AxDafny: Agentic Verified Code Generation in Dafny

Benjamin Breen, Austin Letson, Borja Requena Pozo, Leopoldo Sarra · 2026

Read AxDafny: Agentic Verified Code Generation in Dafny by Benjamin Breen, Austin Letson, Borja Requena Pozo, Leopoldo Sarra (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-32007.txt.

Description

Breen and three others study agents that must write Dafny code together with the proofs that verify it. Dafny programs carry formal specifications, and a verifier checks them by machine. Writing the invariants, assertions and termination arguments is the hard part that usually demands expert effort.

Their framework, AxDafny, loops writing against verifier feedback, fixing code and proof artifacts together. A new benchmark, LCB-Pro-Dafny, turns 250 competition-style problems into Dafny with specifications and a verifier harness. AxDafny verifies 56.4 percent of those tasks, far above the 11.6 percent managed by direct GPT-5.5 writing in their comparison. On DafnyBench it verifies 725 of 782 instances, beating the strongest reported proof-hint baseline by 6.5 points.

Catalog record

CREATORBenjamin Breen, Austin Letson, Borja Requena Pozo, Leopoldo Sarra
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.32007

Browse the full library · About the Tower of Babel