What was done
Kohli and colleagues built CHISEL, a loop that uses compiler feedback and a coverage-guided fuzzer to refine an LLM's recovery of source code from Ghidra pseudo-C.
Varun Kohli, N Raghava, Biplab Sikdar and Dinil Mon Divakaran published CHISEL, a framework that iteratively recovers source code from Ghidra-generated pseudo-C without needing test suites. The paper was accepted to ACM CCS SURE 2026.
How it was done
CHISEL asks an LLM, Gemma4:31b in the paper, to rewrite the pseudo-C. Compiler feedback acts as static analysis, and a coverage-guided fuzzer acts as differential analysis against the original. The framework keeps a memory of past divergences and retains the best candidate. It was tested on 120 ExeBench functions compiled for x86-64 at four optimisation levels, stripped and unstripped.
Results
The authors report about 96.1% re-compilability and 79.8% re-executability on average, in 2.1 iterations on average. The loop recovers about 26% of first-generation execution errors, and the feedback oracle wrongly accepts about 9.4% of candidates. They say it beats two recent LLM-assisted decompilation approaches. No code link is listed.