Topic Models Drop the Word List: Directions in LLM Activations Track Short and Abstract Texts Better

A topic is no longer described by a list of words but by interpretable directions inside a large language model's activation space, so topics can name abstract qualities such as style, tone, and discourse patterns that word co-occurrence misses. Earlier topic models treat documents as bag-of-words counts, and neural variants that use pretrained embeddings still express each topic as a list of words weighted by importance. Across eight datasets, the topic judge gives an MTM the highest score in 14 of 16 dataset-representation rows, and MTMs beat their word-based counterparts in 42 of 48 pairwise comparisons.

The pipeline turns each document into counts of sparse-autoencoder features, then lets a topic model learn topics and per-document topic proportions over those features. A sparse autoencoder is a single-layer network that reconstructs and sparsifies LLM activations; its decoder columns are the feature directions, and each direction can be labeled with an automatically generated text description. Experiments use the 16k sparse autoencoder from layer 10 of Gemma 2-9B, feature descriptions from Neuronpedia, and K = 50 topics on the smallest datasets and K = 100 on the rest.

A learned topic also yields a steering vector that modulates how strongly that topic appears in generated text. The steering vector weights feature directions by their topic weights and normalizes them; generation removes the activation component parallel to that direction and adds back a scaled steering vector. Topic relevance win rates exceed 84% on all eight datasets and reach 99% on Bills, while perplexity of steered generations stays close to the unsteered baseline.

The report covers adapting topic models to sparse-autoencoder features, plus the topic judge, standard coherence and diversity metrics, a topic novelty analysis, and steering experiments. It is aimed at researchers and engineers who explore text collections, describe topics, or control generation by topic. The source proposes training or finetuning sparse autoencoders on the target corpus as a mitigation, but reports no experiment for that direction.

The source states that topic quality is sensitive to the choice of sparse-autoencoder layer, the data used to train it, and the accuracy of automatically generated feature descriptions, so topics may be poor if features are not relevant or labels are inaccurate. The preprocessing pipeline is heavier than for word-based models, needs a GPU able to hold the LLM and the sparse autoencoder, and feature filtering adds another LLM call. On Bills with the TopFeatures representation, all three word-based models outscore their mechanistic counterparts, which the authors attribute to the judge finding feature descriptions less specific than keywords; mBERTopic learns overlapping topics on PoemSum and WritingPrompts, with diversity dropping to 0.38 and 0.36.