Indirect Object Identification (IOI) task
Sparse Autoencoders Find Highly Interpretable Features in Language Models — inherited
A previously-studied model behavior (Wang et al., 2022) in which a language model completes sentences such as 'Alice and Bob went to the store. Alice gave a snack to ___' by correctly naming the indirect object. Used as this paper's case-study task for testing how precisely dictionary features localize a known circuit, via activation patching and KL-divergence comparisons against PCA.
Most claims about what a learned feature does can only be checked against themselves — an automatically generated explanation is judged self-consistent, never independently verified. Indirect Object Identification breaks that pattern: a template sentence-completion task whose underlying circuit inside a language model had already been mapped by earlier researchers, working head by head, well before this paper's methods were ever pointed at it.
This page covers the task itself and the circuit behind it, and why testing a new patching procedure against an already-published answer gives its results a kind of external check that most of the paper's other interpretability claims simply don't have.
The known circuit that lets the paper check its own ruler
The Indirect Object Identification task is the proving ground for the causal-proof-by-patching theme: every other technique in Section 4 -- Activation patching, Dictionary-feature patching procedure, the ACDC-borrowed feature ranking -- is only as convincing as the task it is tested on, and IOI is chosen precisely because its answer is already known from independent prior work rather than from this paper's own methods.
Wang et al. (2022)'s circuit, in general terms
IOI is a template task introduced by Wang, Variengien, Conmy, Shlegeris and Steinhardt (2022) in "Interpretability in the Wild": a model is given sentences of the form "Then, [Name A] and [Name B] went to the [place]. [Name A] gave a [object] to ___" and must complete them with the correct indirect object, the person who did not perform the giving action. Wang et al. hand-reverse-engineered a multi-part circuit in GPT-2 small responsible for this behavior, involving specific classes of attention heads (Duplicate Token Heads, S-Inhibition Heads, Name Mover Heads among them) that together track which name has already appeared and which should be named again; the paper became a canonical worked example in mechanistic interpretability and one of the benchmark tasks ACDC itself was later validated against.
This paper's usage: a task the dictionaries were never trained for
This paper studies model behaviour on IOI using sentences like "Then, Alice and Bob went to the store. Alice gave a snack to ___", chosen because it "captures a simple, previously-studied model behaviour" (sparse-autoencoders, §"4 Identifying causally-important dictionary features for indirect object identification", p. 6). The patching procedure then drives a counterfactual variant of that sentence, with the indirect object swapped (e.g. "Bob" replaced by "Vanessa") (sparse-autoencoders, §"4.1 Adapting activation patching to dictionary features", p. 6). Crucially, the dictionary features under test were trained with no knowledge of IOI at all: "the training of our feature dictionaries does not emphasize any particular task" (sparse-autoencoders, §"4 Identifying causally-important dictionary features for indirect object identification", p. 6).
Why an externally-verified task matters here
Most of the paper's interpretability claims rest on autointerpretability scores, which have no independent ground truth to check against -- a feature's automatically-generated explanation can only be judged self-consistent, not verified correct. IOI breaks that limitation: because its circuit was already published by Wang et al. (2022) before this paper existed, the patching results on it -- fewer features and smaller edits reaching a given KL divergence than PCA (sparse-autoencoders, §"4.2 Precise localisation of IOI dictionary features", p. 6) -- are a claim that could in principle have come out looking arbitrary or wrong, and didn't.