A new paper argues that transformer interpretability can be boiled down to two assumptions and one graph, instead of a grab bag of separate techniques.
The researchers behind Unpack start from two architectural facts: attention and MLP layers both work by selecting values through a shared lookup form, and every component reads from the same additive residual stream. Freeze those selections at their forward-pass values, and the whole model becomes a static computational graph. From that graph, the same backward attribution procedure answers three questions interpretability researchers usually tackle separately: which components interact, how information routes through the network, and which input tokens drove a given output. The team tested Unpack on models ranging from 160 million to 6.9 billion parameters, checking its answers against established tests for ablation effects, known circuits, and token attribution.
Interpretability research has splintered into separate toolkits for circuit-tracing, attribution, and ablation, each with its own assumptions and blind spots. A single framework that answers all three questions with the same math makes results easier to compare across labs and models, and easier to trust when they disagree. The paper's more interesting finding: on a task with a known circuit, a component's contribution to the output diverged from its causal effect, and that mismatch had a recognizable signature. Real signal-carrying components changed their output when the task was removed from the input, while bias-like components did not.
It's not a new capability so much as tidier plumbing, and the real test will be whether it holds up on today's frontier-scale models rather than the comparatively modest ones benchmarked here.