DSS 2026 / Day 2 / Amber Hall
Bytecode Back to Source: AI Reconstruction of Unverified Contracts
Yannis Smaragdakis
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:00you.
0:04Okay. So I'm Janis from D.D.O.B.
0:08Uh I'm co-founder of D.D.O.B. Maybe
0:11let's get a quick show of hands. Who has
0:13ever used our decompiler? Who has ever
0:15used D.D.O.B.?
0:17Okay. So a bunch of people have. So this
0:20is what we've been known for. So since
0:232019 we've had a decompiler that's been
0:25running over the entire Ethereum mainnet
0:28and many other blockchains.
0:30Uh so our decompiler is code code named
0:33Gigahorse. It's now in version three.
0:35It's been documented in multiple
0:37publications. Right now we have over
0:3915,000 registered users.
0:41Some of those are duplicates. We know
0:43that some of you like to abuse our
0:45service and
0:47uh get lots for free. Uh
0:50we don't like that, but okay. But still
0:53it's it's a very sizable community
0:55that's using our decompiler. And if you
0:57are a user or if you're not uh
1:00if you go on app.ddob.com
1:03right now one of the stages of the
1:05compilation that you will see is AI
1:07reconstruction. So I'm going to talk a
1:08bit about AI reconstruction. Our AI
1:10reconstruction on the site is just a
1:13poor man's just one shot thing.
1:16Obviously you can also get the output of
1:18our decompiler and feed it to your
1:19favorite LLM, but we try to give you
1:22very presentable code if you want to
1:24just use our site. Uh and if you want to
1:26integrate uh much much of our decompiler
1:29is open source. The the heart of it is
1:31open source uh and it's a Gigahorse
1:34toolchain. So you can find it on GitHub
1:37uh down there and you can get all of
1:39that if you just search on D.D.O.B.
1:41Okay. So basically
1:43the big question here, and this is a
1:45very short talk, is can modern LLMs
1:48decompile from bytecode? And you
1:50probably have an idea about that
1:52already. And the answer is well to a
1:55first approximation yes.
1:57They are modern LLMs are obviously
2:00incredibly impressive. They've
2:02superseded us in so many ways. So, of
2:04course, they can decompile from
2:06bytecode. But how do they decompile from
2:08bytecode? So I took that terminology
2:10from a paper that I'm going to cite in
2:12one slide, end-to-end models, no
2:14algorithmic component. You just
2:17bytecode in, decompiled out. So that
2:20would be end-to-end. You just give it
2:21the bytecode.
2:23The refinement model in contrast, which
2:25is what we do and what I will argue is
2:29uh advantageous in many ways, you start
2:32from an algorithmic decompiler. You get
2:34some hacky intermediate language out.
2:37Maybe it looks like Solidity, maybe it's
2:39even lower level. Actually, looking like
2:40Solidity is turns out to be an
2:42advantage. And then you have an LLM
2:44produce the final code. So end-to-end
2:47versus refinement model. Uh now I have
2:50to warn overall, every time you see a
2:54study or every time you see an anecdote,
2:56like someone goes on Twitter and says,
2:58hey, we decompiled that contract using
3:00just an LLM and nobody could decompile
3:02it before or anything like that. So
3:04LLMs are great because of experience,
3:07not because necessarily of smarts. So if
3:10your contract is not too rare, if it's
3:13very similar to past contracts, or if
3:15it's something that has source code
3:17online already, the LLM will do
3:18fantastically well. Every time we've
3:21tried it with completely original
3:23bytecode,
3:25it has not done that tremendously well.
3:27There are clear shortcomings and it's
3:30very easy to get fooled. We've had to
3:32work very hard to get representative
3:34studies. So don't trust people who don't
3:37give you full details on the workload.
3:39And one of the things that you need to
3:40have is prevent network access. The LLM
3:43should not be able to look up the
3:45contract online. And if you have some
3:49way, which is very, very hard typically,
3:51to make sure that the LLM has not seen
3:53the contract before, like if you create
3:55new contracts, then that's the most
3:57representative. But that's an extremely
3:59hard study to run.
4:01Uh also on the side on the
4:04uh caveat side, I will cite this paper
4:07that I mentioned before, uh which is uh
4:10by Syracuse University, When LLM
4:12Decompilers Recompile More and Preserve
4:13Less. It was just recently uploaded a
4:15couple of months ago to Archive. Uh this
4:18is for x86, this is for binary
4:20decompilers, but one very the the main
4:22finding there is that
4:24you have an LLM produce decompiled code.
4:28It's great, it satisfies all the test
4:30cases. If you have other test cases that
4:32it has not seen, most likely the
4:34decompiled code will do differently than
4:36the original code. So the LLM will give
4:38you
4:39fixed code, so to speak, but not
4:41necessarily faithful code compared to
4:44the original thing you're decompiling.
4:45So that's another caveat. So let's get
4:47to more original results on our study.
4:50So on our study, uh refinement model
4:54versus end-to-end model, like if you
4:56give it the output of our algorithmic
4:58decompiler and use versus starting
5:01directly from bytecode and all the
5:03setting of the experiment is up there.
5:05So these are very popular contracts. So
5:08the LLM has likely seen many of those
5:11because they are the thousand contracts
5:12that don't use a Yul pipeline and have
5:14the most transactions in the past couple
5:17of years. So popular contracts. So we
5:20expect inflated numbers, but still the
5:22refinement model is clearly better than
5:26the end-to-end model. So we get uh
5:29recompilation rate uh of over 92% versus
5:3388%.
5:34We have lower cost, obviously in tokens.
5:37And then if you actually look at the
5:39output, the output that has uh how many
5:42of the decompiled contracts contain
5:43assembly, it's a much higher percentage
5:46for the end-to-end model. And the
5:49assembly uh share of the code again is
5:51much higher in the end-to-end model. So
5:54the argument here is that if you try to
5:56be very systematic, there's clear value
5:59in having a refinement model, in having
6:01an algorithmic decompiler as a first
6:03stage and then running a good LLM over
6:06it.
6:06Clearly,
6:08any of those numbers are
6:11very are subject to change. If you have
6:14a million rounds on the LLM and you keep
6:17refining and you give it the right
6:19objectives, it will end up being near
6:22perfect.
6:23But it's going to be very very hard for
6:26it to get near perfect. It's much easier
6:28for it to fool you and run a bad
6:30experiment is my point. If you try to
6:32run a fair experiment, it's always
6:35better to give it high-level code to
6:36begin with.
6:38Okay, so that's uh that's the that's the
6:40brief talk and I think we have time for
6:44brief questions.
6:49One one question.
6:53I think that's you. Yes.
7:06>> Yes.
7:09>> No, this is not identical bytecode
7:11matching because if it's identical
7:13bytecode, then it will pass by
7:14definition all the test cases. So this
7:16is not something that's doing
7:19transaction replay or anything. I mean,
7:20we have statistics on transaction
7:22replay, but the very first test is
7:24re-compilability. And even if you fail
7:27if you even fail re-compilability, then
7:29there's no point in
7:31in trust in trying transaction replay.
7:33But you're asking is it identical
7:34bytecode? No. I mean, it wouldn't be
7:36that high. Like Garrett's numbers from
7:38before, it would be like for
7:40re-compilable, you can I am sorry, for
7:42identical bytecode, you can let the LLM
7:45run for a month. Maybe you'll get to 70%
7:47at this point. Maybe in the future it
7:49will be much better. Maybe you spend
7:50billions more in tokens. You can get to
7:53be much better, but it's going to have a
7:55much lower
7:57upper threshold.
8:01Thank you.
Source and accuracy details
- Names, organizations, technical terms, code and numbers may be misrecognized.
- Source wording and disfluencies are preserved; no editorial corrections were applied.
- Export contains whole-second cue starts only; cue ends and subsecond timing are unavailable.
- A full caption-track export does not establish that every spoken word is represented.
- Automatic YouTube captions; not human-verified against the audio.
Source export SHA256: b8a148a7748c74c02447e96da7d35c05d1ad29cd6fb2a58bd812ac1bc89f4ba8