Abstract
Despite rapid adoption and deployment of large language models (LLMs), the internal computations of these models remain opaque and poorly understood. In this work, we seek to understand how high-level human-interpretable features are represented within the internal neuron activations of LLMs. We train $k$-sparse linear classifiers (probes) on these internal activations to predict the presence of features in the input; by varying the value of $k$ we study the sparsity of learned representations and how this varies with model scale. With $k=1$, we localize individual neurons which are highly relevant for a particular feature, and perform a number of case studies to illustrate general properties of LLMs. In particular, we show that early layers make use of sparse combinations of neurons to represent many features in superposition, that middle layers have seemingly dedicated neurons to represent higher-level contextual features, and that increasing scale causes representational sparsity to increase on average, but there are multiple types of scaling dynamics. In all, we probe for over 100 unique features comprising 10 different categories in 7 different models spanning 70 million to 6.9 billion parameters.
See twitter summary here.
Contributions
In the first part of the paper, we outline several variants of sparse probing, discuss the various subtleties of applying sparse probing, and run a large number of probing experiments. In particular, we probe for over 100 unique features comprising 10 different categories in 7 different models spanning 2 orders of magnitude in parameter count (up to 6.9 billion). The majority of the paper then focuses on zooming-in on specific examples of general phenomena in a series of more detailed case studies to demonstrate:
There is a tremendous amount of interpretable structure within the neurons of LLMs, and sparse probing is an effective methodology to locate such neurons (even in superposition), but requires careful use and follow up analysis to draw rigorous conclusions.
Many early layer neurons are in superposition, where features are represented as sparse linear combinations of polysemantic neurons, each of which activates for a large collection of unrelated $n$-grams and local patterns. Moreover, based on weight statistics and insights from toy models, we conclude that the first 25\% of fully connected layers employ substantially more superposition than the rest.
Higher-level contextual and linguistic features (e.g., \texttt{is\_python\_code}) are seemingly encoded by monosemantic neurons, predominantly in middle layers, though conclusive statements about monosemanticity remain methodologically out of reach.
As models increase in size, representation sparsity increases on average, but different features obey different dynamics: some features with dedicated neurons emerge with scale, others split into finer grained features with scale, and many remain unchanged or appear somewhat randomly.
We will have a follow up post in the coming weeks with what we see as the key alignment takeaways and open questions following this work.
I really like this paper (though, obviously, am extremely biased). I don’t think it was groundbreaking, but I think it was an important contribution to mech interp, and one of my favourite papers that I’ve supervised.
Superposition seems like an important phenomena that affects our ability to understand language models. I think this paper was some of the first evidence that it actually happens in language models, and on what it actually looks like. Thinking about eg why neurons detecting compound words (eg blood pressure) were unusually easy to represent in superposition, while “this text is in French” merited dedicated neurons, helped significantly clarify my understanding of superposition beyond what was covered in Toy Models of Superposition (discussed in Appendix A). I also just like having case studies and examples of phenomena in language models to think about, and have found some of the neuron families in this paper helpful to keep in mind when reasoning about other weirdnesses in LLMs. I largely think the results in this paper have stood the test of time.