back
80 comments
"Models favor monolithic, single-file implementations that diverge sharply from human-written code."

You say! I might have been just an LLM all along without even knowing it since I too prefer single file implementations.

Back in the old VB5/VB6 days Visual Studio had this mode where it showed the different functions in a file almost as if they were separate files. You could not scroll beyond the functions end but you could easily transition between that mode and global file view. I always found that a nice way of working (but admittedly the world was a lot simpler back then).

Also my preference for fewer but longer files is only there when I write the code myself. For working with AI I think smaller files are beneficial for quicker turn around between human and machine.

This is interesting. I have always preferred to make my personal projects single-file (or at least few-massive-file)[1]. I noticed that teams in general, strongly dislike this style of programming (even before LLM-coding-assistants, as far back as 2020).

I wonder how much of the multi-file (and increasingly multi-repo) code-organization is just a manifestation of Conway's Law (https://en.wikipedia.org/wiki/Conway%27s_law).

[1]: It makes navigation and iteration much faster, and obviates the need to use indexers. It also forces you to only put _orthogonal_ programs in external files (I recently had to write a kind of quasi-SAT-solver, and that was code that was complex enough to require its own "namespace", and it was also something that was reusable across projects). One thing I noticed, maybe in 2025, is that LLMs struggled to navigate large single-file programs, but were quite good at navigating multi-file programs. It is interesting that they (according to my 2025 experience and the quote you give) prefer to _write_ code in ways that make it difficult for them to _read_ code.

How often has there been a HN submission for a project 'in a single C header file'?
This has less to do with natural opinions regarding code organization and more to do with the fact that including, modularizing, and distributing C code has historically been a pain in the ass which is ameliorated by shoving everything into a single file.
EDIT: Sorry, I missed the "header" part (and the irony).

At least once, here you go:

https://news.ycombinator.com/item?id=48053570

Ok, I just submitted it myself but I could not believe it never had been submitted before. It is from 1997 and was pretty popular for some time. I think it even was built into Google Picasa for sime time.

This VB feature existed to accommodate programmers coming from the DOS based QB IDE who were used to the one function per screen view there. To my sensibilities, it does not make much sense with the advent of high-resolution desktop environments.
I think it's one (but not the only) reason that makes LLMs work very well with Ruby on Rails
This has been my preference as well. I build everything in one file until it becomes uncomfortable and only then I start breaking up into multiple files... But even then, I try to keep the main business logic fully visible in the main file.
It's a very misleading: they don't provide any meaningful documentation/requirements. Just an executable blackbox.

E.g. a doc for ffmpeg, which I checked by downloading docker image they provide to the model, is a README which basically just says this is ffmpeg and docs can be found online. They do not allow models to get online.

So a model is supposed to reverse-engineer a blackbox using only limited number of tries. I'm not sure even ASI can do this under these constraints (without memorizing the ffmpeg code base, obviously.)

In the only posts one of authors mentions "usage docs". Obviously they had a command-line tool like `grep` in mind -- where a man page sort-of specifies program behavior. But then added sqlite, ffmpeg, php, etc. - where a usage doc is like one millionth of information you need to implement ffmpeg.

And, of course, there's no human baseline. I'd guess making such a baseline would cost billions of dollars.

Ahem. Bitkeeper and Samba were reverse engineered just from the protocol by humans. For free.
i thought the agent can execute real ffmpeg to compare
Nice work once again from Ofir Press and team; this seems to be an idea that's in the air.

> Our 200 tasks range from compact CLI tools to widely used software such as FFmpeg, SQLite, and the PHP interpreter. We evaluate 9 LMs and find that none fully resolve any task

Fwiw, this is very different from what we find in MirrorCode:

> Opus 4.6 successfully reimplements almost every program up to gotree’s size in our benchmark.

https://epoch.ai/blog/mirrorcode-preliminary-results

I don't have time right now to dig in to what could explain the difference (I'm working hard on getting the full MirrorCode out as soon as possible). But I suspect that the ProgramBench authors are either under-eliciting the AIs, or their tasks are unfair/impossible given the constraints, or both.

I hope to look more into it after releasing MirrorCode, and write up my conclusions.

Surely the biggest difference is that you guys are mostly testing LLMs on simpler utilities, mostly involving higher-level languages, whereas ProgramBench are all very complex C programs (and much older programs with much more comprehensive test cases).

Eg cal is totally routine. I would expect most sophomores to be able to write a perfectly good cal. In fact the only program you tested which actually has anywhere close to the complexity of SQLite or FFmpeg is is Pkl, and it looks like Opus 4.6 totally failed.

I think your results are consistent. You're just measuring different things. Your benchmarks mostly tests LLMs ability to write technically routine programs of moderate length - yes the bioinformatics package involves specialized domain knowledge, but not specialized Go engineering. ProgramBench is harder.

I would love to try this out. I have a horrible legacy project that is written in angular by a really amateur developer, full of huge blocks of copy pasted code that has minor modifications in each block. I’ve tried before to get an LLM to rewrite it to something more sensible, but I have not succeeded, usually it just ends up breaking everything. Is there a guide or some system to follow? What’s the best way to accomplish a task like this?
Problem with these types of benchmarks is that it’s 100% certain the LLM has been trained on all that code already, so they’re all tainted since you don’t know whether it’s just benchmarking recall vs actual reasoning.

Same with SWE-bench and others.

Is anyone familiar with gotree? That was mentioned as the most complex piece of code, but the metric was LOC. Based on the high level description gotree might be closer to a set of small programs / algorithms.

Interesting anyway. It will be nice to see these comparisons with open weight models and how do those fare.

I should say one big difference is ProgramBench has 200 target programs while MirrorCode has about 30. We did many manual things to ensure task quality, that would have required huge resources to do at ProgramBench scale.
"But I suspect that the ProgramBench authors are either under-eliciting the AIs, or their tasks are unfair/impossible given the constraints, or both."

I'd go with "impossible":

"Given a gold (reference) executable and its usage documentation, a task worker is asked to write source code and a build script that constructs a candidate executable which should reproduce the behavior of the gold executable."

The test cases are built from an AI doing an examination of the source code and producing test cases, and later text also confirms that the AI during the production phase can't read the original executable so it can't reverse engineer it directly, so the test cases are being drawn from a situation where the tester has vastly more knowledge of the program than the implenter.

That is a losing scenario for anyone, be they human, modern AI, or even some hypothetical perfect programmer. Take ffmpeg as an extreme example. The documentation does not even remotely specify the program. Entire codecs can be missed at a stroke, and each of those codecs is itself a rich set of features that may or may not be used in a given input or output file, but the final tests can freely draw from any of those things. And trying to implement a codec from just some input and output would strain anyone, especially when the input is all but certain to not be sufficiently broad to make the determination for sure.

That sort of issue extends all the way down to even some tiny command-line programs I've written myself. The end-user documentation is never a specification. That's not what end-user documentation is. And even if you did hand the AI all relevant specifications you'd still get an implementation of the specification, but anyone who has ever implemented a non-trivial specification into real-world situations can tell you all about how even the spec is never enough.

I think that's an absolutely ridiculous test. If you handed to me as a human I would simply refuse because I'd tell you straight up front that it is plainly obvious I'm going to utterly and completely fail, so why even bother with the time to try?

I am not surprised but this one sticks out...

> Models favor monolithic, single-file implementations that diverge sharply from human-written code.

Well, all of our code is monolithic with some files close 20K lines of code and we do use coding agents - not for the original code but as of late. I've always had that hunch that splitting everything into tiny files does not improve AI coding agent performance although it feels counterintuitive due to model context constraints.

To me the important parts of a program should be clustered together so the implementation is obvious. Scattering the implementation in various files all over the source tree does not help much building the mental model.

That also closely match how software used to be written in the past too.

> Scattering the implementation in various files all over the source tree

If you treat the source tree seriously, you can communicate a lot with how it is structured

Kinda surprising to me, since i had some trouble with Cursor & Co. once the file went over ~800 lines. It repeatedly failed to edit it, until i split it up into multiple logical components. As it should have been from the beginning...

Though, it was some time ago, so things might have improved?

> Models favor monolithic, single-file implementations that diverge sharply from human-written code.

This isn't the case if models are prompted to actually plan the file architecture beforehand, it's only the case if they're given a dumb monolithic "code this thing" prompt.

this is a big frustration for web code what with HTML, CSS, JS, PHP all spread about

https://htmx.org/essays/locality-of-behaviour/ is a good fight back as exemplified in many stacks, eg https://harcstack.org

> Scattering the implementation in various files all over the source tree does not help much building the mental model.

Yeah, that happens where I work and I hate it. A combination of lint rules and AI reviewer prompts complain about long files and long functions. This means something that could be a 300 line self contained function that could be read linearly, gets split up into 6 functions across 6 files.

It's the illusion of "clean code". If you're casually skimming the code, you feel good. But as soon as you go beyond the surface level it becomes annoying.

> Open internet with cheating detection => cheating is widespread, 20-36% of tasks are flagged for the stronger models, with source code lookup accounting for the majority of the violations.

Therefore:

> blocking internet access entirely is the appropriate default for ProgramBench

The fact that your Anthropic coding assistant has a tendency to search on the Internet code to be inserted into your program may count for an additional copyright violation (besides the possibility of reproducing recognizable fragments of its training data).

(I do not agree that copyright, at least in its current form, should be applicable to computer programs, but it is weird that the same companies who try to exploit copyrights against others also insist on the use of coding assistants that are a workaround against copyright laws, which is the main reason why they can increase programming productivity, because they may cut and paste code that you are not allowed to copy yourself.)

If a photo cannot be copyrighted then dark factory code wont be either.
It’s unfortunate that they didn’t eval using subagents/orchestration for such a complex set of tasks (from what I can tell), e.g. analyze program to produce initial spec -> code -> review and rinse&repeat with each of those steps being a separate subagent allocated

I would be interested to see if there’s a significant quantifiable difference.

This might actually be the whole value prop of this benchmark. Forget their initial scores, take open models (so we can be sure the base doesn't change), and test different combinations of harness + prompts + strategies + whatever memthing is popular today. See if the scores improve. Repeat.
It's interesting that Figure 4 shows that Sonnet and Opus have a very clear distinct curve from all other models, even from GPT 5.4. Anthropic superiority I guess.
In before "but they did not use my agent swarm"
In science N=1 is statistically insignificant. In business it might mean that you have a product.
It’s the annoying thing about AI. If it works, the AI is magic. If it doesn’t work, you’re using it wrong.
I was curious about the variance of the output and I made some runs w/ deepseek v4 flash and found that it was pretty high?

There's also a possibility of strong model memorization on the tasks, I saw a header w/ the authors generated in one of the runs for the cmatrix task (one of 3 of 200 tasks I selected to re-evaluate).

Additionally curious if anyone else thinks that passing in the gold executable as part of the task lowers the usefulness of this benchmark.

caveats: N=5 on my runs and I used my own generalized task prompt

I wonder if a model that does not know anything about a hypothetical programming language X, could write code once given said language X specification, APIs, and SDK tools and their documentation.

Meaning: the model has no idea, no access to examples, no previous codebase trained on, nothing, for language X. But it knows English, it knows how to program in general (training data does contain other programming languages), and everything we expect from LLMs today. It just doesn't know jack about language X.

Neat research. I find figure 11 interesting. The models behave so differently there.

imo the benchmark should be named Can_It_Pull_a_CharDet_Bench

It's funny, because that task is very diverse. Any LLM will use the codebase given as a template(At least in free-tier models)

My software as a contract of behaviors works like a program bench(I even cross tested buildouts) Made an entire corpus layout for multi agent multi platform builds to be compared. Even went ahead and ran 50 contracts for an example. It honestly showed improvable areas, and distinct differences between model code.

{contract_name}/ └── submissions/ └── {date}_{os}_{agent}_{model}_{stack}/ ├── {contract}.osc.md ├── osc.osc.md └── results/ └── {contract}.snapshot.json That's it, compare to the same contract, or find a new contract to use to compare. Lot's of signed/hash pinned files are all you need to reproduce software from nothing, with an LLM.

Programbench is close to that(they have a nice paper/article here. But I don't like the work used. Having software to start with is not a bench of making code but reverse engineering.

github/s1ugh34d/osc

RE: monolithic, single-file implementations

We have a lint that caps source code files at 650 LOC and it works really well.

This is not a serious benchmark, come on.

Tomorrow I'm launching a benchmark where I check if an LLM can build a Airbus A320 from scratch without internet. (Spoiler: no LLM succeeds)

Preinternet people would routinely re-implement unix and get shell scripts working across systems. This benchmark shows that agentic LLMs can't even do that, not just for complex programs and scripts, but for simple programs and simple scripts. 0%. Which fits with claudes' inability to write a c compiler.
Suggested alternative title for the paper:

Can American corporate desires finally kill community based open source once and for all?

I mean, it seems clear to me, companies hate the GPL, and they're willing to play these games to try to get that code into their hands under the MIT license and they're happy to use these thinly disguised methods to get it. I see all these absurd ideas as part and parcel of this larger strategy.

I find the current state of affairs disgusting.

How long until AI is not even writing code but producing machine code?

Think about it, all these compilers, tooling, what a waste!

I imagine a future where chipset makers will provide a model you can just prompt to "act upon that chipset" and voila, "You're absolutely right! Here is your binary."

We won't be developers, we won't be devops, we'll be rollmops! /s

Coding agents can write ASM. But if you mean writing the actual byte-code that will require a very different approach at a very different level of abstraction that LLMs are not designed to do. Keep in mind that all LLMs are trained first on text and then fine-tuned on code.
Good luck reasoning about the output in any meaningful way then. AI introduces a bug? Well, you're fucked.
My hunch is that it would take years of hundreds of thousands of developers working with machine code, posting stackoverflow questions with machine code, and publishing github repos written on it with documentation. Thats all the free labor LLMs leveraged to use high level langs.

>We won't be developers, we won't be devops, we'll be modelops! /s

I can still see this happening with higher level langs. the thing is the compiler is not replaced in the training data, more likely LLMs will give rise to semideterministic layers on the compilers

I could see nvidia achieving this first with how nice the devex is with CUDA