Source
Lindsey et al. (2025), On the Biology of a Large Language Model, LIT-1541, read in NOTE-1181: the section “Planning in Poems” (its four subsections and the table of 25-poem samples).
What was actually shown
The couplet. “A rhyming couplet: / He saw a carrot and had to grab it,” completed “His hunger was like a starving rabbit”. Attributing back from “rabbit”, the important features sit on the newline before the second line: “rhymes with eet/it/et” features, fed from “it”, activate features for “rabbit” and “habit”, which feed “say rabbit” at the last position. The authors “initially hypothesized that we’d observe improvisation”.
Position. Steering negatively on rabbit and habit and positively on a “green” feature (from another poem) changes the predicted final word only when applied at the newline; “the features are only active over the planning token”.
Writing towards the plan. The same rabbit features promote “comparison” features active before “like a” at the intermediate word “like”; suppressing them changes the transition to one suited to “habit”. Resampling the whole line under the interventions (500 samples per experiment) changes its structure to end on the steered word.
Injection across poems. Injecting the “rabbit” and “green” features into “a random sample of 25 poems”: “the model ended its line with the injected planned word in 70% of cases”. The denominator (25 or 50 injections) is not stated.
Frequency. “We found planned word features in about half of the poems we investigated”, which the authors attribute either to the CLT lacking features for the planned words or to the model not always planning.
What could have come out otherwise. The rhyme could have been chosen at the last position, from the line written so far, as the authors expected; steering at other positions could have worked as well as at the newline; injection could have failed to move the final word.
What this does not say
- Not that the model plans in general. One couplet is analysed in depth; the 25-poem test shows that injected word features usually steer the ending, not that the model produced such features itself in those poems; “about half” is a count over an undescribed set.
- Not that the plan is a planning-specific representation. The authors stress the opposite: the features “seem to be ordinary features representing that word”.
- Not about prose. Rhyme is a constraint that makes a target word determinable before the line starts.
- One displayed success is not one. In the article’s table, the “rabbit” injection on “The clouds are gray, the raindrops fall,” gives “Soft and slow like a timid hare.”, which ends in “hare”.