From Transformer Circuits Perspective to Hands-On Tracing of Open-Source Models
PyCon TW 2026 · poster · Anna Liang
For an engineer, a black box has a concrete definition: you cannot set a breakpoint in it.
This poster shows how to set one in a 2B open-source model. Take gemma-2-2b,
give it a small Python function, cut the prompt right after return, and ask
which local variable it will name. Then look inside: an attention heatmap says
where the model looks from the return position, an attribution graph
built with circuit-tracer
says what is moved toward each candidate logit, and steering one pathway down
says whether it matters. Every variable name turns out to have its own pathway to its
own logit; binding sets how much drive a name accumulates; scaling one pathway down flips the
answer. The same probes also show where the model's picture of Python breaks: it holds the
wrong rules for late binding and for which names leak out of a loop.
The theory pages give the background the poster only has room to name: the 2020 to 2025 timeline from Zoom In to attribution graphs, attention as a kernel estimator, and where the word "circuit" comes from. The practice pages hold the experiment itself: prompts, numbers, interventions, the interactive graph, and the reproduction steps.
One axis, two phases. First how to think about a transformer (QK and OV circuits, induction heads, superposition), then how to look inside one (sparse autoencoders, transcoders, attribution graphs). Nine papers, one page each where the poster needs it.
Four prompts, 18 names, one model. An attention heatmap from the return
position, attribution graphs with every direct effect named against its logit, a
dose-response curve with attention frozen and free, ablations, and four traps in reading a graph.
Plus the probes that show where the rules break.