A new framework to turn binary code into readable software instructions
ATLAS uses multiple agents to reconstruct LLVM intermediate representation from optimized binaries.
Researchers built ATLAS, a multi-agent framework. It lifts stripped and heavily optimized binaries to a canonical LLVM intermediate representation (IR).
The framework uses structure-aware, region-level retrieval. This method uses pseudo-probe attribution to align instructions with canonical IR exemplars. A collaborative refinement loop closes the gap between probabilistic generation and deterministic compiler constraints.
A Lift agent creates an initial IR draft. An Advisor agent interprets LLVM verifier and compilation diagnostics to propose repair actions. A Validation agent applies edits and re-runs toolchain verification or runtime checks.
The agents iterate until the IR is verifier-clean or a budget limit is reached. ATLAS achieved 91.8% re-executability and 89.4% CFG similarity across ARM, x86-64, and RISC-V binaries.
This work is essential for security analysis and legacy software migration when source code is unavailable. It enables practical cross-ISA recompilation and verifiable binary-only migration workflows.
Key facts
- ATLAS achieved 91.8% re-executability on HumanEval-Decompile and MBPP.
- The framework works across ARM, x86-64, and RISC-V binaries.
- ATLAS outperforms seven baselines in lifting LLVM IR.