Sparse Autoencoders for Hypothesis Generation

arXiv:2502.04382v3 [cs.CL] 6 Jun 2025 Sparse Autoencoders for Hypothesis Generation Rajiv Movva * 1 Kenny Peng * 2 3 Nikhil Garg 3 Jon Kleinberg 2 Emma Pierson 1 Abstract We describe HYPOTHESAES, a general method to hypothesize interpretable relationships between text data (e.g., headlines) and a target variable (e.g., clicks). HYPOTHESAES has three steps: (1) train a sparse autoencoder on text embeddings to produce interpretable features describing the data distribution, (2) select features that predict the target variable, and (3) generate a natural lan- guage interpretation of each feature (e.g., men- tions being surprised or shocked) using an LLM. Each interpretation serves as a hypothesis about what predicts the target variable. Compared to baselines, our method better identifies reference hypotheses on synthetic datasets (at least +0.06 in F1) and produces more predictive hypotheses on real datasets (∼twice as many significant find- ings), despite requiring 1-2 orders of magnitude less compute than recent LLM-based methods. HYPOTHESAES also produces novel discoveries on two well-studied tasks: explaining partisan dif- ferences in Congressional speeches and identify- ing drivers of engagement with online headlines. 1. Introduction Large language models (LLMs) show promise as a tool for hypothesis generation. Discovering relationships between text data and a target variable is an important and funda- mental task with diverse applications in economics, political science, sociology, medicine, and business (Grimmer, 2010; Rathje et al., 2021; Sun et al., 2022; Gentzkow & Shapiro, 2010; Monroe et al., 2009; Nelson, 2020; Ranard et al., 2016; Ting et al., 2017). What features of a restaurant re- view predict a low rating? What features of a social media post predict whether it will go viral? What features of a patient’s clinical notes predict if they will develop cancer? *Equal contribution 1UC Berkeley 2Cornell University 3Cornell Tech. Correspondence to: Rajiv Movva <rmovva@berkeley.edu>, Kenny Peng <kennypeng@cs.cornell.edu>. Proceedings of the 42 nd International Conference on Machine Learning, Vancouver, Canada. PMLR 267, 2025. Copyright 2025 by the author(s). Automated approaches to answering these questions have the potential to significantly expand and accelerate scientific discovery (Ludwig & Mullainathan, 2023). Indeed, multiple lines of work—spanning decades of research—have sought to extract text features that can predict a target variable (Blei et al., 2003; Monroe et al., 2009). Recent LLM-based hypothesis generation methods are especially exciting since they operate directly at the level of human-interpretable natural language concepts (Zhou et al., 2024; Batista & Ross, 2024). Concretely, hypothesis generation methods take a dataset of texts linked to a target variable (e.g., headlines and their engagement level) and hypothesize natural language concepts that predict the target variable (e.g., mentions being surprised or shocked or asks a question to the reader). However, basic hurdles impede the full usage of LLMs for hypothesis generation. Consider two natural approaches. The first prompts an LLM with positive and negative ex- amples and asks it to hypothesize concepts that distinguish them (Zhou et al., 2024; Batista & Ross, 2024; Ludan et al., 2024). However, due to context window and reasoning limi- tations, we can only give LLMs a few training examples at a time, making this hard to scale. A second approach aims to interpret a model fine-tuned to predict the target variable (Zhong et al., 2024), leveraging the full training dataset. However, neurons are often hard to interpret (Elhage et al., 2022). In summary, efficient and scalable learning of statis- tical relationships requires numerical representations of text (e.g., neurons), but interpreting neurons is hard. Our first contribution is a theoretical framework that con- nects model interpretation and hypothesis generation, clari- fying the challenge above. Proposition 3.1 gives a “triangle inequality” for hypothesis generation: if a neuron is predic- tive of the target variable, then a sufficiently high-fidelity interpretation of the neuron is also predictive. This result di- rectly motivates a general hypothesis generation procedure: 1. (Feature generation) Learn interpretable neurons (i.e., that tend to fire in the presence of human concepts). 2. (Feature selection) Select neurons that predict the target variable. 3. (Feature interpretation) Generate high-fidelity natural language interpretations of these neurons. 1

Sparse Autoencoders for Hypothesis Generationactive neurons per input. Sparsity encourages interpretability. k 1. Feature Generation Train sparse autoencoder on text embeddings. Hidden neurons serve as interpretable features. 2. Feature Selection Find neurons with activations that predict the target variable 3. Feature Interpretation Find natural language concept that correlates with neuron activation Given texts (e.g., headlines) labeled with target variable (e.g., clicks), hypothesize interpretable features that predict target variable. High-Activation Examples • “They Went To Rescue A Hurt Little Dog Hiding In A Bush. What They Found Was A Big Surprise.” • “Over 20 Years Ago, A Little Girl Stunned People From Around The World With 5 Minutes Of WOW” Low-Activation Examples • “I Dare You To Kick Off Those Mental Training Wheels And Make Your Next Bike Ride This Inspiring” • “The Cable Package You Chose To Use Might All Be An Elaborate Illusion” “mentions being surprised or shocked” LLM Hypothesis Sampled Texts for Interpretation Predictive neurons Sparse Hidden Layer Input Layer Output Layer “I Was Surprised At The Largest Buyer Of Elections In My State” Text Input Pre-Trained Embedding Model Figure 1. HYPOTHESAES is a general method that outputs natural language hypotheses from text datasets. Our second contribution is a method, HYPOTHESAES, that successfully implements this procedure. A primary chal- lenge is learning the interpretable neurons in step 1—after all, a long line of work has established the challenge of inter- preting neural networks (Olah et al., 2017; Kim et al., 2018; Elhage et al., 2022). To overcome this, we train a sparse autoencoder (SAE; Makhzani & Frey (2014)) on pre-trained text embeddings from our dataset. Neurons in the SAE hidden layer are highly interpretable, while retaining the explanatory power of embeddings (Cunningham et al., 2023; Bricken et al., 2023; Templeton et al., 2024; Gao et al., 2024; O’Neill et al., 2024). Intuitively, this architecture leverages the fact that natural language is sparse; only a small fraction of all human concepts are expressed at a time. In step 2, we select neurons that are predictive of the target variable (e.g., using Lasso). In step 3, we automatically generate high-fidelity interpretations of these predictive neurons by prompting an LLM with texts that activate the neuron, build- ing on recent work in autointerpretability (Bills et al., 2023). These interpretations serve as hypotheses. We evaluate HYPOTHESAES against recent state-of-the- art methods. We consider three synthetic tasks, as well as three real-world tasks of practical interest: hypothesizing the relationship between headlines and engagement, speech text and political party, and review text and rating. Example HYPOTHESAES hypotheses are given below: Task Example Hypotheses (Abbreviated) Headline Engagement (+) mentions being surprised or shocked (-) asks a question directly to the reader Speaker Party (Rep.) discusses illegal immigration (Dem.) criticizes tax breaks for the wealthy Restaurant Rating (+) mentions plans to return to the restaurant (-) mentions issues related to food safety HYPOTHESAES produces many more significant hypothe- ses than three baseline methods. On three real-world tasks, 45/60 hypotheses generated by our method are significant, compared to at most 24 for the baselines. On two datasets, we identify new hypotheses which previous detailed analy- ses have not uncovered. Intuitively, by decoupling feature generation from the prediction task, HYPOTHESAES is able to explore a wider range of hypotheses. HYPOTHESAES is also more than 10× faster and cheaper than recent LLM baselines, with similar runtime and cost to BERTOPIC, a standard embedding-based topic model. Filtering hypotheses based on the predictiveness of neu- rons is significantly cheaper than directly evaluating natural language concepts, since all neuron activations of an input can be computed from a single forward pass of an SAE (in comparison to one LLM call per natural language feature). 2. Related Work 2.1. Hypothesis Generation Classical text analysis methods, such as comparing n- gram frequencies between groups (Fightin’ Words; Monroe et al. (2009)) or topic modeling (LDA, Blei et al. (2003); BERTopic, Grootendorst (2022)) remain standard tools to discover relationships between text and target variables (Grimmer et al., 2022). A challenge with these methods is that their outputs—lists of words or documents—are not immediately interpretable by humans (Chang et al., 2009). As a result, applying these methods for scientific discovery requires sufficient domain expertise to prespecify hypothe- ses and assess whether the data support them (Demszky et al., 2019; Gentzkow et al., 2016; Sun et al., 2022). Recently, several LLM-based methods aim to automate inter- 2

Sparse Autoencoders for Hypothesis Generation ahead of the 24, 23, and 20 for NLPARAM, BERTOPIC, and HYPOGENIC respectively. When evaluating overall predictive performance, HYPOTHE- SAES beats baselines on 8 of 9 comparisons, demonstrating strongly predictive hypotheses. The exception is that HY- POGENIC achieves 2% higher R2 on Yelp. However, HY- POGENIC’s most predictive hypotheses, such as expresses disappointment with the overall dining experience, essen- tially restate the rating prediction task. While most LLMs can perform such a prediction, these hypotheses do not reveal specific insights about the task. In contrast, our hy- potheses—in Table 6 for YELP—tend to be more specific: we find that negative reviews mention food poisoning, rude and aggressive staff, or dishonest business practices. HYPOTHESAES’s outputs are rated as qualitatively helpful and interpretable. The results of our qualitative evaluation (Figure 2) confirm the above intuition: most of the predictive concepts generated by HYPOTHESAES are rated helpful (24 out of 30) and interpretable (29 out of 30)5. Several hypotheses from NLPARAM and HYPOGENIC, de- spite being predictive, do not pass these checks. The total count of hypotheses that are predictive, helpful, and inter- pretable generated by HYPOTHESAES (23) far exceeds any baseline (BERTOPIC is second-best with 7). Figure 2. HYPOTHESAES produces a larger number (left) and per- centage (right) of significant hypotheses that are rated helpful and interpretable. We report majority vote of 3 ratings per hypothesis. 6.3. Non-SAE Ablations To isolate the value of the SAE, we conduct ablation ex- periments. Because our method differs from baselines in multiple ways—e.g., our neuron interpretation procedure differs from that of NLPARAM—it is worth asking what gain the SAE specifically provides. To answer this question, we run the following ablations on the real-world datasets: • Embedding: We skip training an SAE and directly select 20 features, via Lasso, from text embeddings. The rest of our method is unchanged. 5To reduce annotator workload, our qualitative evaluation only included results from HEADLINES and CONGRESS, since these are more well-studied social science tasks than YELP. • PCA: Same as Embedding, but we first use PCA to re- duce the embedding dimension from 1536 down to 256. • Bottleneck: Inspired by bottleneck models (Yang et al., 2023), we fit a neural network to predict y using text embeddings as input and a single hidden layer with 20 neurons. We generate hypotheses by interpreting the hid- den neurons (in the same way we interpret SAE neurons). We show results in Table 7. In addition to the count of significant hypotheses and overall predictiveness, we also report average neuron fidelity. We establish two benefits of the SAE: the neuron interpreta- tions are (1) higher fidelity and (2) cover a wider variety of non-redundant concepts. Regarding (1), interpretations of SAE neurons achieve a mean F1 of 0.84, compared to 0.64 for the best ablation (bottleneck). Regarding (2), we see im- provements to downstream metrics: the SAE produces many more significant hypotheses (45/60 across the three datasets, vs. 30/60 for the best ablation) and achieves slightly higher predictiveness. Qualitatively, the SAE neurons yield more specific interpretations, while the ablations tend to yield generic ones. Both benefits (1) and (2) are consistent with the SAE producing monosemantic neurons, which are easy to interpret and specific enough to avoid redundancy. 6.4. Evaluating Hypothesis Novelty Beyond quantitative comparisons against baselines, we eval- uate if HYPOTHESAES discovers new hypotheses even on datasets that have been thoroughly analyzed in prior work. CONGRESS. A classic study from Gentzkow et al. (2016) identifies bigrams and trigrams in Congressional speeches that mark Republican (R) or Democrat (D) speakers. We compare to their work by computing counts, in each speech, of their full reported n-gram lists (150 R / 150 D); we also run HYPOTHESAES to produce 100 hypotheses, of which 33 are significant. A regression using only n-gram counts achieves AUC 0.61, while including our 33 hypotheses in- creases AUC to 0.74. Notably, 28 out of 33 hypotheses re- main significantly predictive when controlling for n-grams, showing that our hypotheses add signal to classical methods. Beyond predictive performance, HYPOTHESAES offers two qualitative advantages. First, our hypotheses are inherently interpretable. For example, while Gentzkow et al. (2016) report that “oil companies” correlates with Democrats, it requires context to interpret: are speakers referring to oil companies positively or negatively? In contrast, HYPOTHE- SAES’s corresponding hypothesis is criticizes oil compa- nies or oil industry practices. Second, our hypotheses cap- ture nuanced patterns that n-grams cannot expresss, like criticizes government resource allocation, highlighting dis- parities between domestic needs and actions abroad. 8