from 2025–2026, however, “IQA” and “AgenticIQA” also appear in a broader sense of intelligent quality assessment for agentic AI systems. A formalization paper introduces a dual-model framework consisting of a Host Agent Model and a Task Lifecycle Model, with 31 temporal-logic properties spanning liveness, safety, completeness, and fairness (Allegrini et al., 15 Oct 2025). A survey of agentic AI architectures frames evaluation through the CLASSic dimensions—Cost, Latency, Accuracy, Security, and Stability—and places IQA within a broader taxonomy of perception, memory, planning, action, tool use, and collaboration (V et al., 18 Jan 2026). An architectural analysis of goal-directed systems emphasizes typed tool interfaces, governance, observability, and reproducibility as first-class requirements for production agents (Alenezi, 11 Feb 2026). Process-centric work on Graphectory argues that quality analysis should not be outcome-centric alone, but should also encode temporal and semantic structure in agent trajectories (Liu et al., 2 Dec 2025).

This creates a terminological bifurcation. In one lineage, AgenticIQA is a specific perceptual framework for image quality assessment (Zhu et al., 30 Sep 2025). In another, IQA denotes formal, operational, and governance-oriented evaluation of agentic systems more generally (Allegrini et al., 15 Oct 2025, V et al., 18 Jan 2026, Alenezi, 11 Feb 2026, Liu et al., 2 Dec 2025). This suggests that the image-domain framework is also a concrete instantiation of broader agentic design commitments: decomposition of complex tasks, explicit tool mediation, structured evidence integration, and evaluation beyond a single scalar endpoint.