The Rhyme Was Planned Before the Line Ended
A poetry case study for Anthropic’s NLA: models plan rhymes ahead—and editing the explanation can causally change what they write next.
When a model writes a rhyming couplet, it can look as if it invents the second line only after finishing the first. Mechanistic work on Claude has long suggested otherwise: by the time the first line lands on a word like “grab it,” the model may already be holding “rabbit” as the planned rhyme. Natural Language Autoencoders (NLA) make that early plan readable—and, more importantly, editable.
Reading the plan at the line break
In the reported case, researchers used a couplet prompt that sets up a rabbit rhyme. With NLA on Claude Opus, the explanation at the newline token after the first line already described a plan to finish with “rabbit.” That matters because the second line had not been written yet. The “thought” was sitting in activations, not in the visible draft.
If you can edit the explanation, you can move the output
NLA is not only a viewer. Researchers rewrote the explanation—swapping rabbit for mouse, carrot for cheese—then used the Activation Reconstructor to turn the edited text back into an activation direction. Injecting the difference into the residual stream, at sufficient strength, steered the model away from rabbit toward mouse or house. That is a causal check: if the explanation were empty storytelling, editing it should not reliably change completions.
Planning is not only for poems
Poetry is a clean demo because rhyme is crisp and easy to spot. The same pattern shows up elsewhere: a model may pre-compute an arithmetic answer before a tool returns, or weigh a reward rule while never naming it. Advance planning means the next token is not pure improvisation; parts of the future answer are already represented. For builders, that reframes “hallucination” and “stubbornness” as sometimes being early commitments, not last-second mistakes.
What to take away without overclaiming
NLA does not prove that models “think in English” the way people do. It shows a practical loop: verbalize an activation, test whether that verbalization tracks behavior, and—when needed—steer by editing the words and writing them back. For a general reader, the memorable point is simpler: by the time a line looks finished on the page, the model may already know how it wants the next line to land.