Closing-parenthesis dictionary feature
Sparse Autoencoders Find Highly Interpretable Features in Language Models — introduced
A case-study dictionary feature in layer 5 whose unembedding directly predicts closing-parenthesis tokens, used to illustrate the paper's feature-circuit-detection method by tracing which upstream features (detecting dates, acronyms, and other parenthesis-preceding phrases) cause it to activate.
Testing a method that traces which features cause another feature to fire is only convincing if you can independently check whether the result makes sense — hard to do when the target feature's own meaning is itself uncertain. This case study sidesteps that by picking a target whose identity is about as unambiguous as it gets: a final-layer feature whose top tokens, read directly off the model's own output weights, are unmistakably variants of a closing parenthesis.
The page covers why this feature's position gives it that strong, weights-based confirmation before any tracing begins, what the ablation-based procedure finds sitting upstream of it, and why it was chosen to carry the paper's hardest analytical claim rather than the feature used elsewhere.
The case study that carries circuit analysis, not input or output
The closing-parenthesis feature is one half of the paper's two-feature case-study pair in single-features-under-the-microscope, and it is assigned a specific, narrow job: where the apostrophe feature demonstrates monosemanticity through what activates a feature and what a feature causally does to the logits, the closing-parenthesis feature exists to demonstrate the third leg of Section 5's methodology, tracing which upstream features cause a feature to fire at all.
Feature 52027, layer 5: meaning given almost for free
This is feature 52027 in layer 5, chosen because layer 5 is the model's final layer, meaning its role is to write directions that the unembedding matrix reads out directly as output tokens; unembedding the feature confirms its top tokens are all closing-parenthesis variants (sparse-autoencoders, §"5.3 Intermediate features: dictionary features allow automatic circuit detection", p. 8). That gives the feature's semantic identity an unusually strong, weights-based confirmation before any circuit tracing happens at all -- a luxury a mid-network feature, whose meaning can only be inferred from the text that activates it, does not have.
What tracing its causes actually found
Using the paper's ablation-based circuit-detection procedure -- find the feature's maximum activation $M$, sample 20 activating contexts in $[M/2, M]$, ablate each layer-4 feature in turn and rank by the resulting drop in the target's own activation -- the upstream causes uncovered for the closing-parenthesis feature are features for dates, acronyms, and other phrases that typically precede a closing parenthesis (sparse-autoencoders, §"5.3 Intermediate features: dictionary features allow automatic circuit detection", p. 8). Figure 5 renders the result as a causal tree, with edge thickness encoding the strength of each ablation-measured causal effect and human interpretations attached to every node.
Why this feature, and not a harder one
Because the closing-parenthesis feature's own identity is nearly unambiguous, the sensibleness of the circuit found upstream of it -- date-like phrases and acronyms plausibly do precede closing parentheses -- can be judged directly, rather than trusted on faith. That makes it a credible validation case for Automatic feature circuit detection rather than a mere illustration, and it explains the paper's division of labor between the two case-study features: the apostrophe feature covers input and output analysis, and this feature alone covers the intermediate-feature, circuit-tracing leg.