TRAIL pairs a translator agent against a challenger agent and gains 23.1% relative syntax accuracy on C-to-Rust translation

arXiv·medium signalXBlueskyLinkedInCopy link

A September 14 arXiv paper attacks the problem that per-trace debugging experience in agentic code translation does not accumulate into reusable knowledge. TRAIL runs two collaborating agents adversarially, a Translator and a Challenger, distilling trace-specific experience into generalizable translation rules. Against the strongest baseline LLM translator it reports 23.1% relative improvement in syntax accuracy and 15.9% in semantic accuracy on CRUST-Bench and SmartC2Rust-Bench, with the refined insights transferring across both benchmarks.