Nikhil Garg

Rajiv Movva, Kenny Peng, Nikhil Garg et al.

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 language interpretation of each feature (e.g., "mentions 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 findings), 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 differences in Congressional speeches and identifying drivers of engagement with online headlines.

Kenny Peng, Rajiv Movva, Jon Kleinberg et al.

While sparse autoencoders (SAEs) have generated significant excitement, a series of negative results have added to skepticism about their usefulness. Here, we establish a conceptual distinction that reconciles competing narratives surrounding SAEs. We argue that while SAEs may be less effective for acting on known concepts, SAEs are powerful tools for discovering unknown concepts. This distinction cleanly separates existing negative and positive results, and suggests several classes of SAE applications. Specifically, we outline use cases for SAEs in (i) ML interpretability, explainability, fairness, auditing, and safety, and (ii) social and health sciences.

Elliot Kim, Avi Garg, Kenny Peng et al.

Diversity in training data, architecture, and providers is assumed to mitigate homogeneity in LLMs. However, we lack empirical evidence on whether different LLMs differ meaningfully. We conduct a large-scale empirical evaluation on over 350 LLMs overall, using two popular leaderboards and a resume-screening task. We find substantial correlation in model errors -- on one leaderboard dataset, models agree 60% of the time when both models err. We identify factors driving model correlation, including shared architectures and providers. Crucially, however, larger and more accurate models have highly correlated errors, even with distinct architectures and providers. Finally, we show the effects of correlation in two downstream tasks: LLM-as-judge evaluation and hiring -- the latter reflecting theoretical predictions regarding algorithmic monoculture.

Nikhil Garg, Paul Uriarte Vicandi, Yanming Zhang et al.

The increasing deployment of wearable sensors and implantable devices is shifting AI processing demands to the extreme edge, necessitating ultra-low power for continuous operation. Inspired by the brain, emerging memristive devices promise to accelerate neural network training by eliminating costly data transfers between compute and memory. Though, balancing performance and energy efficiency remains a challenge. We investigate ferroelectric synaptic devices based on HfO2/ZrO2 superlattices and feed their experimentally measured weight updates into hardware-aware neural network simulations. Across pulse widths from 20 ns to 0.2 ms, shorter pulses lower per-update energy but require more training epochs while still reducing total energy without sacrificing accuracy. Classification accuracy using plain stochastic gradient descent (SGD) is diminished compared to mixed-precision SGD. We analyze the causes and propose a ``symmetry point shifting'' technique, addressing asymmetric updates and restoring accuracy. These results highlight a trade-off among accuracy, convergence speed, and energy use, showing that short-pulse programming with tailored training significantly enhances on-chip learning efficiency.

Wenshuo Guo, Karl Krauth, Michael I. Jordan et al.

Recommender systems play a crucial role in mediating our access to online information. We show that such algorithms induce a particular kind of stereotyping: if preferences for a set of items are anti-correlated in the general user population, then those items may not be recommended together to a user, regardless of that user's preferences and rating history. First, we introduce a notion of joint accessibility, which measures the extent to which a set of items can jointly be accessed by users. We then study joint accessibility under the standard factorization-based collaborative filtering framework, and provide theoretical necessary and sufficient conditions when joint accessibility is violated. Moreover, we show that these conditions can easily be violated when the users are represented by a single feature vector. To improve joint accessibility, we further propose an alternative modelling fix, which is designed to capture the diverse multiple interests of each user using a multi-vector representation. We conduct extensive experiments on real and simulated datasets, demonstrating the stereotyping problem with standard single-vector matrix factorization models.