Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Mechanistic Topic Models learn themes from LLM internals, not words.

desk verdict Genuinely useful SAE-feature topic models with an honest evaluation of topic quality, but the steering claim is built on a closed loop and needs a human check before it can be taken at face value. read the letter →

arxiv 2507.23220 v2 pith:4JYXTXLY submitted 2025-07-31 cs.CL cs.LG

Carolina Zheng , Nicolas Beltran-Velez , Sweta Karlekar , Claudia Shi , Achille Nazaret , Asif Mallik , Amir Feder , David M. Blei This is my paper
classification cs.CLcs.LG
keywords mechanistictopicmodelssparseautoencodersSAEfeaturesmodelingLLMsteeringvectorsjudgecontrollabletextgenerationinterpretabledirections
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

What this paper is trying to establish is that topic models do not have to describe themes as word lists: by representing documents as counts of interpretable sparse-autoencoder features extracted from a language model's internal activations, a topic can be a concept, or even a steering direction, rather than a bag of words. The paper introduces Mechanistic Topic Models (MTMs), three adaptations of LDA, ETM, and BERTopic to this feature space, and an LLM-based pairwise evaluation framework called topic judge. Across eight datasets, MTMs match or beat word-based baselines on coherence and diversity, are preferred by both LLM and human judges in most comparisons, and, uniquely among topic models, let a discovered topic be turned into a steering vector that biases generation toward that theme while keeping perplexity near baseline. If this holds, the practical consequence is that abstract semantics that word co-occurrence misses, such as tone, discourse, and sentiment, become first-class citizens of topic modeling, and the same topics that describe a corpus can directly control the text an LLM produces.

What carries the argument

SAE feature count vectors, defined by thresholding per-feature activation at the 80th percentile of the feature's training distribution, replace document word counts everywhere. These counts feed three adapted models: mLDA (multinomial Dirichlet over features), mETM (features drawn binomially from softmax-transformed topic activations), and mBERTopic (document embeddings formed by summing feature directions, then UMAP/HDBSCAN clustering with class-based TF-IDF). Two quality-control stages keep the vocabulary reliable: an LLM filters out spurious or topic-irrelevant features before training, and an optional LLM step prunes mislabeled features after training. The steering vector is defined as $\mathbf{s}_k = \sum_i \beta_{k,i}\mathbf{w}_i / \|\sum_i \beta_{k,i}\mathbf{w}_i\|_2$, and the intervention removes the parallel component of the activation along the steering direction before adding $\lambda\mathbf{s}_k$.

What would settle it

Run MTM and a word-based baseline on a corpus whose domain is far from the SAE's training distribution, such as dense medical records or specialist legal filings, and have human readers judge which topic sets better describe each document; the claimed generality fails if humans do not prefer MTM topics there. As a sharper check, randomly shuffle or corrupt the SAE feature descriptions and show that topic-judge preference collapses, which would confirm the descriptions are load-bearing rather than incidental to the pipeline.

Watch

Extended reading notes

Core claim

Mechanistic Topic Models claim that the right vocabulary for topic modeling is not words but the sparse, human-interpretable feature directions that sparse autoencoders recover from an LLM's activations. Documents are featurized by counting how often each feature activates above its 80th-percentile threshold, then passed to standard topic-model machinery: a multinomial LDA variant, an embedded-topic variant with logistic-normal priors, and a clustering variant with class-based TF-IDF. Topics are described by the automatically generated text labels of their top features, and each topic yields a steering vector, the normalized, weight-scaled sum of its feature directions. The paper reports that an MTM variant received the highest topic-judge Elo in 14 of 16 dataset–representation rows, MTMs beat their word-based counterparts in 42 of 48 pairwise comparisons, the advantage is largest on short and abstract datasets (about 195 Elo points on average), and steering achieves topic-relevance win rates above 84% with perplexity close to unsteered generations.

Load-bearing premise

The entire approach assumes that the pretrained SAE feature directions and their auto-generated text labels are a semantically faithful vocabulary for the corpus being modeled; if those labels or directions are wrong or irrelevant to the domain, topic descriptions and steering vectors inherit the error.

Editorial extensions

If this is right

  • MTMs let a single model output both a topic description and a topic steering vector, so the discovered themes can directly control what an LLM writes; the paper reports win rates above 84% and near-baseline perplexity.
  • Abstract corpora are where MTMs gain most: on emotion, poetry, and writing-prompt data the margin over word-based models is about 195 Elo points.
  • Topic judge supplies a label-free way to compare topic models across different vocabularies, and its LLM Elo agrees strongly with human Elo (Spearman $\rho$ between 0.76 and 0.95).
  • Standard metrics are not hurt: MTMs are roughly on par with LDA, ETM, and BERTopic on coherence and diversity, and are no less stable across random seeds.
  • SAE feature spaces emerge as a reusable semantic representation for unsupervised tasks beyond interpretability, such as featurizing documents for clustering and generation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Corpus-specific SAE training, which the paper floats as a mitigation, would likely sharpen both topic descriptions and steering vectors without changing any MTM algorithm, an extension the paper leaves to future work.
  • Feature-layer and SAE-width choices interact with dataset genre; the paper fixes one layer, but a principled layer-selection rule could push MTM gains onto the benchmark datasets where it currently only ties.
  • Because steering vectors live in the same activation space, combinations of topic vectors could permit compositional control, e.g., 'legal + emotion', a capability the paper does not test.
  • The topic judge protocol is not tied to MTMs; the same pairwise LLM-judge setup could rank any two document-summarization representations, which would be a broader evaluation tool than the paper claims.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This paper introduces Mechanistic Topic Models (MTMs), a family of topic models that operate on SAE feature counts rather than word counts or raw embeddings. Three instantiations are proposed, mLDA, mETM, and mBERTopic, adapting LDA, ETM, and BERTopic to SAE feature representations. The paper also introduces 'topic judge,' an LLM-based pairwise comparison framework for evaluating topic models across vocabularies, and demonstrates topic steering via SAE feature directions. Experiments across eight datasets report coherence, topic diversity, alignment, topic-judge Elo scores, and steering win rates, with a human study validating the LLM judge on two datasets.

Significance. If the paper's claims hold, MTMs constitute a genuinely new class of topic models that leverage SAE feature spaces, with the distinctive capability of topic-based controlled generation. The paper is methodologically transparent: it provides detailed algorithms, reports bootstrap confidence intervals for Elo scores, uses t-tests for coherence differences, and commits to releasing code and data. The topic judge framework, with its human validation, is a potentially reusable evaluation tool. The main weakness is that the steering evaluation is self-referential and the fluency claim is unsupported at the operating strengths, so the paper's most distinctive contribution requires additional evidence.

major comments (3)
  1. [Section 5.4, Eq. (16), Eq. (20), Figure 2, Table 4] The TWR evaluation is a closed loop: the topic summary shown to the judge is an LLM summarization of the top SAE feature descriptions, and the steered texts are generated by adding the same SAE feature directions, so a high TWR only confirms that steering activates the features whose descriptions appear in the prompt, not that the generated text is topically relevant as a human would judge it. The likelihood-difference metric in Eq. (20) is similarly self-referential because the 'on-topic' document sets Dk are selected using the same model's document-topic proportions θd. The human study in Section 5.1 (Figure 2) validates topic descriptions, not steering, so the abstract's claim of 'effective LLM steering' is not yet supported by independent evidence.
  2. [Section 5.4.1, Tables 4 and 10] The statement that steering 'preserves the fluency of generated text' is not supported at the steering strengths responsible for most wins. Table 10 shows that λ∈{30,40,50} account for roughly 80% of mLDA TWR wins, while Table 4 reports perplexities at these strengths of 7.5–20.1 versus a 6.2–6.3 baseline, i.e., up to a 220% increase; no significance test or human fluency rating is provided, so 'close to baseline' is an unverified claim.
  3. [Abstract and Table 2] The claim that MTMs 'match or exceed' baselines on coherence metrics is not consistently supported by the data. In Table 2, word-based BERTopic achieves the highest rating on 20NG, Wiki, Yelp, AGNews, and GoEmotions, and the paper itself concedes in Section 5.2 that 'word-based models outperform MTMs in some cases'; the abstract should be qualified to reflect the mixed evidence rather than claiming uniform match-or-exceed.
minor comments (6)
  1. [Abstract] There is a typo in the abstract: 'proposestopic judge' should be 'proposes topic judge'.
  2. [Section 4.1, Eq. (5)] The 80th percentile threshold qi is a free parameter, but no sensitivity analysis is reported over its value; a brief robustness check would strengthen the paper.
  3. [Appendix A.2 and Figure 10(i)] The post-training refinement is claimed to remove mislabeled features, but Figure 10(i) shows a failure case where spurious features persist despite refinement; the paper should quantify the frequency of such failures across datasets.
  4. [Section 5.3 and Figure 3] The topic novelty analysis relies on visual inspection of correlation heatmaps; the paper should report the fraction of topics with correlations below 0.5 and 0.2 numerically for all datasets.
  5. [Related Work] The paper cites concurrent work by Girrbach and Akata (2026) viewing SAEs as topic models but does not describe how MTMs differ in their objectives or outputs; a brief comparison would clarify the novelty.
  6. [Section 5.1] The human validation of topic judge covers only Wiki and PoemSum; the paper should state this coverage limitation when claiming that the LLM judge is a reliable proxy for human preferences.

Circularity Check

2 steps flagged · score 6.0 of 10

Steering evaluation is self-referential: the LLM judge sees summaries of the same SAE features used to build the steering vectors, and the likelihood-difference ground truth comes from the model's own topic assignments.

  1. self definitional [Section 5.4, 'Topic Relevance Win Rate'; Eq. 16; Section 4.3; Appendix F.4]
    "We then construct the textual topic representation tk from the automatically generated descriptions di corresponding to these n features. An LLM judge then selects the text most representative of the topic according to its summarized description tk."

    The topic summary tk is built from the descriptions of the top SAE features in the topic-feature vector βk, and the steering vector sk is constructed from exactly those same feature directions: Eq. 16 sets sk to the normalized weighted sum of wi with weights βk,i. The TWR prompt in Appendix F.4 asks the judge to match generated texts to tk, so a win only demonstrates that adding sk makes the model produce text that activates the features whose descriptions appear in the prompt. The metric therefore reduces to a self-consistency check between the steering intervention and the feature vocabulary that defines the topic, rather than an independent test of topical relevance. No human evaluation of steered outputs is reported.

  2. self definitional [Section 5.4, 'Topic Likelihood Difference'; Eq. 20; Appendix E.3]
    "Formally, let Dk contain documents highly associated with topic k and D−k be an equally sized random sample from the set of documents highly associated with another topic (see Section E.3)."

    Appendix E.3 defines 'highly associated' using the MTM's own document-topic proportions: Dk = {d : θd,k ≥ τ} for threshold τ = 0.5. Thus the on-topic and off-topic document sets are selected by the very model whose topic-feature weights βk and steering vector sk are being evaluated. The likelihood difference in Eq. 20 then measures whether sk increases the likelihood of documents the model already assigned to topic k, so the ground-truth topic labels are generated by the method under test. This is a closed evaluation loop, though not a strict mathematical identity because the steering could still fail to shift likelihoods.

full rationale

The main topic-quality evaluation is not circular: it uses standard coherence metrics (ratings and intrusion) and topic judge, and the LLM judge is validated against a human study on Wiki and PoemSum with Spearman rho 0.76-0.95. No fitted parameter is renamed as a prediction, and there is no load-bearing self-citation chain or imported uniqueness theorem. The circularity is localized to Section 5.4, where the abstract's claim that MTMs 'enable effective LLM steering' is evaluated self-referentially: the judge is given topic summaries built from the same SAE feature descriptions that define the steering vectors, and the likelihood-difference experiment selects on-topic documents using the model's own θd. These two loops make the steering claim partially reduce to the model's own feature vocabulary and topic assignments, so the paper is partially circular rather than wholly so.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The method does not invent new physical or conceptual objects; it repurposes existing SAE features. The load-bearing input that the reader did not pay for upstream is the quality of the pretrained SAE and its labels, plus several hand-set thresholds that shape every count vector. The free parameters are mostly evaluation and preprocessing constants rather than fitted scientific constants, but they are numerous and not covered by sensitivity analysis.

free parameters (7)
  • Threshold quantile q_i = 0.80 (80th percentile)
    Equation 5 counts a feature as active when its activation exceeds its 80th percentile on the SAE training data; the quantile is chosen without sensitivity analysis and directly shapes all feature count vectors.
  • Feature filtering thresholds = 1% activation frequency; 90% document frequency; description categories (programming, math, grammar, formatting, stop…
    Appendix A.1 removes features using three hand-set criteria; the paper says filtering is 'crucial' because without it the models do poorly, so the central results depend on these cutoffs.
  • Topic refinement counts n=10, m=2 = n=10, m=2
    Appendix A.2 keeps 10 features and lets an LLM delete up to 2; this post-processing is applied to all models and influences coherence and topic judge results.
  • Topic judge parameters (q, p, T) = q=2, p=0.75, T=100
    Section G: pairwise judge shows up to 2 topics per document, cumulative mass threshold 0.75, 100 comparisons per model pair; changing these likely changes Elo rankings.
  • Number of topics K = 50 for AGNews, GoEmotions, PoemSum; 100 otherwise
    Section 5 Setup: K is set by dataset size, not derived from data or theory.
  • SAE layer and width = Gemma 2-9B layer 10, 16k features
    Section 5 Setup: layer 10 is chosen because it was 'found empirically that it provides the most useful features'; no search over layers is reported.
  • Bayesian optimization hyperparameters = per-model ranges in Appendix D
    25 iterations per model-dataset pair optimize NPMI coherence times topic diversity; these fitted hyperparameters can inflate later coherence figures.
assumptions (5)
  • domain assumption Linear Representation Hypothesis: LLM activations decompose as sparse sums of near-orthogonal, human-interpretable linear directions (Definition 1).
    Section 3 defines LRH and uses it to justify treating SAE feature directions as topic atoms; if many concepts are not linearly represented, feature descriptions and steering vectors lose meaning.
  • domain assumption GemmaScope SAE features and Neuronpedia auto-generated descriptions are accurate enough to serve as a topic vocabulary.
    Section 5 and Appendix A rely on downloaded SAE metadata and descriptions; mislabeled features are acknowledged and partially fixed by refinement, but no independent audit of the feature space is provided.
  • domain assumption The multinomial (mLDA) and binomial (mETM) count models approximate the distribution of thresholded SAE activation counts.
    Equations 8 and 12 impose specific likelihoods on features that activate at most once per token (mETM) or ignore activation magnitude (mLDA); these modeling assumptions are carried over from word-based topic models.
  • domain assumption GPT-4.1 pairwise judgments are a reliable proxy for human preferences about topic-document fit.
    Section 5.1 builds the topic judge method on this premise; a 68-participant human study on Wiki and PoemSum shows Spearman rho 0.76-0.95, but the proxy is not validated on the other six datasets or on steering evaluations.
  • standard math UMAP and HDBSCAN recover meaningful topic clusters from SAE-count document embeddings (mBERTopic).
    Equation 14 constructs document embeddings, then Section 4.2.3 applies UMAP/HDBSCAN; the method inherits the clustering algorithm's assumptions about manifold structure and density.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders." pith.science (2026). https://pith.science/paper/4JYXTXLY

@misc{pith2026250723220,
  author       = {Pith},
  title        = {Pith review of: Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4JYXTXLY}},
  note         = {Machine review of arXiv:2507.23220}
}
read the original abstract

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 \textit{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.

Figures

Figures reproduced from arXiv: 2507.23220 by the authors.

Figure 6
Figure 6. The features identified as spurious in topic [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 5
Figure 5. Screenshot of the evaluation interface shown [PITH_FULL_IMAGE:figures/full_fig_p022_5.png] view at source ↗
Figure 7
Figure 7. Topic cross-correlation heatmaps (mLDA vs. LDA) for the three additional datasets. Conventions as in [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figures from the paper (3 more)
Figure 8
Figure 8. Figure 8: Document log-likelihood difference ∆ℓ(λ) for mETM (above) and mBERTopic (below). For mETM, ∆ℓ generally increases as λ increases, confirming bias toward the target topic. In contrast, mBERTopic does not always show positive ∆ℓ shifts. We think this may be because it is…
Figure 9
Figure 9. Figure 9: Document log-likelihood difference ∆ℓ(λ) for mLDA on AGNews, Yelp, and WritingPrompts. As in [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: Illustrative examples of topic sets from MTMs (mLDA, mETM, mBERTopic) across diverse datasets. [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Stories in Space: In-Context Learning Trajectories in Conceptual Belief Space

    cs.CL 2026-05 unverdicted novelty 6.0 of 10

    LLMs perform in-context learning as trajectories through a structured low-dimensional conceptual belief space, with the structure visible in both behavior and internal representations and causally manipulable via inte...

Reference graph

Works this paper leans on

22 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [1]

    For each topic k, retrieve the top n+m fea- tures by weight (wheremis small, e.g., 2)

    work page

  2. [2]

    Valérie Costa, Thomas Fel, Ekdeep Singh Lubana, Bahareh Tolooshams, and Demba Ba

    Chatbot Arena: An open platform for evaluating LLMs by human preference.arXiv preprint arXiv:2403.04132. Valérie Costa, Thomas Fel, Ekdeep Singh Lubana, Bahareh Tolooshams, and Demba Ba. 2025. From flat to hierarchical: Extracting sparse repre- sentations with matching pursuit.arXiv preprint arXiv:2506.03093. Hoagy Cunningham, Aidan Ewart, Logan Riggs, Ro...

    arXiv 2025

  3. [3]

    InProceedings of the Conference on Language Modeling

    Inspecting and editing knowledge repre- sentations in language models. InProceedings of the Conference on Language Modeling. Matthew Honnibal, Ines Montani, Sofie Van Lan- deghem, and Adriane Boyd. 2020. spaCy: Industrial-strength natural language processing in Python. Alexander M. Hoyle, Pranav Goel, Andrew Hian- Cheong, Denis Peskov, Jordan L. Boyd-Grab...

    work page 2020

  4. [4]

    Qintong Li, Leyang Cui, Lingpeng Kong, and Wei Bi

    LLMs-as-judges: A comprehensive sur- vey on LLM-based evaluation methods.arXiv preprint arXiv:2412.05579. Qintong Li, Leyang Cui, Lingpeng Kong, and Wei Bi. 2025. Exploring the reliability of large lan- guage models as customized evaluators for di- verse NLP tasks. InProceedings of the Interna- tional Conference on Computational Linguistics, pages 10325–1...

    arXiv 2025

  5. [8]

    InProceedings of the Associa- tion for Computational Linguistics, pages 15504– 15522

    Steering Llama 2 via contrastive activa- tion addition. InProceedings of the Associa- tion for Computational Linguistics, pages 15504– 15522. Mansi Sakarvadia, Aswathy Ajith, Arham Khan, Daniel Grzenda, Nathaniel Hudson, André Bauer, Kyle Chard, and Ian Foster. 2023. Mem- ory injections: Correcting multi-hop reasoning failures during inference in transfor...

    work page 2023

  6. [10]

    Gemma 2: Improving open language models at a practical size.arXiv preprint arXiv:2408.00118. Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, Brian Chen, Adam Pearce, Craig Citro, Emmanuel Ameisen, Andy Jones, Hoagy Cunningham, Nicholas L Turner, Callum McDougall, Monte MacDiarmid, C. Daniel Freeman, Theodore R. Sumers, Ed- war...

    arXiv

  7. [11]

    Tech- nical blog post

    Scaling monosemanticity: Extracting in- terpretable features from Claude 3 Sonnet. Tech- nical blog post. Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J. Vazquez, Ulisse Mini, and Monte MacDiarmid. 2023. Steer- ing language models with activation engineering. arXiv preprint arXiv:2308.10248. Felipe Viegas, Antonio Pereira, Washing...

    arXiv 2023

  8. [12]

    Tianlong Wang, Xianfeng Jiao, Yinghao Zhu, Zhongzhi Chen, Yifan He, Xu Chu, Junyi Gao, Yasha Wang, and Liantao Ma

    Exploiting contextual embeddings in hi- erarchical topic modeling and investigating the limits of the current evaluation metrics.Compu- tational Linguistics, 51(3):843–883. Tianlong Wang, Xianfeng Jiao, Yinghao Zhu, Zhongzhi Chen, Yifan He, Xu Chu, Junyi Gao, Yasha Wang, and Liantao Ma. 2025. Adaptive activation steering: A tuning-free LLM truthful- ness ...

    work page 2025

Show all 22 references
  1. [14]

    Prompt an LLM to identify and remove up to m features that appear irrelevant or mislabeled relative to the other features

  2. [15]

    New York

    Retain the resulting top n features as the final topic representation. Automated refinement is cost-effective since the number of topics and features per topic is typically small. The specific prompt used for this task is provided in Section F.1. We set n= 10 and m= 2 in our e...

    2020

  3. [16]

    Do not use a full sentence or a complete thought

  4. [17]

    Use your world knowledge to help you decide what the topic is about

  5. [18]

    In particular, do not rely on lists in your response, or include specifics that only pertain to a few items in the topic

    The summary should be general, capturing the commonalities of the items as a single main theme. In particular, do not rely on lists in your response, or include specifics that only pertain to a few items in the topic

  6. [19]

    What is this group of [word | feature]s about?

    If unsure, err on the side of being more general rather than too specific in your summary. User prompt.Topic: [topic] F.3 Ratings and Intrusion System prompt.The goal of this task is to evaluate a list of [word | feature]s produced by an automatic method. We call this list a "...

    1965

  7. [20]

    Compute correlations:For models m and m′ with K topics each, compute the K× K correlation matrix C where Ci,j = corr(θm :,i, θm′ :,j )

  8. [21]

    Greedy matching:Starting with all topics unmatched, iteratively: • Find the column j∗ with highest maximum across all rows: j∗ = arg maxj∈remaining maxi Ci,j • Among remaining rows, find the best match for this column: i∗ = arg maxi∈remaining Ci,j∗ • Add pair (i∗, j∗) to align...

  9. [22]

    Legal terminology and concepts surrounding property rights, ownership, environmental conservation, and financial transactions in real estate and land management

    Visualize:Reorder topics by alignment order to create the heatmap, with strongest column matches appearing first along the diagonal. High correlations along the diagonal indicate shared concepts between models, while low- correlation regions (dashed boxes) reveal novel topics ...

    1924

  10. [2013]

    InProceedings of the North American Chapter of the Association for Computational Linguistics, pages 746–751

    Linguistic regularities in continuous space word representations. InProceedings of the North American Chapter of the Association for Computational Linguistics, pages 746–751. Rajiv Movva, Kenny Peng, Nikhil Garg, Jon Klein- berg, and Emma Pierson. 2025. Sparse autoen- coders f...

    2025

  11. [2017]

    Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger

    hdbscan: Hierarchical density based clustering.Journal of Open Source Software, 2(11):205. Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. 2018. UMAP: Uniform man- ifold approximation and projection.Journal of Open Source Software, 3(29):861. Stephen Merity, ...

    2018

  12. [2023]

    Are the current topic modeling evaluation metrics enough?

    Automated evaluation of written dis- course coherence using GPT-4. InProceedings of the Workshop on Innovative Use of NLP for Building Educational Applications, pages 394– 403. David Newman, Jey Han Lau, Karl Grieser, and Timothy Baldwin. 2010. Automatic evaluation of topic co...

    2010 arXiv

  13. [2024]

    InAd- vances in Neural Information Processing Sys- tems, volume 37, pages 49519–49551

    Personalized steering of large language models: Versatile steering vectors through bi- directional preference optimization. InAd- vances in Neural Information Processing Sys- tems, volume 37, pages 49519–49551. Jonathan D. Chang, Jordan L. Boyd-Graber, Sean Gerrish, Chong Wang...

    2009

  14. [2025]

    Techni- cal blog post

    Negative results for SAEs on downstream tasks and deprioritising SAE research (GDM mech interp team progress update #2). Techni- cal blog post. Dominik Stammbach, Vilém Zouhar, Alexander Hoyle, Mrinmaya Sachan, and Elliott Ash. 2023. Revisiting automated topic model evaluation...

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.