AutoManifold: Agentic Design of Data Visualisation Algorithms via Manifold Embedding
Organizations: Department of Computer Science; University of Manchester; Manchester, M13 9PT, UK · X32 Studio; Abu Dhabi, UAE
Abstract
Manifold embedding for data visualisation is dominated by a small set of hand-designed state-of-the-art (SOTA) algorithms. Their design targets are fixed at development time, resulting in inflexible algorithmic behaviour that cannot be readily steered towards user-specified structural priorities. Hand-designing new embedding algorithms to address user priority requires highly specialised expertise and is time consuming. To address autonomous algorithm design tailored to user preference, we introduce an agentic algorithm generation pipeline AutoManifold. It composes new manifold embedding algorithms from a constrained vocabulary of affinity, cost, and optimisation primitives, supported by multi-agent large-language-model (LLM) orchestration. AutoManifold conditions every stage of its design on user-specified structural preservation preferences, and iteratively refines the algorithm configuration through an LLM-guided, metric-grounded iterative loop. We compare the generated algorithms against the three strongest and most frequently used SOTA methods (t-SNE, UMAP, and PaCMAP) on real-world datasets spanning different difficulty regimes under identical evaluation infrastructure. The results demonstrate that LLM-driven, priority-conditioned algorithm synthesis can move beyond hyperparameter tuning, producing genuinely new algorithms that outperform traditional hand-designed methods in their ability to steer towards user-specified properties without compromising the intrinsic structure in the original data. The code of AutoManifold is available at https://github.com/buraksusam/AutoManifold_NeurIPS2026.
Figures & tables
| Dataset | Score | t-SNE | UMAP | PaCMAP | Auto (w/o) | Auto-T | Auto-C | Auto- | Auto-Sil | Auto-S | Auto-L | Auto-local | Auto-global |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| iris ( Unwin and Kleinman, 2021 ) | 0.9848 | 0.9810 | 0.9483 | 0.9865 | 0.9838 | 0.9858 | 0.9918 | 0.9200 | 0.9899 | 0.9866 | 0.9804 | 0.9907 | |
| 0.9842 | 0.9837 | 0.9689 | 0.9888 | 0.9816 | 0.9843 | 0.9946 | 0.9247 | 0.9922 | 0.9883 | 0.9858 | 0.9930 | ||
| Spearman ( ) | 0.8077 | 0.9602 | 0.7332 | 0.9629 | 0.8792 | 0.8845 | 0.9967 | 0.8535 | 0.9945 | 0.9297 | 0.9438 | 0.9927 | |
| Silhouette ( ) | 0.6383 | 0.6056 | 0.3978 | 0.6000 | 0.6165 | 0.6031 | 0.5238 | 0.7810 | 0.5274 | 0.6320 | 0.6454 | 0.5402 | |
| Stress ( ) | 0.5164 | 0.2205 | 0.5408 | 0.1689 | 0.3985 | 0.4769 | 0.0334 | 0.4730 | 0.0458 | 0.4527 | 0.2870 | 0.0734 | |
| LOF ( ) | 1.0088 | 1.0093 | 1.0155 | 1.0131 | 1.0106 | 1.0129 | 1.0162 | 1.0146 | 1.0163 | 1.0080 | 1.0100 | 1.0166 |
| Sil | RNX | Shep | kNN | NMI | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Auto- | 0.874 | 0.896 | 0.639 | 0.428 | 0.446 | 1.011 | 0.281 | 0.460 | 0.967 | 0.709 |
| std | .005 | .010 | .022 | .067 | .016 | .001 | .019 | .019 | .008 | .020 |
| Auto- | 0.858 | 0.898 | 0.806 | 0.303 | 0.509 | 1.013 | 0.229 | 0.734 | 0.900 | 0.531 |
| std | .013 | .010 | .104 | .097 | .241 | .008 | .015 | .152 | .028 | .251 |
| removed | Sil | RNX | Shep | kNN | NMI | |||||
|---|---|---|---|---|---|---|---|---|---|---|
| full | 0.988 | 0.987 | 0.968 | 0.652 | 0.308 | 1.008 | 0.700 | 0.906 | 0.973 | 0.806 |
| Difficulty | -0.007 | -0.008 | -0.140 | -0.064 | +0.007 | +0.004 | -0.013 | -0.093 | -0.007 | +0.079 |
| Neighbourhood | -0.012 | -0.010 | -0.023 | -0.088 | +0.032 | +0.006 | -0.109 | -0.022 | N | -0.101 |
| Affinity | -0.003 | -0.002 | -0.003 | -0.074 | -0.160 | +0.002 | +0.033 | +0.056 | N | N |
| Objective | -0.004 | -0.002 | -0.003 | -0.097 | -0.166 | +0.007 | +0.042 | +0.059 | N | N |
| Hyperparam-opt | N | +0.002 | +0.008 | -0.048 | -0.183 | -0.001 | -0.015 | +0.070 | N | -0.012 |
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Score | Symbol | Lens | Direction | Role |
| Spearman rank correlation | global (distance) | design | ||
| Stress-1 | global (distance) | design | ||
| Silhouette | (Sil) | global (cluster) | design | |
| Trustworthiness area | local | design | ||
| Continuity area | local | design | ||
| Median local outlier factor | outlier | design |
| Option (label) | Prompt passed to the agents |
|---|---|
| Balance local & global | optimize a balance of Trustworthiness@ , Continuity@ , Spearman, Silhouette, LOF outlier %, Stress; no single metric prioritized |
| Prioritise local | optimize Trustworthiness@ and Continuity@ above all else; others secondary |
| Prioritise global | optimize Spearman and Stress above all else; others secondary |
| Maximum Trustworthiness | optimize Trustworthiness@ above all else; others secondary |
| Maximum Continuity | optimize Continuity@ above all else; others secondary |
| Maximum Spearman | optimize Spearman above all else; others secondary |
| Dataset | Bracket | UMAP | t-SNE | UMAP Sil. | t-SNE Sil. |
|---|---|---|---|---|---|
| Iris | Easy | 0.981 | 0.985 | 0.654 | 0.640 |
| Aging Human Skin | Moderate | 0.982 | 0.985 | 0.291 | 0.223 |
| Human Kidney | Moderate | 0.991 | 0.993 | 0.333 | 0.176 |
| PBMC 3k | Moderate | 0.859 | 0.872 | 0.567 | 0.440 |
| Cortex Development | Hard | 0.972 | 0.976 | -0.023 | -0.001 |
| Fashion-MNIST | Hard | 0.975 | 0.979 | 0.188 | 0.182 |
| Bracket | Dataset | Domain | |||
| Easy | Iris | 150 | 4 | 3 | botany |
| Moderate | Aging Human Skin (scRNA-seq) | 8,000 | 50 | 13 | genomics |
| Moderate | Human Kidney (scRNA-seq) | 8,000 | 50 | 13 | genomics |
| Moderate | PBMC 3k scRNA-seq | 2,700 | 50 | 8 | genomics |
| Hard | Cortex Development (scRNA-seq) | 4,129 | 50 | 47 | genomics |
| Hard | Fashion-MNIST (sub.) | 5,000 | 784 | 10 | vision |
| method | Sil | |||||
|---|---|---|---|---|---|---|
| t-SNE (default) | 0.875 | 0.901 | 0.661 | 0.440 | 0.444 | 1.011 |
| t-SNE-S | 0.857 | 0.879 | 0.654 | 0.318 | 0.385 | 1.007 |
| UMAP-S | 0.848 | 0.883 | 0.669 | 0.481 | 0.389 | 1.014 |
| PaCMAP-S | 0.846 | 0.863 | 0.587 | 0.132 | 0.393 | 1.005 |
| Auto-S | 0.857 | 0.900 | 0.837 | 0.241 | 0.290 | 1.007 |
| RNX | kNN | Shep | Sil | NMI | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| K, t-SNE | 0.984 | 0.491 | 0.979 | 0.961 | 0.508 | 0.504 | 0.175 | 0.706 | 0.423 | 1.015 |
| K, t-SNE- | 0.985 | 0.495 | 0.981 | 0.962 | 0.525 | 0.522 | 0.180 | 0.690 | 0.415 | 1.017 |
| K, Auto- | 0.952 | 0.376 | 0.963 | 0.784 | 0.931 | 0.910 | -0.360 | 0.366 | 0.284 | 1.018 |
| C, t-SNE | 0.969 | 0.467 | 0.961 | 0.711 | 0.621 | 0.602 | -0.001 | 0.672 | 0.358 | 1.011 |
| C, t-SNE- | 0.972 | 0.466 | 0.964 | 0.625 | 0.649 | 0.726 | -0.005 | 0.666 | 0.359 | 1.012 |
| C, Auto- | 0.947 | 0.354 | 0.958 | 0.390 | 0.883 | 0.873 | -0.209 | 0.488 | 0.296 | 1.028 |
| Dataset | Backbone | Sil | |||||
|---|---|---|---|---|---|---|---|
| Iris | GPT-5.4 | 0.9583 | 0.9690 | 0.9311 | 0.4619 | 0.2166 | 1.0131 |
| Gemini-3.1-Pro | 0.9620 | 0.9709 | 0.8239 | 0.4229 | 0.3427 | 1.0139 | |
| Claude-Opus-4.8 | 0.9261 | 0.8962 | 0.7110 | 0.3483 | 0.4214 | 1.0440 | |
| Aging | GPT-5.4 | 0.5673 | 0.4992 | 0.1919 | 0.3797 | 0.9149 | 1.0148 |
| Gemini-3.1-Pro | 0.9661 | 0.9733 | 0.4250 | 0.1455 | 0.5128 | 1.0179 | |
| Claude-Opus-4.8 | 0.5293 | 0.5322 | 0.0272 | 0.6660 | 0.9774 | 1.0127 |
| Dataset | Score | Auto (w/o) | Auto- | Auto- | Auto- | Auto-Sil | Auto- |
|---|---|---|---|---|---|---|---|
| Iris | 0.0568 | 0.1201 | 0.0072 | 0.0248 | 0.3610 | 0.0016 | |
| 0.0560 | 0.1168 | 0.0021 | 0.0245 | 0.3680 | 0.0023 | ||
| 0.0636 | 0.0366 | 0.0950 | 0.2143 | 0.8029 | 0.0020 | ||
| Sil | 0.1036 | 0.2787 | 0.0497 | 0.0036 | 0.8109 | 0.0013 | |
| 0.1063 | 0.0987 | 0.1143 | 0.3324 | 0.1876 | 0.0124 | ||
| 0.0141 | 0.0700 | 0.0003 | 0.0019 | 0.0248 | 0.0008 |
| -focused | -focused | |||||
|---|---|---|---|---|---|---|
| Score | SingleLLM- | NoSkelAgents- | Auto- | SingleLLM- | NoSkelAgents- | Auto- |
| 0.8614 | 0.8433 | 0.8645 | 0.8075 | 0.8584 | 0.8570 | |
| 0.9140 | 0.8786 | 0.9056 | 0.8174 | 0.9012 | 0.9004 | |
| 0.8766 | 0.7944 | 0.8552 | 0.6750 | 0.8362 | 0.8365 | |
| 0.8175 | 0.4450 | 0.3191 | 0.5162 | 0.2904 | 0.2901 | |
| Sil | 0.3443 | 0.4240 | 0.3876 | 0.1762 | 0.2414 | 0.2409 |
| Easy–Hard agreement | Exact E, M, H agreement | |||||
|---|---|---|---|---|---|---|
| Dataset | ||||||
| PBMC3k | 0.952 | 0.987 | 0.978 | 0.45 | 0.70 | 0.53 |
| Kidney | 0.862 | 0.896 | 0.817 | 0.38 | 0.38 | 0.35 |
| Cortex | 0.914 | 0.958 | 0.952 | 0.46 | 0.52 | 0.61 |
| Fashion | 0.900 | 0.983 | 0.940 | 0.44 | 0.58 | 0.50 |
Data Visualization
Data visualization studies how visual representations help people explore, interpret, communicate, and reason about data.
See all 127 papers