EvoQuality: Balancing Quality and Diversity

• EvoQuality is a framework integrating quality and diversity to generate high-performing, varied solutions across behavioral or feature spaces.
• It utilizes archive-based methods like MAP-Elites, surprise search, and local Pareto sets to overcome premature convergence and enrich exploration.
• The paradigm extends to self-supervised image quality assessment, iteratively refining vision-language models through consensus-driven ranking.

EvoQuality denotes, across the cited literature, a family of ideas organized around the joint treatment of quality and diversity in search, and, in a later image quality assessment setting, the name of a specific self-evolving framework for vision-LLMs. In the quality-diversity literature, the central objective is not merely to recover a single optimum, but to generate a collection of diverse and high-performing solutions defined over a behavioral or feature space; in the IQA setting, the same label is used for an iterative, self-supervised process that improves ranking-based quality judgments through self-generated consensus (Chatzilygeroudis et al., 2020, Wen et al., 30 Sep 2025).

1. Conceptual scope

In quality-diversity optimization, the search target differs fundamentally from both classical single-objective optimization and multimodal optimization. Traditional optimization seeks a single global optimum, while multimodal optimization searches for multiple peaks in parameter space. Quality-Diversity instead works in behavioral space or feature space, and attempts to fill the whole behavior space, even if a niche is not a peak in the fitness landscape. The resulting archive provides a holistic view of how high-performing solutions are distributed throughout a search space, and the literature repeatedly describes this as illumination rather than mere optimization (Chatzilygeroudis et al., 2020).

This framing emerged partly as a response to premature convergence. In evolutionary robotics, traditional evolutionary algorithms tend to converge to local optima when the problem is complex or the search space is large. Quality Diversity algorithms try to overcome premature convergence by introducing additional measures that reward solutions for being different while not necessarily performing better. In that sense, EvoQuality is not simply diversity preservation as an auxiliary heuristic; it is a search regime in which diversity is part of the problem definition itself (Nordmoen et al., 2020).

A common misconception is that EvoQuality is equivalent to maximizing diversity for its own sake. The cited work does not support that reading. The persistent theme is the production of high-performing solutions distributed across meaningful behavioral variation. This suggests that EvoQuality is best understood as a family of methods for structuring exploration so that performance is retained while search does not collapse onto a narrow behavioral mode.

2. Formalization of quality, diversity, and archive-level scores

A compact statement of the QD objective is to find, for each point in behavior space, the highest-performing solution that exhibits that behavior:

$$\forall\ \boldsymbol{b}\in\mathcal{B}:\quad \boldsymbol{\theta}^* = \underset{\boldsymbol{\theta}}{\arg\max}\ f_{\boldsymbol{\theta}} \quad \text{s.t.}\ \boldsymbol{b} = \boldsymbol{b}_{\boldsymbol{\theta}}$$

This formulation makes explicit that quality is conditioned on behavior rather than optimized only globally (Chatzilygeroudis et al., 2020).

In modular evolutionary robotics, the paper “Quality and Diversity in Evolutionary Modular Robotics” defines quality as locomotion speed, operationalized by the fitness of a robot as the straight-line distance its root module travels in the XY-plane during simulation. Diversity is defined morphologically through the explicit descriptor

$b' = (m_i, j_i),$

where $m_i$ is the number of non-movable modules and $j_i$ the number of movable joint modules. For NSGA-II, diversity is quantified as the average pairwise difference in morphological descriptors within the population:

$D(x) = \frac{1}{|P_n|} \sum_{y \in P_n} d(x, y),$

with

$d(x, y) = 1.0 - e^{- |(m_x, j_x) - (m_y, j_y)| }.$

The same study uses coverage as the number of unique niches filled and defines a Quality-Diversity Score as

$\text{QD-score}(m) = \sum_{x \in m} Q_x,$

where $m$ is the repertoire map and $Q_x$ the fitness stored in each niche (Nordmoen et al., 2020).

Multi-objective extensions preserve the same archive logic while replacing scalar quality with local Pareto structure. In “Multi-Objective Covariance Matrix Adaptation MAP-Annealing”, each archive cell stores a local Pareto Set, and the archive-wide score becomes

$$\textrm{MOQD-Score} = \sum_{e=1}^{|A|} \mathrm{HV}(F_e),$$

where $b' = (m_i, j_i),$0 is the hypervolume of the local Pareto Front in cell $b' = (m_i, j_i),$1. This shifts EvoQuality from one quality value per niche to a set of trade-offs per niche, while retaining the diversity mandate over behavioral cells (Zhao et al., 27 May 2025).

3. Algorithmic families and search regimes

MAP-Elites is the canonical archive-based QD algorithm in the cited literature. In modular robotics, it is compared with a single-objective evolutionary algorithm and with NSGA-II. All three algorithms evolve high-performing individuals, but MAP-Elites is reported as better adept at filling all niches with high-performing solutions, with the fastest and most consistent increase in coverage and very low variance between runs. The difference from NSGA-II is attributed to MAP-Elites’ storage and maintenance mechanism, which archives the best individual per niche rather than relying on Pareto-front competition alone (Nordmoen et al., 2020).

A second line of work broadens the diversity signal itself. “Quality Diversity Through Surprise” introduces Surprise Search with Local Competition (SS-LC), Novelty-Surprise Search with Local Competition (NSS-LC), and Novelty Search–Surprise Search–Local Competition (NS-SS-LC). In robot navigation across 60 highly deceptive mazes, the combination of surprise and novelty is reported to yield significantly higher efficiency, speed, and robustness than novelty search with local competition alone. The paper’s central claim is that novelty and surprise are orthogonal and synergistic: novelty rewards deviation from previously seen behaviors, while surprise rewards deviation from predicted behaviors, enabling efficient back-tracking in behavior space (Gravina et al., 2018).

Other methods reinterpret the balance between quality and diversity rather than using a full external archive. “Quality Evolvability ES” optimizes for task performance and evolvability simultaneously, but it aims to find a single individual with a diverse and well-performing distribution of offspring. By contrast, “Quality with Just Enough Diversity” does not attempt to uniformly populate behavior space; it learns a Gaussian Process model of the relationship between behavior and fitness, selects promising target behaviors from the Pareto front over GP mean and variance, and then uses Evolution Strategies locally. The stated objective is to use diversity tactically, only where it helps reach higher fitness values (Katona et al., 2021, Templier et al., 2024).

At the multi-objective end of the spectrum, “MO-CMA-MAE” integrates CMA-ES into Multi-Objective Quality-Diversity. Candidates are ranked by incremental hypervolume improvement within an archive cell, and a threshold-accepting annealing mechanism keeps under-explored cells attractive for longer. The reported result is significant improvement in MOQD-score and coverage over baselines such as MOME, NSGA-II, SMS-EMOA, and COMO-CMA-ES on Sphere and Rastrigin, while performing equivalently to MOME on Arm and Overcooked (Zhao et al., 27 May 2025).

Taken together, these algorithms show that EvoQuality is not tied to a single archive structure, a single notion of diversity, or a single optimization granularity. The literature supports at least four regimes: explicit cell archives, novelty/surprise-based divergence, targeted diversity for single-best search, and local Pareto-set illumination.

4. Descriptor construction, learned behavior spaces, and controllable exploration

A major methodological question in EvoQuality is how behavioral descriptors are obtained. Hand-coded descriptors are common in classical MAP-Elites, but “Unsupervised Behaviour Discovery with Quality-Diversity Optimisation” introduces AURORA, which learns behavioral descriptors directly from raw sensory data through an auto-encoder. AURORA alternates between a QD phase and an encoder update phase, periodically retraining the encoder on archive contents and remapping all descriptors. Across maze navigation, hexapod locomotion, and air-hockey, it performs similarly to traditional hand-coded approaches without the requirement to provide any hand-coded behavioral descriptor, and it finds behaviors that are novel with respect to more features than its hand-coded baselines (Grillotti et al., 2021).

The same paper also addresses descriptor dimensionality. Its Container Size Control (CSC) mechanism dynamically adjusts the novelty threshold so that archive size remains near a target value, and the reported finding is that CSC remains stable for latent spaces of dimension up to at least 10, whereas Volume Adaptive Threshold suffers from the curse of dimensionality for $b' = (m_i, j_i),$2. This is important because it changes the practical meaning of diversity: behavioral variation need not be limited to what a designer can enumerate a priori (Grillotti et al., 2021).

EvoQuality has also been extended to interactive design. “Controllable Exploration of a Design Space via Interactive Quality Diversity” introduces User-Controlled MAP-Elites (UC-ME). Here, the user is shown only a few alternatives sampled from a small window of the behavioral map; after a selection, the window is recentered on the chosen behavior characterization, parent selection is restricted to that window, and evolution proceeds locally. For constrained architectural layout generation, the method uses a two-archive scheme with feasible and infeasible archives, and the reported result is that global exploration is not as pronounced as in MAP-Elites, but the system finds more appropriate solutions to the user’s taste (Sfikas et al., 2023).

These developments challenge the assumption that EvoQuality requires a fixed, globally defined behavior space. The cited work shows learned descriptor spaces, local windows, and interactive recentering, all while preserving the central logic of quality-conditioned diversity.

5. Evaluation practice and application domains

Because EvoQuality methods return repertoires rather than single incumbents, evaluation typically combines performance, coverage, and distributional indicators. The Evogil platform is explicitly designed to support rigorous assessment of multi-objective and hybrid evolutionary algorithms, with quality indicators including Hypervolume (HV), Generational Distance (GD), Inverted Generational Distance (IGD), Average Hausdorff Distance (AHD), Epsilon, Extent, Spacing, and Pareto Dominance Indicator (PDI). It also provides replications, result caching, incremental updates, plotting tools, benchmark suites, and chunked persistence, and the paper states that these practices align with well-regarded evaluation frameworks such as “EvoQuality” (Idzik, 2019).

The application range of EvoQuality is broad. The survey chapter on QD optimization highlights robotics, deep learning, and reinforcement learning as successful areas, while the procedural content generation survey emphasizes that QD can create multiple systematically varying solutions, enabling new approaches to creative human-AI interaction and adaptivity. In game-related PCG, behavior metrics define niches for levels, decks, structures, or weapons, and quality is imposed through playability, balance, or local competition within niches (Chatzilygeroudis et al., 2020, Gravina et al., 2019).

A related application outside canonical archive illumination is “EvolGAN”. It uses a quality estimator and a $b' = (m_i, j_i),$3 evolution strategy to search the latent space of pre-trained GANs at inference time, seeking higher-quality images while preserving the original generator’s diversity through locality in latent space. Human raters preferred EvolGAN outputs with frequency 83.7pc for Cats, 74pc for FashionGen, 70.4pc for Horses, and 69.2pc for Artworks, with minor improvements for already excellent GANs for faces (Roziere et al., 2020).

Later LLM-guided systems suggest further generalization. “EvoLattice” represents a population of programs or agent behaviors within a single directed acyclic graph whose nodes store multiple persistent alternatives, and its paper states that the resulting dynamics resemble quality-diversity optimization, emerging implicitly from the internal multi-alternative representation rather than an explicit external archive. This suggests that EvoQuality-like behavior can arise from representation and persistence mechanisms even when a MAP-Elites-style archive is absent (Yuksel, 15 Dec 2025).

6. EvoQuality as self-supervised image quality assessment

In “Self-Evolving Vision-LLMs for Image Quality Assessment via Voting and Ranking”, EvoQuality is the name of a specific framework for image quality assessment. It is described as a fully self-supervised method that enables a vision-LLM to refine its quality perception capabilities without any ground-truth labels. The method adapts self-consistency to the ranking-based nature of IQA: in an offline stage, the model generates pairwise pseudo-labels through majority voting over repeated comparisons; in an online stage, those pseudo-rankings define a fidelity reward used to update the policy through Group Relative Policy Optimization (GRPO) (Wen et al., 30 Sep 2025).

The pairwise pseudo-label is defined by majority voting as

$b' = (m_i, j_i),$4

and the fidelity reward for a sampled trajectory is

$b' = (m_i, j_i),$5

This reward encourages predicted pairwise preference probabilities to align with the model’s own consensus-derived pseudo-labels. The paper also reports that increasing the number of candidate reasoning paths for majority voting improves performance, and that the ranking-based self-training variant is more robust than a regression-style pseudo-labeling baseline termed EvoEstimate (Wen et al., 30 Sep 2025).

Empirically, EvoQuality improves the zero-shot base VLM by 31.8\% on PLCC and 33.7\% on SRCC, and it is reported to outperform supervised VLM-based IQA models on 5 out of 7 benchmarks. The framework is therefore distinct from earlier EvoQuality usages in evolutionary computation: it does not illuminate a behavioral archive, but it retains the core motif of iterative improvement through structured variation and selective retention. A plausible implication is that the term “EvoQuality” now spans both a general paradigm of quality-conditioned diversity and a concrete self-supervised algorithm for perceptual ranking.