CHISEL-ing Back Source Code with AI-enabled Iterative Recovery

1 Introduction

• Traditional decompilers such as Ghidra and IDA Pro produce pseudo-C that is difficult to read, rarely compilable, and not guaranteed to preserve semantics; LLM-based approaches include non-iterative and feedback-guided iterative methods.Pseudo-C can contain synthetic types, machine-level artifacts, and ABI scaffolding.Source: §1 · p. 1
• Prior methods’ reliance on native test suites limits semantic-divergence feedback when reversing binaries of unknown provenance; asymmetric sanitization and symbolic-execution path explosion pose further challenges.In-loop feedback has also typically been limited to compiler errors and return-value divergence.Source: §1 · p. 1
• CHISEL is a test-suite-free iterative decompiler whose feedback comes from a compiler and coverage-guided fuzzer operating on the original binary.It adds rich observables for grounded differential analysis, cross-iteration divergence memory, and best-candidate retention.Source: §1 · pp. 1–2
• Its six-arm ablation builds from single-shot generation to the end-to-end design and evaluates recovery, feedback-oracle soundness, and iteration overhead.The axes include re-compilability, re-executability, regression and recovery metrics; false-acceptance and false-rejection; and average iterations.Source: §1 · p. 2
• Across 120 ExeBench functions, CHISEL reports 96.1% RC, 79.8% RE, 9.4% FA, and 26.2% recovery after first-generation failure, outperforming LLM4Decompile and Agent4Decompile.The evaluation covers x86-64 binaries at O0–O3, both stripped and unstripped, using Gemma4:31b.Source: §1 · p. 2

2 Related Work

• LLM4Decompile [18], Nova, and Idioms generate code without behavioral feedback, while DeGPT uses multi-role edits guarded by static symbolic checks.Source: §2 · p. 2
• DecLLM combines compiler repair, test-seeded AFL++ divergence feedback, and candidate sanitization, while PCodeTrans [7] uses official tests and breakpoint-matched differential tracing.The paper characterizes sanitizing only generated code, unlike the original, as unfaithful decompilation.Source: §2 · p. 2
• Agent4Decompile [20] checks syntax, compilation, and execution through three agent tiers but assumes supplied test cases; prior work generally has limited observable coverage except PCodeTrans [7].Source: §2 · p. 2
• CHISEL uses compiler and type-aware coverage-guided fuzzing feedback without supplied tests, plus rich observables, cross-iteration divergence memory, and best-candidate retention.Its differential analysis observes return values, output buffers, stdout, exit codes, timeouts, and crashes.Source: §2 · p. 2

3.1 Iterative Enhancement Algorithm

• Each round prompts the LLM for one candidate, uses compiler diagnostics to guide compilability, and applies differential fuzzing to find inputs that expose behavior divergence.If neither oracle flags an issue, the candidate is returned as recovered source.Source: §3.1 · p. 3
• Differential fuzzing mines inputs with coverage-guided entropic libFuzzer from type-aware signature-derived seeds, then samples accumulated divergences as feedback within capped feedback budgets.At most d_c compilation errors and d_f divergences are passed directly; remaining counts are summarized to stay within token budgets.Source: §3.1 · p. 3
• Chisel’s observe feature compares rich outputs and signals, memory replays deduplicated divergences across iterations, and retain_best returns the best-performing candidate.Observables include return values, output buffers, stdout, timeouts, exit codes, and crashes; pointer results are compared by content rather than raw address.Source: §3.1 · p. 3

3.2 Prompts and Parameters

• The self-contained prompt first includes task instructions and pseudo-C, then appends the previous candidate and available feedback in later rounds; runs use K=5, temperature 0, reasoning off, and O0 analysis.Each round tests against 200 type-aware seeds and 2,000 libFuzzer-mined inputs with a 2-CPU-second per-input timeout; feedback caps are d_c=5 and d_f=10.Source: §3.2 · pp. 3–4

4.1 Hardware, Software, and Dataset

• Experiments use 120 ExeBench functions compiled for x86-64 at O0–O3, in stripped and unstripped variants.The functions were selected from 150 longest real-world functions based on availability of at least 200 functional test cases; each selected function has 46–265 lines of code. Runs use an x86-64 Linux host with four NVIDIA H200 GPUs, Gemma4:31b3 at int-4 quantization, Ghidra, gcc -c, and libFuzzer.Source: §4.1 · p. 4

4.2 Experiments and Evaluation Metrics

• A six-arm incremental ablation evaluates recovery, feedback-oracle soundness, and iteration overhead, with the final arm representing the complete CHISEL framework.Metrics are scored on the loop’s final returned candidate against a held-out test suite; compilation requires building, and re-execution requires matching the original on every suite input. The measures include compile and re-execution rates, recovery from initial failures, oracle false acceptance and rejection, regressions, and mean iterations.Source: §4.2 · p. 4

4.3 Baselines

• The baselines are LLM4Decompile-9B-v2 [18] and Agent4Decompile [20], evaluated without test-suite access during iterations.Agent4Decompile uses Gemma4:31 to align the harness comparison; both baselines use Ghidra and their specified prompts and input formats. DecLLM and PCodeTrans [7] are excluded because their code is not public.Source: §4.3 · p. 5

Ablation

• Differential fuzzing raised RE to 79.3% and recovered 7.5% of compiled-but-incorrect first generations; observe halved FA from 17.6% to 9.4%.Memory raised Pass over +observe, while best-candidate retention increased mean suite pass-rate from 50.9% to 52% compared with the previous arm.Source: §5 · p. 5, Figure 2 · p. 5

Soundness and residual errors

• FA is 16.4% on stripped binaries versus 1.7% on unstripped, and residual 9.4% FA is dominated by signature-recovery errors, predominantly in stripped samples.Because candidates are tested against their recovered signatures, argument mis-recovery can pass the differential oracle yet fail the original-signature suite; the differential oracle also covers divergences outside that suite, while FR approaches 45% with a behavioral oracle and falls with optimization.Source: §5 · pp. 5–6, Table 4 · p. 6, Table 5 · p. 6

Comparison with baselines

• LLM4Decompile-9B-v2 achieves 17% RE versus 73.0% for CHISEL’s LLM-only Gemma arm; CHISEL also outperforms A4D at lower iteration cost.The comparison uses the same 120 functions and test suite.Source: Table 6 · p. 6, §5 · p. 6

Results across optimizations and stripping conditions

• CHISEL re-executes 86.7% of unstripped and 72.9% of stripped functions on average; RE falls from 93.3% to 80.8% and 78.3% to 68.3% from O0 to O3, respectively.The passage reports that R_exec and R_tot confirm that the iterative loop repairs a growing share of the re-execution gap.Source: §5 · p. 6, Table 4 · p. 6, Table 5 · p. 6

6 Limitations, Risks and Ethics

• Signature recovery from stripped binaries remains difficult; experiments used deterministic 0-temperature, non-reasoning settings, though higher temperatures and reasoning may improve generation and feedback response.The authors suggest that more creative generation, critical thinking, and better responses to feedback may aid recovery.Source: §6 · p. 6
• Using Ghidra pseudo-C as ground truth leaves decompiler resilience unevaluated, while I/O sampling only approximates functional similarity and cannot establish semantic equivalence.Source: §6 · p. 6
• Because ExeBench may overlap LLM training data, future work could use newer datasets such as [14].Source: §6 · p. 6
• The authors warn that the work could be adopted by threat actors for unauthorized activities and urge ethical use.Source: §6 · p. 6

7 Conclusion

• Chisel achieved 96.1% RC and 79.8% RE, repaired 26.2% of first-generation failures, and cut FA from 17.6% to 9.4% versus return-only differential fuzzing.The evaluation used 120 ExeBench functions compiled for x86-64 at four optimization levels, with stripped and unstripped variants.Source: §7 · p. 6
• On the same subset, Chisel’s RE was 79.8% versus 17.0% for LLM4Decompile-9B-v2, and it outperformed Agent4Decompile on RE, FA, and overhead under the same LLM and test-suite assumptions.Source: §7 · p. 6
• Chisel is a test suite-free iterative decompiler combining compiler-based static analysis and coverage-guided differential fuzzing, augmented with rich observables, cross-iteration divergence memory, and best-candidate recovery.Source: §7 · p. 6
• The results identify richer observation, rather than more search, as the decisive gain and highlight the feedback harness as an important lever alongside the LLM.Source: §7 · p. 6

• Generative AI tools Claude and Gemini helped improve manuscript language, generate verified plotting code, and write prototype code.Source: A Generative AI Usage · p. 7
• The authors produced experiments, datasets, and numerical results through their own pipeline, regenerated figures and tables from raw outputs, and reviewed AI-assisted code while taking responsibility for the work’s accuracy, originality, and integrity.Source: A Generative AI Usage · p. 7