Announcing EVMDec-3: Our Latest Smart Contract Decompiler Model
We are excited to introduce EVMDec-3, the newest version of our smart contract decompiler. Starting today, PLUS and PRO subscribers can access this model, which sets a new baseline in AI-driven decompilation quality.
This is also the first release where we are publishing detailed benchmark results. To ensure transparency, all evaluations are performed on a fixed dataset — EVMDec-Bench v1 — consisting of 46,040 Ethereum smart contracts that were never seen during training.
Evaluation Metrics
We benchmark EVMDec-3 against four core dimensions of decompilation quality:
Completeness
Whether the model produces full outputs that follow the expected contract format without truncation.
Note: incomplete outputs may still provide value in practice, but completeness is a strong indicator of robustness.
Recompilation
Whether the decompiled output can be recompiled back into valid EVM bytecode. This is a stringent measure of structural fidelity.
ABI Recovery (Functions)
Among recompilable contracts, we measure how many function ABIs (selectors and signatures) match the originals.
We focus on functions — not events or errors — since these are the primary interfaces for interacting with contracts.
Textual Similarity
We measure how close the vocabulary distribution of the decompiled output is to the original contract, using TF-IDF similarity. This reflects readability and alignment with developer conventions.
Results
Completeness
- 100% completeness: all 46,040 outputs were fully generated.
- No truncated or malformed results were observed.
Recompilation
🎯 Industry First: EVMDecompiler is the very first smart contract decompiler to support recompilation back to valid EVM bytecode.
- 49.8% success rate: nearly half of the outputs compiled successfully into valid EVM bytecode.
- This is a tough benchmark — small inconsistencies in typing or structure can cause compilation failure — so approaching 50% is a strong starting point.
- No other decompiler in the market currently offers this capability, making this a groundbreaking achievement in the field.
ABI Recovery
When recompilation succeeds, ABI recovery shows a wide spread.
- Average ABI match rate: 31.1%
- Median ABI match rate: 13.0%
- 19.1% of contracts achieve >95% ABI recovery, with many reaching perfect matches.
ABI Recovery Distribution (CDF)
Textual Similarity
Textual similarity scores indicate that most outputs are stylistically close to the original source:
- Average similarity: 0.834
- Median similarity: 0.939
- 46.2% of contracts scored above 0.95.
Textual Similarity Distribution (CDF)
Availability
- PLUS users: EVMDec-3 Model (256K tokens)
- PRO users: EVMDec-3 Model (256K tokens)
Key Takeaways
EVMDec-3 achieves 100% completeness across the benchmark.
Nearly 50% recompilation success, demonstrating structural correctness at scale.
Strong ABI recovery on a significant subset of contracts, including many perfect matches.
High textual similarity, making the outputs both accurate and human-readable.
With EVMDec-3 and EVMDec-Bench v1, we now have a transparent baseline. Future releases will be compared against the same benchmark, ensuring measurable progress over time.