Explainable Iterative Data Visualization Refinement via an LLM Agent

arXiv:2604.15319 · cs.HC, cs.AI · Submitted 2026-03-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Explainable Iterative Data Visualization Refinement via an LLM Agent".

Jane: Here's the gist of the paper. Exploratory analysis of high-dimensional data relies heavily on projecting data into low-dimensional spaces (typically 2D or 3D) to uncover latent structures.

Tom: The authors propose an agentic AI pipeline that leverages a large language model (LLM) to bridge the gap between rigorous quantitative assessment and qualitative human insight.

Jane: The Core Problem Addressed Exploratory data analysis of high-dimensional datasets relies heavily on projecting data into low-dimensional spaces (typically 2D or 3D) to uncover latent structures.

Tom: Finding a suitable algorithm configuration, particularly hyperparameter setting, to produce a visualization plot that faithfully represents the underlying reality and encourages pattern discovery remains challenging.

Paper summary: Tom: So, this paper is about taking all that high-dimensional data, right? We project it down into low dimensions like two or three dimensions so we can actually look at it and find the patterns. But finding the right settings for those projection algorithms, those hyperparameters, to make the resulting picture actually represent what's going on in the data is really hard work.

Jane: That sounds like a problem we all run into when we’re visualizing complex datasets. So what does this paper propose? It seems to be using an AI agent, specifically a large language model or LLM, to help bridge the gap between those hard numbers and what a human actually sees.

Lu: Right. The main idea is treating the visualization evaluation and that hyperparameter optimization process as a kind of semantic task for the LLM <ref:2604.15319#pg2>. They’re building an agentic pipeline to automate that iterative refinement loop so you don't have to do all the tedious tuning yourself.

Meng: Automating hyperparameter selection sounds cool, but what does it actually *do*? Does it just spit out a list of settings? Because if it doesn't actually fix the visualization, we just have a fancy report that doesn't help us debug.

Tom: The paper claims this system generates a multi-faceted report that contextualizes the hard metrics with descriptive summaries and suggests actionable recommendations for algorithm configuration <ref:2604.15319#pg1>. It’s not just guessing; it uses these LLMs to reason about what the visualization is missing.

Jane: So, how does this agent actually know what's missing? It seems they feed the LLM four different types of input data: quantitative metrics like Trustworthiness or Silhouette scores, cluster dendrograms, the embedding coordinates themselves, and then the visualization plot it’s looking at <ref:2604.15319#pg3>.

Lu: That multi-modal input is key because it lets the LLM act as a visual diagnostician <ref:2604.15319#pg3>. It gets to compare the raw numbers against what’s visually happening, which is where the real power seems to be.

Meng: Comparing those numbers to a picture sounds like heavy computation, but what are we looking at in terms of specific metrics? Are we talking about just Trustworthiness and Stress scores? Because I need to know if that’s something practical for production data.

Tom: They use a few, including Trustworthiness T of k, Continuity C of k, Stress S, and Silhouette scores <ref:2604.15319#pg3>. But the paper points out a big issue with those numbers: they often lack explanatory power <ref:2604.15319#pg2>.

Jane: That’s exactly what the paper highlights—a low Trustworthiness score might indicate a problem, but it doesn't tell you where or why the distortion is happening <ref:2604.15319#pg2>. The agent seems designed to get past that limitation by producing an Overall Assessment and Actionable Recommendation <ref:2604.15319#pg3>.

Paper summary: Lu: The iterative loop itself is what’s ambitious here. In each round, the DR engine computes new embeddings based on the previous iteration's suggestions, and those new embeddings are fed back to the LLM to dictate the next set of parameters <ref:2604.15319#pg3>.

Meng: So it’s like a self-correcting system? If we start with a bad setting, the AI figures out how to adjust for the next try, instead of us having to guess every time. But how much computation does that take per iteration? I need to know if this is feasible on our existing infrastructure.

Tom: The paper shows it rapidly navigates the hyperparameter search spaces for algorithms like t-SNE, UMAP, and PaCMAP <ref:2604.15319#pg4>. And they show that when comparing implicit LLM reasoning against explicit mathematical scoring, the agent successfully identifies visual artifacts like artificial fragmentation or class dominance <ref:2604.15319#pg4>.

Jane: That’s really interesting because it suggests the LLM isn't just following a strict math formula; it’s using qualitative reasoning to spot those kinds of distortions that a pure number might miss <ref:2604.15319#pg2>.

Lu: The comparative analysis shows that while explicit metric weighting offers a predictable path, implicit holistic reasoning actually captures and corrects those nuanced visual artifacts better <ref:2604.15319#pg4>. For example, one LLM prioritized local neighborhood preservation by assigning weights of zero point three zero to Trustworthiness and zero point three zero to the Silhouette Score <ref:2604.15319#pg4>.

Meng: So, it’s telling us that sometimes you have to trust the holistic judgment over just following a strict mathematical objective function for the best result <ref:2604.15319#pg3>. What about those other algorithms? Does this work equally well for things like UMAP when we’re dealing with continuous trajectories instead of discrete clusters?

Tom: The paper also shows that improving global topology, which is one goal, often comes at the direct expense of local structure <ref:2604.15319#pg5>. The agent seems to navigate that trade-off dynamically based on what it learns about the data structure <ref:2604.15319#pg5>.

Jane: It sounds like this framework is trying to solve those five core problems in data visualization and understanding, which are things like lack of explanatory power and ambiguity in diagnosis <ref:2604.15319#pg2>. The whole point is to make the visualizations more faithful representations of the high-dimensional reality <ref:2604.15319#pg1>.

Lu: And they show that when they compare both the implicit scores and the explicit quantitative metrics, there’s a strong correlation between them over multiple iterations <ref:2604.15319#pg5>. That suggests the LLM's implicit score is actually a reliable proxy for using those explicit mathematical optimization techniques <ref:2604.15319#pg5>.

Paper summary: Meng: If that correlation holds, then we might be able to use less intense methods for the initial tuning phase and just let the agent refine it later on <ref:2604.15319#pg3>. That would save a lot of compute time before we even get to the final, high-fidelity plot.

Tom: This whole system is designed to produce a high-quality visualization plot in full automation <ref:2604.15319#pg4>. It’s about accelerating pattern discovery without needing exhaustive manual tuning <ref:2604.15319#pg1>.

Jane: So, the core message of this paper, "Explainable Iterative Data Visualization Refinement via an LLM Agent," is that we can use AI agents to combine rigorous quantitative assessment with qualitative human insight to create better visualizations automatically <ref:2604.15319#pg1>. It moves us away from purely mathematical optimization toward a system that understands the visual data itself.

Lu: And they emphasize explainability by providing a clear rationale for every single parameter adjustment, which is important because we need to know *why* the system chose that setting <ref:2604.15319#pg5>.

Meng: So, this means if we use this framework, we can get visualizations that are both biologically faithful and interpretable with less guesswork <ref:2604.15319#pg4>. It sounds like a lot of overhead to build the initial agent setup though.

Tom: The paper does acknowledge that they have limitations, which is important because it’s not a magic fix <ref:2604.15319#pg2>. One limitation they point out is that these metrics are often reductive because a single scalar value can't capture the nuance of complex topologies <ref:2604.15319#pg2>.

Jane: Exactly, and they also note that relying on clustering-specific metrics implicitly assumes the data should form discrete groups, which penalizes valid non-convex structures or continuous trajectories that are really important for understanding the data <ref:2604.15319#pg2>.

Lu: The system stops working when it hits those inherent limitations of using scalar values to capture complex topology <ref:2604.15319#pg2>. That’s the caveat they flag about not being able to fully capture the nuance <ref:2604.15319#pg2>.

Meng: So, if you were deploying this for real traffic, I'd worry about how robust it is when the underlying data structure doesn't neatly fit into a cluster <ref:2604.15319#pg2>. It seems like its strength is in structured data where we *expect* some kind of grouping.

Tom: The authors are pushing toward agentic AI systems because they see a shift from traditional automated machine learning to this kind of reasoning system <ref:2604.15319#pg3>. That’s the broader implication for how we build these visualization tools <ref:2604.15319#pg3>.

Jane: So, in conclusion, they are proposing an explainable iterative framework that uses an LLM agent to refine visualizations automatically by balancing hard metrics with qualitative understanding <ref:2604.15319#pg1>. It’s a tool for accelerating pattern discovery while trying to keep the final plot faithful to the high-dimensional data <ref:2604.15319#pg5>.

Conclusion: Segment: Conclusion**

Tom: So, to wrap up this paper, they’re proposing an agentic pipeline that uses an LLM to iteratively refine data visualizations by balancing strict mathematical metrics with human insight.

Jane: It's essentially taking that complicated process of picking the right settings for a projection algorithm and automating it using reasoning from a large language model.

Lu: They formalize it as an iterative loop where the LLM agent looks at the current visualization, generates a report on its quality using various metrics, and then suggests new parameters for the next attempt.

Meng: So you’re telling me this system can basically self-correct to produce a better plot without someone manually tuning every single setting? That sounds like it could save a ton of time in production environments where we need quick results.

Tom: Exactly. The authors emphasize that they provide a clear rationale for every single parameter adjustment, which is important because you get to see the reasoning behind the change.

Jane: They’re showing that this approach successfully addresses five core problems in data visualization, like the lack of explanatory power and ambiguity when diagnosing what’s wrong with a plot.

Lu: And they found a strong correlation between the LLM's implicit scoring—its holistic judgment—and traditional explicit mathematical metrics over multiple iterations.

Meng: If that correlation holds up, it means we might be able to use less intense methods for initial tuning and just let the agent refine it later on. That could cut down on compute time significantly.

Tom: It’s about moving away from just following a fixed mathematical formula toward a system that actually understands the visual data and what makes it look right.

Jane: The authors conclude that this explainable, iterative framework bridges the gap between rigorous math and qualitative human insight for visualization refinement.

Lu: They also point out the limitations, saying these scalar metrics can't fully capture every nuance of complex data topologies, which means it’s not a magic fix for every single scenario.

Meng: So the big question now is how robust this thing is when we deal with data that doesn't neatly fit into clusters or simple structures.

Tom: That’s what we need to keep an eye on, because if it breaks down on complex data, it stops being a reliable tool for us.

Jane: Next up, we're going to look at how they tested this pipeline on those single-cell RNA sequencing datasets and what the concrete results showed.

Burak Susam, Tingting Mu

University of Manchester

cs.HC, cs.AI

Submitted: 2026-03-02

Updated: 2026-11-10

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: Exploratory analysis of high-dimensional data relies heavily on projecting data into low-dimensional spaces (typically 2D or 3D) to uncover latent structures.

Key concepts

Agentic AI Pipeline
This is an automated workflow where an LLM acts as an intelligent agent to perform complex tasks. It doesn't just answer questions; it plans a sequence of actions—like generating a visualization, evaluating it against metrics, and then deciding on the next parameter change—to achieve a final goal without constant human intervention.
Multi-Modal Input
The LLM agent receives four different types of data simultaneously: hard numbers (quantitative metrics like Silhouette scores), visual structures (cluster dendrograms and embedding coordinates), the visualization plot itself, and textual descriptions. This allows the agent to make a comprehensive diagnosis by looking at both the mathematical properties and the visual appearance of the data representation.
Iterative Optimization Loop
The system operates in a repeating cycle: generate a plot $ ightarrow$ evaluate it using quantitative and qualitative metrics $ ightarrow$ use an LLM agent to suggest hyperparameter changes for the next iteration. This loop continues until a high-quality, well-balanced visualization is achieved, ensuring that local structure preservation and global fidelity are optimized simultaneously.

Terminology

Summary

Exploratory analysis of high-dimensional data relies heavily on projecting data into low-dimensional spaces (typically 2D or 3D) to uncover latent structures. p. 0

Where the paper says it

By implementing an iterative optimization loop of this process, the system is able to produce rapidly a high-quality visualization plot, in full automation. E xploratory data analysis of high-dimensional datasets relies heavily on projecting data into low-dimensional spaces (typically 2D or 3D) to uncover latent structures.

The authors propose an agentic AI pipeline that leverages a large language model (LLM) to bridge the gap between rigorous quantitative assessment and qualitative human insight. p. 0
Where the paper says it

However, finding a suitable algorithm configuration, particularly hyperparameter setting, to produce a visualization plot that faithfully represents the underlying reality and encourages pattern discovery remains challenging. To address this challenge, we propose an agentic AI pipleline that leverages a large language model (LLM) to bridge the gap between rigorous quantitative assessment and qualitative human insight.

The Core Problem Addressed

Exploratory data analysis of high-dimensional datasets relies heavily on projecting data into low-dimensional spaces (typically 2D or 3D) to uncover latent structures. Finding a suitable algorithm configuration, particularly hyperparameter setting, to produce a visualization plot that faithfully represents the underlying reality and encourages pattern discovery remains challenging. p. 0

Where the paper says it

org preprint Explainable Iterative Data Visualisation Refinement via an LLM Agent Burak Susam, Department of Computer Science, University of Manchester, Manchester, M13 9PT, UK Tingting Mu, Department of Computer Science, University of Manchester, Manchester, M13 9PT, UK Abstract—Exploratory analysis of high-dimensional data relies on embedding the data into a low-dimensional space (typically 2D or 3D), based on which visualization plot is produced to uncover meaningful structures and to communicate geometric and distributional data characteristics. However, finding a suitable algorithm configuration, particularly hyperparameter setting, to produce a visualization plot that faithfully represents the underlying reality and encourages pattern discovery remains challenging.

The paper addresses the challenge of finding an informative and high-quality visualization plot by proposing an iterative optimization loop using an LLM agent. p. 1
Where the paper says it

Our goal is to develop a visualization framework that is easy to use and efficient at producing an informative and high-quality visualization plot, addressing the above identified limitations. Specifically, we propose a novel framework for iterative hyperparameter optimization and visualization refinement using AI agents, taking advantage of the strong pattern recognition, contextual reasoning, and broad domain knowledge encoded in a modern large language model (LLM).

The Agentic Pipeline Design

The proposed framework formalizes the visualization generation process as an iterative optimization loop guided by an LLM agent. This pipeline consists of three phases, including visualization generation, multi-modal evaluation, and hyperparameter refinement. p. 4

Where the paper says it

The agent continuously refines the DR algorithm configuration by taking into account (i) the generated visualization plot, (ii) the 2D embedding source used for producing the plot, and (iii) a collection of metrics and extracted quantities summarizing the geometric and distributional information of the data. Iterative LLM-Guided Pipeline Our proposed pipeline consists of three phases, including visualization generation, multi-modal evaluation, and hyperparameter refinement, which is illustrated in Figure 1.

In each iteration, a DR & visualization engine computes 2D embeddings from the highdimensional data using a specific DR configuration, and produces the corresponding visualization plot using the embeddings. p. 4
Where the paper says it

Iterative LLM-Guided Pipeline Our proposed pipeline consists of three phases, including visualization generation, multi-modal evaluation, and hyperparameter refinement, which is illustrated in Figure 1. In each iteration, a DR & visualization engine computes 2D embeddings from the highdimensional data using a specific DR configuration, and produces the corresponding visualization plot using the embeddings.

The resulting embeddings are then fed into the LLM agent to produce a structured JSON report that dictates the parameter search for the subsequent iteration. p. 5
Where the paper says it

processed by a quantitative analyst engine and a clustering engine. The obtained results, together with the embedding coordinates and visualization plot, are fed into the LLM agent to produce a structured JSON report that dictates the parameter search for the subsequent iteration.

Multi-Modal Input and Reasoning

To enable the LLM agent to act as a comprehensive visual diagnostician, it is supplied with four distinct data modalities. These modalities include quantitative metrics, cluster dendrograms, embedding coordinates, and the visualization plot itself. p. 5

Where the paper says it

This process repeats until the visualization and embedding quality meets the required standard or stabilizes. Multi-Modal Input of LLM Agent To enable the LLM agent to act as a comprehensive visual diagnostician, we supply it with a rich context comprising four distinct data modalities for it to reason: Quantitative Metrics: The quantitative analyst engine implements a collection of data analysis tools as explained earlier.

The input schema for the LLM agent includes quantitative metrics such as Trustworthiness T(k), Continuity C(k), Stress S, and Silhouette scores. p. 5
Where the paper says it

Multi-Modal Input of LLM Agent To enable the LLM agent to act as a comprehensive visual diagnostician, we supply it with a rich context comprising four distinct data modalities for it to reason: Quantitative Metrics: The quantitative analyst engine implements a collection of data analysis tools as explained earlier. It computes the trustworthiness T(k), continuity C(k), Stress S, and Silhouette scores, as well as LOF-based outlier statistics to quantify embedding quality at both local and global scales.

Furthermore, the agent is prompted to produce a structured JSON response that includes an Overall Assessment & Quality Score, Structural & Visual Inspection, Outlier Forensics, and Actionable Recommendation. p. 6
Where the paper says it

Outlier Forensics: An assessment of LOF distributions is asked, which should flag potentially concerning patterns or data quality issues. Actionable Recommendation: The most importantly, the LLM agent is prompted to produce a prioritized list of hyperparameter adjustments, e.

Experimental Validation and Findings

The efficacy of the proposed framework was validated on complex, high-dimensional single-cell RNA sequencing datasets. The results demonstrated that the agentic pipeline rapidly navigates the hyperparameter search spaces of diverse dimensionality reduction algorithms, including tSNE, UMAP, and PaCMAP. p. 1

Where the paper says it

To validate the efficacy of the proposed framework, we evaluate our system on complex, high-dimensional single-cell RNA sequencing datasets. Our experimental results demonstrate that the agentic pipeline rapidly navigates the hyperparameter search spaces of diverse dimensionality reduction algorithms, including tSNE, UMAP, and PaCMAP, achieving high-quality convergence typically within a few iterations.

Comparative analysis between implicit LLM reasoning and explicit mathematical scoring reveals that the agent successfully identifies and resolves detailed visual artifacts such as artificial fragmentation and class dominance. p. 1
Where the paper says it

Our experimental results demonstrate that the agentic pipeline rapidly navigates the hyperparameter search spaces of diverse dimensionality reduction algorithms, including tSNE, UMAP, and PaCMAP, achieving high-quality convergence typically within a few iterations. Furthermore, comparative analysis between implicit LLM reasoning and explicit mathematical scoring reveals that the agent successfully identifies and resolves detailed visual artifacts such as artificial fragmentation and class dominance that rigid objective functions often overlook.

The framework effectively balances local neighborhood preservation with global geometric fidelity, generating biologically faithful, interpretable visualization plots in full automation. p. 1
Where the paper says it

Furthermore, comparative analysis between implicit LLM reasoning and explicit mathematical scoring reveals that the agent successfully identifies and resolves detailed visual artifacts such as artificial fragmentation and class dominance that rigid objective functions often overlook. Ultimately, the framework effectively balances local neighborhood preservation with global geometric fidelity, generating biologically faithful, interpretable visualization plots in full automation while providing clear, actionable rationales for every refinement step.

Optimization Strategies Comparison

The comparative analysis of implicit and explicit scoring mechanisms reveals that while explicit metric weighting offers a predictable, math-driven optimization route, implicit holistic reasoning better captures and corrects nuanced visual artifacts. For instance, GPT-5. p. 1

Where the paper says it

Most critically, these numbers lack explanatory power. For instance, a low Trustworthiness score can indicate a problem, but often fails to explain where or why the distortion is occurring.

2 prioritized local neighborhood preservation and cluster separation by assigning the highest weights to Trustworthiness (0. p. 9
Where the paper says it

Explicit Weighting Strategies: When prompted to act as quantitative strategists, the three LLMs designed noticeably different composite objective functions to evaluate the 2D embeddings. 2 prioritized local neighborhood preservation and cluster separation, assigning the highest weights to Trustworthiness (0.

30) and the Silhouette Score (0. p. 0
Where the paper says it

, Stress, Trustworthiness) and, frequently, clustering metrics (e. , Silhouette score) used as proxies for visual separability.

30). Conversely, Claude-Opus-4. p. 8
Where the paper says it

LLM Agents: To understand the impact of the underlying reasoning engine on hyperparameter search efficiency, we compare the pipeline based on three frontier LLMs, including GPT-5. 2, Gemini-3-Pro-Preview, and Claude-Opus-4.

5 demonstrated a strong bias toward strict clustering and identifying density anomalies, assigning 0. p. 9
Where the paper says it

Claude-Opus4. 5 demonstrated a strong bias toward strict clustering and identifying density anomalies, assigning 0.

35 to the Silhouette Score. p. 0
Where the paper says it

, Stress, Trustworthiness) and, frequently, clustering metrics (e. , Silhouette score) used as proxies for visual separability.

The paper also shows that improving global topology often comes at the direct expense of local structure. p. 9
Where the paper says it

50 Note: Values indicate initial and final converged scores out of 10. Higher scores indicate better overall performance in global topology, local neighborhood and hierarchy preservation.

Conclusion on Framework Utility

Ultimately, this explainable, iterative framework bridges the gap between rigorous mathematical evaluation and qualitative human insight. By providing a clear rationale for every parameter adjustment, the pipeline ensures that the resulting scatter plots faithfully represent the underlying high-dimensional reality. p. 0

Where the paper says it

More sophisticated search and optimization algorithms like Bayesian optimization, evolutionary optimization, and reinforcement learning based search require highly specialized knowledge and dedicated design on algorithms. In addition to the challenge of identifying suitable hyperparameter configuration, evaluating whether a given visualization is able to faithfully represent the high-dimensional reality, covering the underlying geometric, topological and distributional characteristics of the high-dimensional data constitutes another challenge.

p. 0
Where the paper says it

More sophisticated search and optimization algorithms like Bayesian optimization, evolutionary optimization, and reinforcement learning based search require highly specialized knowledge and dedicated design on algorithms. In addition to the challenge of identifying suitable hyperparameter configuration, evaluating whether a given visualization is able to faithfully represent the high-dimensional reality, covering the underlying geometric, topological and distributional characteristics of the high-dimensional data constitutes another challenge.

p. 0
Where the paper says it

More sophisticated search and optimization algorithms like Bayesian optimization, evolutionary optimization, and reinforcement learning based search require highly specialized knowledge and dedicated design on algorithms. In addition to the challenge of identifying suitable hyperparameter configuration, evaluating whether a given visualization is able to faithfully represent the high-dimensional reality, covering the underlying geometric, topological and distributional characteristics of the high-dimensional data constitutes another challenge.

The system is able to produce rapidly a high-quality visualization plot, in full automation. p. 0
Where the paper says it

By treating visualization evaluation and hyperparameter optimization as a semantic task, our system generates a multifaceted report that contextualizes hard metrics with descriptive summaries, and suggests actionable recommendation of algorithm configuration for refining data visualization. By implementing an iterative optimization loop of this process, the system is able to produce rapidly a high-quality visualization plot, in full automation.

p. 0
Where the paper says it

By treating visualization evaluation and hyperparameter optimization as a semantic task, our system generates a multifaceted report that contextualizes hard metrics with descriptive summaries, and suggests actionable recommendation of algorithm configuration for refining data visualization. By implementing an iterative optimization loop of this process, the system is able to produce rapidly a high-quality visualization plot, in full automation.

p. 0
Where the paper says it

By treating visualization evaluation and hyperparameter optimization as a semantic task, our system generates a multifaceted report that contextualizes hard metrics with descriptive summaries, and suggests actionable recommendation of algorithm configuration for refining data visualization. By implementing an iterative optimization loop of this process, the system is able to produce rapidly a high-quality visualization plot, in full automation.

p. 0
Where the paper says it

By treating visualization evaluation and hyperparameter optimization as a semantic task, our system generates a multifaceted report that contextualizes hard metrics with descriptive summaries, and suggests actionable recommendation of algorithm configuration for refining data visualization. By implementing an iterative optimization loop of this process, the system is able to produce rapidly a high-quality visualization plot, in full automation.

The authors conclude that this method accelerates pattern discovery without the need for exhaustive manual tuning.

Pruning redundant reasoning tokens

Exploratory analysis of high-dimensional data relies heavily on projecting data into low-dimensional spaces (typically 2D or 3D) to uncover latent structures. It is a common practice to use dimensionality reduction (DR) algorithms such as PCA, t-SNE [1] and UMAP [2] to compute 2D or 3D embeddings, and then to visualize those points on scatter plots so that it gets down to a humaninterpretable level visually. p. 0

Where the paper says it

E xploratory data analysis of high-dimensional datasets relies heavily on projecting data into low-dimensional spaces (typically 2D or 3D) to uncover latent structures. It is a common practice to use dimensionality reduction (DR) algorithms such as PCA, t-SNE [1] and UMAP [2] to compute 2D or 3D embeddings, and then to visualize those points on scatter plots so that it gets down to a humaninterpretable level visually.

The primary goal of this visualization process is not merely to partition data, but to achieve a holistic understanding of diverse data characteristics, e. p. 0
Where the paper says it

It is a common practice to use dimensionality reduction (DR) algorithms such as PCA, t-SNE [1] and UMAP [2] to compute 2D or 3D embeddings, and then to visualize those points on scatter plots so that it gets down to a humaninterpretable level visually. The primary goal of this visualization process is not merely to partition data, but to achieve a holistic understanding of diverse data characteristics, e.

g., identifying clusters, hierarchies, anomalies, and continuous manifolds. p. 0
Where the paper says it

The primary goal of this visualization process is not merely to partition data, but to achieve a holistic understanding of diverse data characteristics, e. , identifying clusters, hierarchies, anomalies, and continuous manifolds.

However, using DR algorithms for visualization is not a straightforward process as each algorithms involves their own hyperparemeters that govern the computed embeddings, and the outcomes are often highly sensitive to the hyperparameter configuration. p. 0
Where the paper says it

, identifying clusters, hierarchies, anomalies, and continuous manifolds. However, using DR algorithms for visualization is not a straightforward process as each algorithms involves their own hyperparemeters that govern the computed embeddings, and the outcomes are often highly sensitive to the hyperparameter configuration.

How the training signal is built

The LLM agent is prompted to produce a structured JSON report based on the multi-modal input. This JSON format ensures parsing reliability for the automated pipeline. p. 5

Where the paper says it

Structured JSON Diagnostics The LLM agent is prompted to produce a structured JSON report based on the multi-modal input. This JSON format ensures parsing reliability for the automated pipeline.

The diagnostic JSON report includes an Overall Assessment & Quality Score, Structural & Visual Inspection, Outlier Forensics, and Actionable Recommendation. p. 5
Where the paper says it

This JSON format ensures parsing reliability for the automated pipeline. The diagnostic JSON report includes: Overall Assessment & Quality Score: The LLM agent is prompted to produce a high-level numerical score supported by rationale in text description.

This structure allows the pipeline to parse the suggested parameters programmatically for the next iteration. p. 6
Where the paper says it

The agent is constrained to output a structured JSON response that articulates the reasoning before making actionable hyperparameter recommendations. This structure allows the pipeline to parse the suggested parameters programmatically for the next iteration, which is detailed below: Listing 2: Schema of Diagnostic Output and Hyperparameter Recommendation.

The LLM is constrained to output a structured JSON response that articulates the reasoning before making actionable hyperparameter recommendations. p. 6
Where the paper says it

} } Output Format (Diagnostic Response) The LLM acts as an expert diagnostician, producing a detailed analysis report by reasoning about the input metrics and the visual plot (processed via vision capabilities). The agent is constrained to output a structured JSON response that articulates the reasoning before making actionable hyperparameter recommendations.

Where the gains come from

The iterative optimization loop allows the system to produce rapidly a high-quality visualization plot, in full automation. The agentic pipeline exploits an LLM agent to reason on algorithm behavior and user need, then come up with implicit (or explicit) scoring schemes to guide the hyperparameter search. p. 4

Where the paper says it

Therefore, it motivates us to design an agentic AI pipeline to facilitate this process. For any given DR algorithm A, the pipeline exploits an LLM agent to reason on algorithm behavior and user need, then come up with implicit (or explicit) scoring schemes to guide the hyperparameter search, and rapidly find satisfactory hyperparameter setting in few iterations.

By forcing the LLM to populate the dendrogram comparison and visual inspection fields before generating the recommendations array, the agent engages in a chain-of-thought diagnostic process. p. 6
Where the paper says it

], "follow_up_metrics": [ . ] } Overall, by forcing the LLM to populate the dendrogram_comparison and visual_inspection fields before generating the recommendations array, the agent engages in a chain-of-thought diagnostic process.

The final result is that the framework effectively balances local neighborhood preservation with global geometric fidelity, generating biologically faithful, interpretable visualization plots in full automation. p. 1
Where the paper says it

Furthermore, comparative analysis between implicit LLM reasoning and explicit mathematical scoring reveals that the agent successfully identifies and resolves detailed visual artifacts such as artificial fragmentation and class dominance that rigid objective functions often overlook. Ultimately, the framework effectively balances local neighborhood preservation with global geometric fidelity, generating biologically faithful, interpretable visualization plots in full automation while providing clear, actionable rationales for every refinement step.

p. 1
Where the paper says it

Furthermore, comparative analysis between implicit LLM reasoning and explicit mathematical scoring reveals that the agent successfully identifies and resolves detailed visual artifacts such as artificial fragmentation and class dominance that rigid objective functions often overlook. Ultimately, the framework effectively balances local neighborhood preservation with global geometric fidelity, generating biologically faithful, interpretable visualization plots in full automation while providing clear, actionable rationales for every refinement step.

The convergence is determined using the quality score from the LLM agent or explicit score whose weights defined by the agent. p. 5
Where the paper says it

The agent evaluates the current state of the visualization and outputs a JSON response containing hyperparameter recommendations, which are fed back into the DR and Visualization Engine for the next iteration until the best visualization is achieved. The convergence is determined using the quality score from the LLM agent or explicit score whose weights defined by the agent.

The agents exhibit markedly different exploration strategies and convergence rates, with Gemini-3-Pro-Preview demonstrating the most effective balance of exploration and exploitation. p. 9
Where the paper says it

By plotting both the implicit LLM Quality Score and the explicit Composite Quantitative Score alongside the raw metric values, several distinct behavioral patterns and optimization trade-offs emerge. Model Performance and Convergence Strategies: The three agents exhibited markedly different exploration strategies and convergence rates: • Gemini-3-Pro-Preview (5 iterations): This model 10

p. 9
Where the paper says it

By plotting both the implicit LLM Quality Score and the explicit Composite Quantitative Score alongside the raw metric values, several distinct behavioral patterns and optimization trade-offs emerge. Model Performance and Convergence Strategies: The three agents exhibited markedly different exploration strategies and convergence rates: • Gemini-3-Pro-Preview (5 iterations): This model 10

When looking into a scatter plot of 2D embeddings generated from high-dimensional data by a DR algorithm, one expects to establish a comprehensive understanding of the data, by identifying and analyzing hierarchies of clusters, cluster formation and separation, data density variation and potential anomalies, as well as local neighborhood and manifold properties. Specifically, we can compute a cluster hierarchy by applying a hierachical clustering algorithm, e. p. 2

Where the paper says it

In addition to visual inspection facilitated by color, point size and line/curve based highlights, a visualization should be accompanied by data analysis tools to formalize such patterns of interest. Specifically, we can compute a cluster hierarchy by applying a hierachical clustering algorithm, e.

g., UPGMA (Unweighted Pair Group Method with Arithmetic Mean) [14]. p. 2
Where the paper says it

Specifically, we can compute a cluster hierarchy by applying a hierachical clustering algorithm, e. , UPGMA (Unweighted Pair Group Method with Arithmetic Mean) [14].

In addition to highlighting the clusters in color, the result can be separately visualized as a dendrogram (tree diagram), supplementing the scatter plot. p. 2
Where the paper says it

, UPGMA (Unweighted Pair Group Method with Arithmetic Mean) [14]. In addition to highlighting the clusters in color, the result can be separately visualized as a dendrogram (tree diagram), supplementing the scatter plot.

With respect to cluster arrangement, quantitative metrics that assess cluster quality in terms of intra-cluster tightness and inter-cluster separation, e. p. 2
Where the paper says it

In addition to highlighting the clusters in color, the result can be separately visualized as a dendrogram (tree diagram), supplementing the scatter plot. With respect to cluster arrangement, quantitative metrics that assess cluster quality in terms of intra-cluster tightness and inter-cluster separation, e.

g., Silhouette scores [15], can be presented. p. 2
Where the paper says it

With respect to cluster arrangement, quantitative metrics that assess cluster quality in terms of intra-cluster tightness and inter-cluster separation, e. , Silhouette scores [15], can be presented.

To track data density variation and flag potential outliers, the local outlier factor (LOF) algorithm [16] can be used. p. 2
Where the paper says it

, Silhouette scores [15], can be presented. To track data density variation and flag potential outliers, the local outlier factor (LOF) algorithm [16] can be used.

To examine and quantify characteristics of local neighborhood and manifold structure, we can calculate statistics that measure agreements between the ground truth neighbors indicated by the original high-dimensional data points and the observed neighbors based on the 2D embeddings. p. 2
Where the paper says it

It compares the local data density of a point to the densities of its neighbors, and flags points with less densities as outliers. To examine and quantify characteristics of local neighborhood and manifold structure, we can calculate statistics that measure agreements between the ground truth neighbors indicated by the original high-dimensional data points and the observed neighbors based on the 2D embeddings, and assess consistency between pairwise distances in the original and 2D spaces.

The authors emphasize system explainability by providing rationale for every change, particularly the reasoning and references to related quantitative metrics, hierarchy changes and visual effects for every parameter sweep. The paper demonstrates that when both scores compared, we also demonstrate the strong correlation between them as similar trends appear in both scores over multiple iterations. This correlation suggests that the LLM's implicit score is a reliable proxy for the explicit quantitative metric optimization. The framework successfully addresses the five core problems in data visualization and understanding, including lack of explanatory power and ambiguity in diagnosis. p. 1

Where the paper says it

Yet, this approach is subjective and susceptible to misinterpretation of DR artifacts, where the visual density or inter-cluster distances may not reflect the true high-dimensional relationships [1]. Such disconnect leads to the following five core problems in data visualization and understanding: (i) Lack of Explanatory Power: Quantitative metrics are not self-explanatory and fail to describe the nature of structural distortions.

p. 1
Where the paper says it

Yet, this approach is subjective and susceptible to misinterpretation of DR artifacts, where the visual density or inter-cluster distances may not reflect the true high-dimensional relationships [1]. Such disconnect leads to the following five core problems in data visualization and understanding: (i) Lack of Explanatory Power: Quantitative metrics are not self-explanatory and fail to describe the nature of structural distortions.

p. 1
Where the paper says it

Yet, this approach is subjective and susceptible to misinterpretation of DR artifacts, where the visual density or inter-cluster distances may not reflect the true high-dimensional relationships [1]. Such disconnect leads to the following five core problems in data visualization and understanding: (i) Lack of Explanatory Power: Quantitative metrics are not self-explanatory and fail to describe the nature of structural distortions.

The dynamic navigation through competing spatial constraints by the LLM agents is key to achieving high-quality convergence. p. 1
Where the paper says it

Our experimental results demonstrate that the agentic pipeline rapidly navigates the hyperparameter search spaces of diverse dimensionality reduction algorithms, including tSNE, UMAP, and PaCMAP, achieving high-quality convergence typically within a few iterations. Furthermore, comparative analysis between implicit LLM reasoning and explicit mathematical scoring reveals that the agent successfully identifies and resolves detailed visual artifacts such as artificial fragmentation and class dominance that rigid objective functions often overlook.

p. 1
Where the paper says it

Our experimental results demonstrate that the agentic pipeline rapidly navigates the hyperparameter search spaces of diverse dimensionality reduction algorithms, including tSNE, UMAP, and PaCMAP, achieving high-quality convergence typically within a few iterations. Furthermore, comparative analysis between implicit LLM reasoning and explicit mathematical scoring reveals that the agent successfully identifies and resolves detailed visual artifacts such as artificial fragmentation and class dominance that rigid objective functions often overlook.

p. 1
Where the paper says it

Our experimental results demonstrate that the agentic pipeline rapidly navigates the hyperparameter search spaces of diverse dimensionality reduction algorithms, including tSNE, UMAP, and PaCMAP, achieving high-quality convergence typically within a few iterations. Furthermore, comparative analysis between implicit LLM reasoning and explicit mathematical scoring reveals that the agent successfully identifies and resolves detailed visual artifacts such as artificial fragmentation and class dominance that rigid objective functions often overlook.

The low-dimensional embeddings computed by these algorithms are often evaluated using structural preservation metrics such as Trustworthiness, Continuity and Kruskal’s Stress [7]. These metrics are used to encourage the AI agent to produce more reliable visualization. p. 2

Where the paper says it

Specifically, global metrics assess the preservation of pairwise relationships across the entire dataset, while local metrics assess the fidelity of neighborhood structures around each point. As shown in our later pipeline design, these metrics are used to encourage the AI agent to produce more reliable visualization.

p. 2
Where the paper says it

Specifically, global metrics assess the preservation of pairwise relationships across the entire dataset, while local metrics assess the fidelity of neighborhood structures around each point. As shown in our later pipeline design, these metrics are used to encourage the AI agent to produce more reliable visualization.

Global Reliability Metrics include Spearman Distance Score [17], Stress Score [7], and Mean Distance Ratio focusing on scaling change. p. 3
Where the paper says it

Mean Distance Ratio focuses on assessing the scaling change of pairwise distance in from the highdimensional space to low-dimensional space, computed by R = 1 Np X i<j d LD ij d HD ij , (3) where Np = n(n−1) 2 is the total number of unique point pairs. This ratio examines the compression (or expansion) of the data region coverage before and after DR, providing an interpretable measure of global scaling distortion.

Local Reliability Metrics include Trustworthiness Score [7] and Continuity Score [7] which measure neighborhood preservation. p. 2
Where the paper says it

To examine and quantify characteristics of local neighborhood and manifold structure, we can calculate statistics that measure agreements between the ground truth neighbors indicated by the original high-dimensional data points and the observed neighbors based on the 2D embeddings, and assess consistency between pairwise distances in the original and 2D spaces. Examples of such statistics include trustworthiness, continuity and stress scores [7], which we will further explain in the next section.

These metrics are used to encourage the AI agent to produce more reliable visualization. The paper shows that improving global topology often comes at the direct expense of local structure. p. 9
Where the paper says it

50 Note: Values indicate initial and final converged scores out of 10. Higher scores indicate better overall performance in global topology, local neighborhood and hierarchy preservation.

The authors show that when both scores compared, we also demonstrate the strong correlation between them as similar trends appear in both scores over multiple iterations. This correlation suggests that the LLM's implicit score is a reliable proxy for the explicit quantitative metric optimization. The framework successfully addresses the five core problems in data visualization and understanding, including lack of explanatory power and ambiguity in diagnosis. The dynamic navigation through competing spatial constraints by the LLM agents is key to achieving high-quality convergence. The final result is that the framework effectively balances local neighborhood preservation with global geometric fidelity, generating biologically faithful, interpretable visualization plots in full automation.

The authors conclude that this explainable, iterative framework bridges the gap between rigorous mathematical evaluation and qualitative human insight. By providing a clear rationale for every parameter adjustment, the pipeline ensures that the resulting scatter plots faithfully represent the underlying high-dimensional reality. The system is able to produce rapidly a high-quality visualization plot, in full automation. The agents exhibit markedly different exploration strategies and convergence rates, with Gemini-3-Pro-Preview demonstrating the most effective balance of exploration and exploitation. The paper demonstrates that when both scores compared, we also demonstrate the strong correlation between them as similar trends appear in both scores over multiple iterations. This correlation suggests that the LLM's implicit score is a reliable proxy for the explicit quantitative metric optimization. The framework successfully addresses the five core problems in data visualization and understanding, including lack of explanatory power and ambiguity in diagnosis. The dynamic navigation through competing spatial constraints by the LLM agents is key to achieving high-quality convergence.

The paper introduces a novel framework for iterative hyperparameter optimization and visualization refinement using AI agents. This framework exploits an LLM agent to reason on algorithm behavior and user need, then come up with implicit (or explicit) scoring schemes to guide the hyperparameter search. p. 4

Where the paper says it

Therefore, it motivates us to design an agentic AI pipeline to facilitate this process. For any given DR algorithm A, the pipeline exploits an LLM agent to reason on algorithm behavior and user need, then come up with implicit (or explicit) scoring schemes to guide the hyperparameter search, and rapidly find satisfactory hyperparameter setting in few iterations.

The system is able to produce rapidly a high-quality visualization plot, in full automation. By providing a clear rationale for every parameter adjustment, the pipeline ensures that the resulting scatter plots faithfully represent the underlying high-dimensional reality. The system is able to produce rapidly a high-quality visualization plot, in full automation.

The paper introduces an agentic AI pipleline that leverages a large language model (LLM) to bridge the gap between rigorous quantitative assessment and qualitative human insight. By treating visualization evaluation and hyperparameter optimization as a semantic task, our system generates a multifaceted report that contextualizes hard metrics with descriptive summaries, and suggests actionable recommendation of algorithm configuration for refining data visualization. p. 0

Where the paper says it

To address this challenge, we propose an agentic AI pipleline that leverages a large language model (LLM) to bridge the gap between rigorous quantitative assessment and qualitative human insight. By treating visualization evaluation and hyperparameter optimization as a semantic task, our system generates a multifaceted report that contextualizes hard metrics with descriptive summaries, and suggests actionable recommendation of algorithm configuration for refining data visualization.

By implementing an iterative optimization loop of this process, the system is able to produce rapidly a high-quality visualization plot, in full automation. p. 0
Where the paper says it

By treating visualization evaluation and hyperparameter optimization as a semantic task, our system generates a multifaceted report that contextualizes hard metrics with descriptive summaries, and suggests actionable recommendation of algorithm configuration for refining data visualization. By implementing an iterative optimization loop of this process, the system is able to produce rapidly a high-quality visualization plot, in full automation.

The agentic pipeline takes multimodal input and employs an LLM agent to act as the expert evaluator. p. 4
Where the paper says it

AGENTIC AI PIPELINE DESIGN Our method formalizes the visualization generation process as an iterative optimization loop guided by an LLM agent. Instead of relying on manual trial-anderror hyperparameter selection for a DR algorithm, we propose an automated pipeline that takes multimodal input and employs an LLM agent to act as the expert evaluator.

The resulting visualization generated with the suggested hyperparameter setting by the agent should keep domain-relevant aspects while avoid algorithm-related artifacts. p. 1
Where the paper says it

In our framework, we prompt an LLM agent with a carefully designed context containing mathematical algorithms for computing metrics, hierarchy diagrams, embedding coordinates and visualization plot, and ask the agent to optimize hyperparameters of a given DR algorithm supported by appropriate rationale and reasoning process. The resulting visualization generated with the suggested hyperparameter setting by the agent should keep domain-relevant aspects while avoid algorithm-related artifacts.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Implement an iterative optimization loop guided by an LLM agent to refine data visualization plots in full automation, as described in Section 6. This allows for rapidly a high-quality visualization plot by treating visualization evaluation and hyperparameter optimization as a semantic task rather than relying solely on manual trial-and-error.

  2. Enable the LLM agent to act as an expert visual diagnostician by providing it with multimodal input, including Quantitative Metrics, Cluster Dendrograms, and the Visualization Plot. This equips the agent to reason about Gestalt patterns and visual artifacts that numerical metrics alone cannot capture.

  3. Develop a structured JSON output format for the LLM agent's diagnostic response, ensuring it includes fields like "Overall Assessment & Quality Score and a prioritized list of Actionable Recommendation," which are parsed programmatically for the next iteration. This ensures that suggested hyperparameter settings are directly tied to resolving specific structural distortions or visual artifacts identified in the current iteration.

  4. Compare implicit scoring versus explicit scoring mechanisms to determine the most effective strategy for LLM evaluation, as shown in Table 2 and Figure 4. This allows researchers to understand the inherent tension between mathematically rigid objective functions and holistic visual evaluation and select the prompting strategy that best captures nuanced visual artifacts like crowding or artificial cluster bridges.

  5. Integrate quantitative reliability metrics such as Spearman Distance Score, Stress Score, Trustworthiness Score, Continuity Score, and Mean Distance Ratio into the LLM's context for continuous evaluation. This provides a comprehensive collection of results to supplement the visualization plot and 2D embedding coordinates, helping the agent assess both global structure preservation and local neighborhood structures.

Abstract

Exploratory analysis of high-dimensional data relies on embedding the data into a low-dimensional space, based on which a visualization plot is produced to uncover meaningful structures and to communicate geometric and distributional data characteristics. Finding a suitable embedding algorithm configuration, particularly the hyperparameter setting, to produce a visualization plot that faithfully represents the underlying reality and encourages pattern discovery is labor-intensive and requires domain knowledge. To ease this burden, we propose an agentic AI pipeline that integrates rigorous quantitative assessment with large language model (LLM) based approximation of qualitative human judgment and insight. By treating visualization evaluation and hyperparameter optimization as a semantic task, our system generates a multi-faceted report that contextualizes quantitative metrics with descriptive summaries, and suggests actionable recommendations of algorithm configuration for refining data visualization. Through an iterative refinement loop, the system rapidly produces a high-quality visualization plot in full automation.

Sources

Related papers