/p/2026-10-05 · explainer
Paper explainer · 2610.00622 · Majdoub, Hamid, Ben Charrada et al.

The logic was fine. The import line was not.

Hand-check a hundred files a model wrote against a third-party library and 84% of them carry at least one library fault — a path that moved, an import that was never written, a package that does not exist, a call removed two releases ago. The fix tested here adds no cleverness: name the libraries the task needs, fetch the official documentation, pass the verified import lines into generation, compile, check the imports back against the docs, and loop. Across five models and 300 tasks drawn from fast-moving agent frameworks it removed 38.1% to 54.6% of the library faults, and the documentation fetch was the stage that mattered most.

01 · The problem

Five ways an import goes wrong, and how often each one does

count

02 · What each one looks like

The same five faults, in the frameworks you actually import illustrative


      

03 · The pipeline

Four stages, and the one that is carrying them

remove it and measure

04 · Across five models

The weaker the model, the more the documentation buys

05 · In your own product

What it removes from your review queue illustrative

model behind your codegen

Results

What the paper actually measured

What it does not show

In practice