Some of the examples in the paper seem to be wrong.
For django-31056, they claim the AI-generated patch is "incomplete" because it's "missing critical parts of this logic, such as the try-except block and the check for a running event loop.". But if you look at the diff, that's clearly wrong. The try-except block and running check were already there before the patch. The human patch just indented them, making them appear as both - and +, while the AI patch didn't. To me, the AI patch seems correct. It's slightly less efficient than the human patch when DJANGO_ALLOW_ASYNC_UNSAFE is set, but slightly more efficient when it isn't (which is the common case!). The human patch does feel more natural, but the AI patch is fine. I'd grade it a tie between human and AI.
For django-32517, they claim that the human and AI patches "produce entirely different outputs", but actually they do exactly the same thing. The human version has `reversed(self.dict)`, while the AI version has `reversed(self.dict.keys())`. `reversed` treats the object as an iterator, and iterating over a dictionary in Python just gives you the keys, so it doesn't matter whether you call `.keys()` first. The human patch is more idiomatic, but it's also more confusing, as shown by the fact that it confused the authors of this paper. I'd grade it another tie.
Edit: I tried to sign up for OpenReview so I could leave a comment about this, but the system wouldn't let me register without completing a form that assumes you have an academic position. Perhaps I should email the authors.
The entire premise of this paper is false. They claim that the "hints_text" is used and leaks the answer in Section 2.1.1; however, the authors of SWE-Bench themselves state that this is not used anywhere (Issue #133 on the official SWE-Bench GitHub).
According to the paper:
> 1. Solution leak: represents instances where the solution to the issue is clearly outlined in the issue
description or comments on GitHub. Since both the issue descriptions and comments (referred to
as hints_text in the SWE-Bench study) are provided as input to the models, these LLM models can
extract the solutions directly from this information instead of generating it independently.
And yet, the SWE-Bench authors themselves explicitly state:
> In short, for participating on the SWE-bench leaderboard, using hints_text in any manner is not allowed. Although we don't explicitly say this in the original paper, we also do not make any mention of using the hints_text anywhere.
So, it's a made up issue that would only occur if you deviated from the paper implementation and explicitly added a field called "hints" that isn't used anywhere.
Hmm. For the example they give of solution leakage, sympy issue 16669 aka sympy__sympy-16766[1], the solution actually appears in problem_statement, so it seems to be genuine leakage. But you're right that they claim that hints_text is used, so they may have improperly winnowed out other instances where the solution only appears in hints_text.
[1] Don't ask me why they cited the issue number, 16669, instead of the pull request number, 16766, when only the latter appears in the dataset. This confused me for a bit.
Although I agree with your analysis and it doesn't look great for the authors, this issue (https://code.djangoproject.com/ticket/32517) arguably falls into their "Solution leak" category anyways, as the following text appears in the issue description (and so I think directly in `problem_statement` rather than `hints_text`):
> Currently, OrderedSet isn't reversible (i.e. allowed to be passed as an argument to Python's reversed()). This would be natural to support given that OrderedSet is ordered. This should be straightforward to add by adding a __reversed__() method to OrderedSet.
It isn't the exact code though, so I suppose it could be argued instead that the issue is just extremely easy.
Interesting analysis! I hadn't dug into the specific patch details like that. It's a good reminder that "correctness" isn't always the only dimension to evaluate these AI-generated patches – readability and idiomatic style definitely matter too, even if the functional outcome is the same.
I've been playing around with some automated code review tools recently, and it's surprising how often they flag things that are technically correct but just... unusual. Style matters, especially for maintainability.
I can only confirm two mistakes in the apper: 1) As you say, the reversed(self.dict) is actually correct; 2) as another poster below said, hints are not part of the input. These two mistakes are so egregious given the objective of the paper that I'm convinced the authors are not qualified to write it.
IMHO, it is probably better to discard this paper, and wait for someone else to cover this important topic.
> When we filtered out these problematic issues, the resolution rate of SWE-Agent+GPT-4 dropped from 12.47% to 3.97%.
This matches my intuition about the coding performance of these models a lot better. I don't think any current coding benchmark accurately measures coding performance.
Anecdotal but I was always shocked to see Claude 3.5 perform so poorly in the benchmarks, when it generates 80% of my code in Cursor (and in cases it fails, no other model succeeds)
Different people seem to get wildly different results here, and I'm not sure what percentage is down to the type of software being built vs the usage patterns.
In my case, I would guess less than 10% of the code I get out of AIs is useful.
What sort of code are you getting those results with? Is it yet-another-react-frontend-button? Is it ebpf programs? Is it a parser in rust?
For the latter two, I've found AI to have pretty low rates, and for the former I haven't had the desire to try.
I've been using LLMs for tab autocomplete for a while and just recently started trying out agentic coding AI (Copilot Edits and Cline). I think the disappointing shortfall of agentic AIs (at least for me) comes from the feedback loop being so much looser than the autocomplete style. With autocomplete, I don't have to actively think about what context to feed it, and I can gently correct it if it goes in the wrong direction on a line-by-line basis. With AI agents, they have a lot more leeway to generate a ton of code and reason themselves off the rails before you're able to step in and correct them. Now granted, I am also not very good yet at managing context and crafting prompts, but it feels a lot harder to get good at than simply dropping an AI autocompleter into an existing programming workflow. It's a new paradigm.
I think the big thing overlooked is how much the human steering the models matters. If you know what you’re doing and what changes you need, cursor and other tools make you so productive.
If you don’t know what you’re doing, these things can sometimes produce good code, and sometimes produce things that don’t work at all
That's been my experience too, but I would guess the problem of "here is a ton of context, produce a small amount of code" is significantly better suited for LLMs than "here is a problem, produce a ton of code".
I write a lot of Python and personally I find Claude significantly worse than OpenAI’s reasoning models. I really feel like this varies a ton language to language.
I personally use Aider's Polyglot Benchmark [0] which is a bit low-key and not gamed just yet. It matches my experience too where Claude Sonnet 3.5 is the best and still beats the new reasoning models like o3-mini, DeepSeek, etc.
Quite the corpus for Exercism tasks that were almost certainly trained on, which could lead this to doing what we know LLM/LRM's are good at...approximate retrieval.
> where the resolution rates of the models drop significantly, which are 0.73%, 0.55%, and 3.83%, respectively.
Matches my experience pretty well as too. It'll usually output something that a novice would assume is correct but an expert can clearly identify as "know it all teenager forum post" level stuff.
Yep anecdotally that's basically spot-on. It's also one of the reasons that I still find copilot vastly more useful than highly autonomous AI tooling (cursor, roocode, avante, etc.)
It is worth reflecting, as much as HN seems to hate the social sciences, on this point. But the difficulty of measuring intelligence is a challenge that several fields have struggled with for decades. It is inherently hard because defining intelligence and building intelligence are very closely coupled. This both makes it hard to make unbiased measures as well making measures that don't affect the phenomenon basically NP hard, or known as the Flynn effect[0].
It also goes to how a lot of people misunderstand the replication crisis. 'Hard science' really should replicate - we should be able filter out sources fo error and variance because the phenomena (generally) isn't affected by our attempts to measure it. Making social science replicate often requires so much control that it is deabstracted from reality, meaning the effort at replication reduces the value and usefulness of the knowledge. Generalizable claims are hard because the sources of variance are so much larger adn more complex. Speaking as someone who went through a transition from engineering to social sciences, it is the concept that made it hard. I started my time in social sciences with a cool idea of a whole carrer based on just doing replication studies, because science. That was...useful and stupid at the same time.
I find the models very useful to chat about library documentation or high level algorithm concepts, but I find the code it generates to be… I don’t know how else to say it… really bad and often out of context.
I know developers who blindly follow the hype and use them to generate production code. That scares the poop emoji out of me, and the code reads like an asset flipped 3D game.
Yeah, that's true in many fields with these AI agents. They demo well, but when you put them to actual work they fall right on their face. Even worse, the harder the task you set for them the more they lie to you. It's like hiring a junior dev from one of those highly regimented societies where it's more important to save face than to get the job done.
It's almost as if they're not trying to market to the people actually using the products, but trying to convince investors of features that don't exist
Your last sentence feels kind of spot on. The lack of transparency around confidence in the answer makes it hard to use (and I know it would not be simple to add such a thing)
To be totally fair, using PhD as a barometer of anything without specifying what is like claiming that LLMs have encyclopedic knowledge while meaning a children's encyclopedia.
1. Did the benchmark authors not review the issues and make sure the solution was not present in the issue?
2. Are the issues locked after they’re included in the dataset? You’d think they would be immutable for reproducibility.
3. For the agents writing patches, is test running part of their inner loop validation? If they write a patch that makes the test pass, then the jobs done. Or is that validation step kept secret from the agent? I don’t see how unless the tests aren’t part of the repo.
>1. Did the benchmark authors not review the issues and make sure the solution was not present in the issue?
I looked at a bunch of issues in the dataset when SWE-verified first game out and I was trying to make scaffolding to solve it and I don't remember a single time where the solution existed verbatim in the issue. I'm not saying it never happens, but it would have to be rare.
> 2. Are the issues locked after they’re included in the dataset?
No one changes the issues in the dataset but of course the original issue on github will have been resolved long ago. The models don't have access to this in their context, but if they were trained on github there's a very real risk that they've seen the solution.
> 3. For the agents writing patches, is test running part of their inner loop validation? If they write a patch that makes the test pass, then the jobs done. Or is that validation step kept secret from the agent? I don’t see how unless the tests aren’t part of the repo.
The tests aren't provided to the model, they are run after the model has proposed its final answer.
This was also my first thought, but reading [1] again, what they did was labeling like:
> Whether we consider the issue description to be underspecified and hence unfair to be testing on.
> Whether the FAIL_TO_PASS unit tests filter out valid solution
and a bit more. This is pointed out in the linked paper too.
The moral of the story to me is that, don't believe the paid human annotator. You can (hopefully) still believe the PhD students doing these unpaid jobs as their research ;-)
Submitted title was "SWE-Bench tainted by answer leakage; real pass rates significantly lower". Normally we'd replace that with the article title, in keeping with the site guideline ("Please use the original title, unless it is misleading or linkbait; don't editorialize."), but in this case the article title is so generic that this is arguably misleading as well, so I took a representative phrase from the abstract instead. That's preferable, because it's better to use the authors' own representation of their article.
If anyone can find a better title (i.e. more accurate and neutral, preferably using language from the article itself) we can change it again.
So what we need is something like a versioned crowdsourced coding LLM eval dataset.
Every quarter, you have a couple thousand volunteers provide 2 GitHub issues from the past 3 months, which are nontrivial to resolve, and where there exists strong test cases. Each volunteer then cross-checks 2 issues from other volunteers. The volunteers get 1 month free subscription to some AI service in return.
This dataset is then published as SWE-UberBench-2025-02 or something. People can then only evaluate their coding LLM on datasets published after their training period.
And how would you ensure that all of them were really volunteers and not colluding with the vendors? Like, tech companies cheating on benchmarks is an old, old story (personal favourite: in the dark ages, before 3D acceleration, some graphics card drivers, on detecting a 2D acceleration benchmark, would _simply draw the wrong thing_), and I wouldn’t trust at least three of the major players as far as I could throw them.
Right, so that AI companies can freely throw this significantly more valuable training data into a model and then turn around and advocate for clamping down on the freedom of models.
> 32.67% of the successful patches involve cheating as the solutions were directly provided in the issue report or the comments.
Looking at the benchmark, https://www.swebench.com/, about half of scored submissions score under 1/3 correct? So they're either not cheating, or not cheating effectively?
LLMs do not reliably reproduce their training data. This is quite easy to demonstrate, every LLM has been trained on all of wikipedia (at minimum) and yet there if you ask it a niche fact mentioned once on wikipedia it is highly likely to get it wrong.
This is why I’m a bit skeptical of the o3 results. If it’s spending a bunch of time reasoning aren’t the chances of it simply regurgitating a solution it saw in its training data at some point in its output stream higher? It still needs to be clever enough to identify it as the correct answer but it’s not as impressive as an original solution.
that comment refers to the test time inference, i.e. what the model is prompted with, not to what it is trained on. this is, of course, also a tricky problem (esp over long context, needle in a haystack), but it should be much easier than memorization.
anyways, another interpretation is that the model needs to also make a decision on if the code in the issue is a reliable fix or not too
Readit News
For django-31056, they claim the AI-generated patch is "incomplete" because it's "missing critical parts of this logic, such as the try-except block and the check for a running event loop.". But if you look at the diff, that's clearly wrong. The try-except block and running check were already there before the patch. The human patch just indented them, making them appear as both - and +, while the AI patch didn't. To me, the AI patch seems correct. It's slightly less efficient than the human patch when DJANGO_ALLOW_ASYNC_UNSAFE is set, but slightly more efficient when it isn't (which is the common case!). The human patch does feel more natural, but the AI patch is fine. I'd grade it a tie between human and AI.
For django-32517, they claim that the human and AI patches "produce entirely different outputs", but actually they do exactly the same thing. The human version has `reversed(self.dict)`, while the AI version has `reversed(self.dict.keys())`. `reversed` treats the object as an iterator, and iterating over a dictionary in Python just gives you the keys, so it doesn't matter whether you call `.keys()` first. The human patch is more idiomatic, but it's also more confusing, as shown by the fact that it confused the authors of this paper. I'd grade it another tie.
Edit: I tried to sign up for OpenReview so I could leave a comment about this, but the system wouldn't let me register without completing a form that assumes you have an academic position. Perhaps I should email the authors.
According to the paper:
> 1. Solution leak: represents instances where the solution to the issue is clearly outlined in the issue description or comments on GitHub. Since both the issue descriptions and comments (referred to as hints_text in the SWE-Bench study) are provided as input to the models, these LLM models can extract the solutions directly from this information instead of generating it independently.
And yet, the SWE-Bench authors themselves explicitly state:
> In short, for participating on the SWE-bench leaderboard, using hints_text in any manner is not allowed. Although we don't explicitly say this in the original paper, we also do not make any mention of using the hints_text anywhere.
So, it's a made up issue that would only occur if you deviated from the paper implementation and explicitly added a field called "hints" that isn't used anywhere.
[1] Don't ask me why they cited the issue number, 16669, instead of the pull request number, 16766, when only the latter appears in the dataset. This confused me for a bit.
Although I agree with your analysis and it doesn't look great for the authors, this issue (https://code.djangoproject.com/ticket/32517) arguably falls into their "Solution leak" category anyways, as the following text appears in the issue description (and so I think directly in `problem_statement` rather than `hints_text`):
> Currently, OrderedSet isn't reversible (i.e. allowed to be passed as an argument to Python's reversed()). This would be natural to support given that OrderedSet is ordered. This should be straightforward to add by adding a __reversed__() method to OrderedSet.
It isn't the exact code though, so I suppose it could be argued instead that the issue is just extremely easy.
I've been playing around with some automated code review tools recently, and it's surprising how often they flag things that are technically correct but just... unusual. Style matters, especially for maintainability.
IMHO, it is probably better to discard this paper, and wait for someone else to cover this important topic.
Deleted Comment
Deleted Comment
This matches my intuition about the coding performance of these models a lot better. I don't think any current coding benchmark accurately measures coding performance.
In my case, I would guess less than 10% of the code I get out of AIs is useful.
What sort of code are you getting those results with? Is it yet-another-react-frontend-button? Is it ebpf programs? Is it a parser in rust?
For the latter two, I've found AI to have pretty low rates, and for the former I haven't had the desire to try.
If you don’t know what you’re doing, these things can sometimes produce good code, and sometimes produce things that don’t work at all
0. https://aider.chat/docs/leaderboards/
https://github.com/search?q=Exercism&type=repositories
Matches my experience pretty well as too. It'll usually output something that a novice would assume is correct but an expert can clearly identify as "know it all teenager forum post" level stuff.
It also goes to how a lot of people misunderstand the replication crisis. 'Hard science' really should replicate - we should be able filter out sources fo error and variance because the phenomena (generally) isn't affected by our attempts to measure it. Making social science replicate often requires so much control that it is deabstracted from reality, meaning the effort at replication reduces the value and usefulness of the knowledge. Generalizable claims are hard because the sources of variance are so much larger adn more complex. Speaking as someone who went through a transition from engineering to social sciences, it is the concept that made it hard. I started my time in social sciences with a cool idea of a whole carrer based on just doing replication studies, because science. That was...useful and stupid at the same time.
[0] https://en.wikipedia.org/wiki/Flynn_effect
I find the models very useful to chat about library documentation or high level algorithm concepts, but I find the code it generates to be… I don’t know how else to say it… really bad and often out of context.
I know developers who blindly follow the hype and use them to generate production code. That scares the poop emoji out of me, and the code reads like an asset flipped 3D game.
OAI, xAI, Antropic, Google all score incredibly well, then you go to try and write code and its just okay.
They claim it can do PHD level reasoning, but here I am not trusting it on basic computational thinking.
Not sure that's really the claim. I think they claim that performance on benchmarks like GPQA indicate PhD level knowledge of different fields.
1. Did the benchmark authors not review the issues and make sure the solution was not present in the issue?
2. Are the issues locked after they’re included in the dataset? You’d think they would be immutable for reproducibility.
3. For the agents writing patches, is test running part of their inner loop validation? If they write a patch that makes the test pass, then the jobs done. Or is that validation step kept secret from the agent? I don’t see how unless the tests aren’t part of the repo.
I looked at a bunch of issues in the dataset when SWE-verified first game out and I was trying to make scaffolding to solve it and I don't remember a single time where the solution existed verbatim in the issue. I'm not saying it never happens, but it would have to be rare.
> 2. Are the issues locked after they’re included in the dataset?
No one changes the issues in the dataset but of course the original issue on github will have been resolved long ago. The models don't have access to this in their context, but if they were trained on github there's a very real risk that they've seen the solution.
> 3. For the agents writing patches, is test running part of their inner loop validation? If they write a patch that makes the test pass, then the jobs done. Or is that validation step kept secret from the agent? I don’t see how unless the tests aren’t part of the repo.
The tests aren't provided to the model, they are run after the model has proposed its final answer.
> Whether we consider the issue description to be underspecified and hence unfair to be testing on. > Whether the FAIL_TO_PASS unit tests filter out valid solution
and a bit more. This is pointed out in the linked paper too.
The moral of the story to me is that, don't believe the paid human annotator. You can (hopefully) still believe the PhD students doing these unpaid jobs as their research ;-)
[1] https://openai.com/index/introducing-swe-bench-verified/
If anyone can find a better title (i.e. more accurate and neutral, preferably using language from the article itself) we can change it again.
https://news.ycombinator.com/newsguidelines.html
Every quarter, you have a couple thousand volunteers provide 2 GitHub issues from the past 3 months, which are nontrivial to resolve, and where there exists strong test cases. Each volunteer then cross-checks 2 issues from other volunteers. The volunteers get 1 month free subscription to some AI service in return.
This dataset is then published as SWE-UberBench-2025-02 or something. People can then only evaluate their coding LLM on datasets published after their training period.
1) No known solutions, so there's no "ground truth" dataset to train on
2) Presumably hard to solve
3) But easy to verify a solution if one is provided.
This, of course, is easier done on the STEM side of things, but how do you automatically test creativity, or philosophical aptitude?
Looking at the benchmark, https://www.swebench.com/, about half of scored submissions score under 1/3 correct? So they're either not cheating, or not cheating effectively?
anyways, another interpretation is that the model needs to also make a decision on if the code in the issue is a reliable fix or not too
Smaller ones don't.