Recent topic models leverage pretrained embeddings, but neural architectures produce latent representations without grounding in specific texts, and clustering-based pipelines assign representative documents only post hoc, relying on absolute distances distorted by hubness and anisotropy in high-dimensional spaces. We introduce MARETopic, a training-free framework that casts topic discovery as rank-based prototype selection. After projecting embeddings onto a low-dimensional manifold, MARETopic builds ranked lists encoding ordinal neighborhood structure. A greedy algorithm selects exactly K exemplar documents, real corpus texts, whose neighborhoods cover the corpus. Two variants share this criterion. MARETopicCorr scores candidates with a query performance predictor and a rank correlation measure, leading Purity and NMI on the two benchmarks with the most categories, ahead of both neural and clustering-based topic models. MARETopicDiff scores them with a rank-based diffusion matrix, needs neither measure, and runs 1.7 to 1.9 times faster. Without a single gradient update, MARETopic leads topic coherence on two of three datasets. A novel inter-topic Maximal Marginal Relevance step raises vocabulary diversity at little cost in coherence. Our code is available at https://github.com/thcastilho/maretopic.
Figures & tables
Figure 1: Spatial illustration of a ranked list for a query document di in an embedding space. Rank-based methods operate on this relative ordering rather than on absolute distances, providing robustness to the geometric distortions of high-dimensional spaces.
Figure 2: Overview of the MARETopic framework.
Generative
Clustering-based
Ours
Metric
LDA
NMF
ETM
ECRTM
FASTopic
BERTopic
Top2Vec
MARETopic
MARETopic
( Blei et al., 2003 )
( Lee and Seung, 1999 )
( Dieng et al., 2020 )
( Wu et al., 2023 )
( Wu et al., 2024a )
( Grootendorst, 2022 )
( Angelov, 2020 )
Diffusion
Correlation
20 Newsgroups
C v
0.5352 ± .007
0.5573 ± .016
0.4522 ± .004
0.6116 ± .013
0.5993 ± .019
0.5942 ± .007
0.6649 ± .011
0.6708 ± .015
0.6818 ± .003
NPMI
0.0472 ± .003
0.0613 ± .007
-0.0171 ± .003
-0.0372 ± .021
-0.0689 ± .002
0.0842 ± .003
0.1032 ± .003
0.0958 ± .005
0.1027 ± .002
Table 1: Topic modeling evaluation across three datasets (mean ± std of 3 runs, K=50 ). Columns are grouped by paradigm. Best value per metric per dataset in bold ; runner-up underlined . Time is end-to-end wall clock in seconds on an Apple M4 with 16 GB of memory, averaged over three runs, and lower is better.
20NG
NYT
WoS
MARETopic Corr
C v
0.6817 − 0.026
0.6449 − 0.053
0.7189 − 0.028
TD
0.8854 +0.025
0.8764 +0.016
0.8542 − 0.014
Purity
0.6161 +0.017
0.7102 +0.061
0.8094 − 0.002
NMI
0.5398 +0.031
0.3982 +0.041
0.4697 +0.002
MARETopic Diff
Table 2: Effect of UMAP pre-processing. Each cell shows the UMAP-enabled value with the delta from raw SBERT ( green : improvement; red : degradation).
Figure 3: Sensitivity to neighborhood size k across all four metrics. Solid lines: MARETopic Corr ; dashed lines: MARETopic Diff . Vertical dotted line marks the recommended k=100 .
20 Newsgroups
New York Times
Web of Science
Method
λ
C v
TD
C v
TD
C v
TD
0.0
0.6777
0.6627
0.6398
0.6845
0.7299
0.6938
0.3
0.6674 − 0.010
0.8813 +0.219
0.6384 − 0.001
0.8729 +0.188
0.7206 − 0.009
0.8600 +0.166
0.5
0.6539 − 0.014
0.9538 +0.073
0.6349 − 0.004
0.9582 +0.085
0.6850 − 0.036
0.9533 +0.093
MARETopic Correlation
0.7
0.6257 − 0.028
0.9925 +0.039
0.6043 − 0.031
0.9893 +0.031
0.6689 − 0.016
0.9902 +0.037
0.0
0.6854
0.6502
0.6516
0.7080
0.7287
0.6906
Table 3: Sensitivity to the MMR diversity parameter λ . Only C v and TD are reported, as MMR acts on word extraction and does not affect cluster assignments. Colored deltas show change relative to the previous λ ( green : improvement; red : degradation). Recommended λ=0.3 highlighted (gray).
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
20 Newsgroups
New York Times
Web of Science
Configuration
C v
TD
Purity
NMI
C v
TD
Purity
NMI
C v
TD
Purity
NMI
BERTopic, native words
0.6120
0.7666
0.4261
0.3770
0.6341
0.8067
0.6187
0.3155
0.6903
0.7707
0.7909
0.4488
+ inter-topic MMR, λ=0.3
0.5991 − 0.013
0.9382 +0.172
0.4261
0.3770
0.6445 +0.010
0.9382 +0.132
0.6187
0.3155
0.6531 − 0.037
0.9195 +0.149
0.7909
0.4488
Appendix
Table 4: The proposed diversification applied to BERTopic’s own c-TF-IDF (mean of 3 runs, K=50 ). Deltas are against BERTopic’s native word lists ( green : increase; red : decrease). At λ=0 the step reduces to selecting the highest-scoring words, which is what BERTopic does natively, and reproduces its word lists exactly on every run. Purity and NMI are unchanged by construction, since the diversification acts on word lists after Θ is fixed. This is an independent run, so BERTopic’s native values differ slightly from those in Table 1 .
Generative
Clustering-based
Ours
Metric
LDA
NMF
ETM
ECRTM
FASTopic
BERTopic
Top2Vec
MARETopic
MARETopic
Diffusion
Correlation
20 Newsgroups
C v
0.4817 − 0.053
0.5444 − 0.013
0.4508 − 0.001
0.5751 − 0.037
0.5584 − 0.041
0.5874 − 0.007
0.5995 − 0.065
0.6020 − 0.069
0.6123 − 0.070
TD
0.5678 +0.048
0.4129 − 0.096
0.8111 − 0.036
0.8098 − 0.059
0.9160 − 0.069
0.7218 − 0.051
0.8548 − 0.011
0.8413 − 0.020
0.8138 − 0.065
Appendix
Table 5: The evaluation repeated at K=100 topics (mean of 3 runs). Each cell carries the value at K=100 and, in small type, its change from K=50 ( green : increase; red : decrease). Best value per metric per dataset in bold ; runner-up underlined .
20 Newsgroups
New York Times
Web of Science
Method
Encoder
C v
TD
Purity
NMI
C v
TD
Purity
NMI
C v
TD
Purity
NMI
MARETopic Correlation
MiniLM
0.6818
0.8787
0.6227
0.5419
0.6496
0.8702
0.7084
0.3975
0.7150
0.8547
0.8106
0.4664
MPNet
0.6875
0.8653
0.6141
0.5405
0.6633
0.8840
0.7062
0.3950
0.7033
0.8715
0.8019
0.4647
DistilRoBERTa
0.6686
0.8609
0.6170
0.5326
0.6310
0.8658
0.6997
0.3874
0.7097
0.8591
0.8158
0.4753
MARETopic Diffusion
MiniLM
0.6708
0.8613
0.4144
0.4011
0.6570
0.8831
0.6046
0.3305
0.7198
0.8569
0.6082
0.3660
MPNet
0.6754
0.8707
0.4040
0.3932
0.6684
0.8649
0.6078
0.3341
0.6995
0.8800
0.5894
0.3687
Appendix
Table 6: Robustness to SBERT backbone (all-MiniLM-L6-v2, all-mpnet-base-v2, all-distilroberta-v1), with all other parameters fixed at default ( k=100 , dUMAP=5 , λ=0.3 , K=50 , mean of 3 runs). Best value per method per dataset per metric in bold .
Figure 4: Ten representative topics from MARETopic Corr on 20NG. Each panel is a leader document; bars show its top-5 c-TF-IDF terms. Topic indices follow the ordering from the greedy selection.
Figure 5: Document–topic affinity matrix Θ for MARETopic Corr on the 20NG test set ( N=7,532 , K=50 ). Rows: topic leaders sorted by dominant class, labelled by top-3 words. Columns: test documents sorted by ground-truth class; red lines mark class boundaries. Color: θij (JaccardMax affinity).
Topic modeling extracts latent themes from large text collections, but leading approaches like BERTopic face critical limitations: stochastic instability, loss of lexical precision ("Embedding Blur"), and reliance on a single data perspective. We present TriTopic, a framework that addresses these weaknesses through a tri-modal graph fusing semantic embeddings, TF-IDF, and metadata. Three core innovations drive its performance: hybrid graph construction via Mutual kNN and Shared Nearest Neighbors to eliminate noise and combat the curse of dimensionality; Consensus Leiden Clustering for reproducible, stable partitions; and Iterative Refinement that sharpens embeddings through dynamic centroid-pulling. TriTopic also replaces the "average document" concept with archetype-based topic representations defined by boundary cases rather than centers alone. In benchmarks across 20 Newsgroups, BBC News, AG News, and Arxiv, TriTopic achieves the highest NMI on every dataset (mean NMI 0.575 vs. 0.513 for BERTopic, 0.416 for NMF, 0.299 for LDA), guarantees 100% corpus coverage with 0% outliers, and is available as an open-source PyPI library.
Traditional topic models are effective at uncovering latent themes in large text collections. However, due to their reliance on bag-of-words representations, they struggle to capture semantically abstract features. While some neural variants use richer representations, they are similarly constrained by expressing topics as word lists, which limits their ability to articulate complex topics. We introduce Mechanistic Topic Models (MTMs), a class of topic models that operate on interpretable features learned by sparse autoencoders (SAEs). By defining topics over this semantically rich space, MTMs can reveal deeper conceptual themes with expressive feature descriptions. Moreover, uniquely among topic models, MTMs enable controllable text generation using topic steering vectors. To properly evaluate MTM topics against word list approaches, we propose topic judge, an LLM-based pairwise comparison evaluation framework. Across eight datasets, MTMs match or exceed traditional and neural baselines on coherence metrics, are consistently preferred by topic judge, and enable effective LLM steering.
Carolina Zheng, Nicolas Beltran-Velez, Sweta Karlekar +5
Columbia University · Independent · Google Research
Existing topic modeling methods, from LDA to recent neural and LLM-based approaches, which focus mainly on statistical coherence, often produce redundant or off-target topics that miss the user's underlying intent. We introduce Human-centric Topic Modeling, Human-TM), a novel task formulation that integrates a human-provided goal directly into the topic modeling process to produce interpretable, diverse and goal-oriented topics. To tackle this challenge, we propose the Goal-prompted Contrastive Topic Model with Optimal Transport (GCTM-OT), which first uses LLM-based prompting to extract goal candidates from documents, then incorporates these into semantic-aware contrastive learning via optimal transport for topic discovery. Experimental results on three public subreddit datasets show that GCTM-OT outperforms state-of-the-art baselines in topic coherence and diversity while significantly improving alignment with human-provided goals, paving the way for more human-centric topic discovery systems.
Rui Wang, Yi Zheng, Dongxin Wang +5
School of Computer Science, Nanjing University of Posts and Telecommunications · School of Computer Science and Information Engineering, Hefei University of Technology · School of Electronic Engineering and Computer Science, Queen Mary University of London +1