DeepSWE: A Contamination-Resistant Coding Benchmark — James Shi, Datacurve

Talk transcript

1. 0:00

   [upbeat music] Hey, everyone. Can you guys hear me okay? This is good?

2. 0:15

   Um, yeah, my name is James. I'm one of the founding engineers at Datacurve. Unfortunately, Serena's been out with a fever for the past couple of days. Um, she was supposed to be here giving this talk, um, so I'm just filling in, uh, in her place.

3. 0:28

   But I've been at, uh, Datacurve working on the research and engineering side of things, um, as well as DeepSuite, which is our, uh, frontier long-horizon coding benchmark, which you guys may be familiar.

4. 0:40

   Um, I'll just be going over, you know, some of the most important findings about DeepSuite, um, a brief overview of what it is for those of you who don't-- who, who may not know.

5. 0:49

   Um, and then going deeper into our methodology and exactly how we came about this, uh, uh, frontier coding benchmark. So, um, DeepSuite is a long-horizon software engineering, software engineering benchmark comprised of one hundred and thirteen original software engineering tasks.

6. 1:07

   So this means, unlike something like SWE-bench Pro, we didn't scrape this from, um, existing PRs that have been closed. Um, there's a variety of benefits for this, uh, namely one of them is to resist against contamination and agents being able to cheat, uh, through the course of their rollouts.

7. 1:22

   Um, SWE-bench has-- or SWE-bench Pro, uh, pulls thousands of tasks from only forty repositories. The median task per repository for us is one. So you can see across over a hundred tasks, we pull from nearly a hundred repositories, and the language spans across, uh, TypeScript, JavaScript, Python, Rust, and Go, and we have plans to add, uh, more

8. 1:44

   languages later on. Um, since its, uh, release, we've received, uh, very positive reception. It's replaced SWE-bench Pro in the artificial analysis, uh, coding, um, agent index, as well as being cited by, uh, numerous frontier model labs and us helping with them in, uh, tracking their models on our benchmark as well.

9. 2:05

   So we've been really, really appreciative of that.

10. 2:08

    Uh, a bit of context about us. Datacurve works on building, uh, training data for high-ceiling domains, including coding as well as coding-adjacent fields. Uh, we also are trying to answer the very, uh, elusive question of what exactly makes, uh, good data, what is data quality, and how can we demonstrate that tr-- our training data, in fact, moves

11. 2:31

    the needle. So DeepSuite is one in a long line of initiatives that we have towards answering this question.

12. 2:40

    So why did we create DeepSuite? Well, it's very clear that the existing benchmarks are not hitting the mark. Um, with benches like SWE-bench Pro, uh, top models are clustering at the top.

13. 2:51

    It's very hard to differentiate between, uh, which one is good because they all have overlapping confidence intervals. Uh, contamination is also rampant because, again, all of these tasks are mined from public PRs, so all the solution tests, even the discussion around the PRs, those are all available out in the wild for these agents to access.

14. 3:10

    The verifiers are also very, very brittle because we're anchoring them to a specific implementation, often derived from the PR that was merged in. And oftentimes, you also have tests that check for private helpers and functions, uh, created by the task author, which is, uh, very opinionated, right?

15. 3:27

    And it's not something that models should have to adhere to. Um, and finally, uh, leakage. So one thing about SWE-bench Pro is, uh, for very, uh, insightful models such as Claude, they're able to directly run Git log and then go through the commit hashes and cherry-pick the ones out that contain the golden patches, which again, very, very

16. 3:45

    serious issue. So this is DeepSuite. This is the updated leaderboard as of July first. You can see, um, I was mentioning before, uh, the problem of differentiating, uh, but you can see on DeepSuite here, there is a very clear difference.

17. 3:59

    There is a very clear performance gap between the top-performing models versus, you know, at tenth place you have Gemini 3.1 Pro. Um, also within the Claude and the GPT models as well, we're able to see some, uh, deviance.

18. 4:12

    And, um, yeah, if you go on deepsuite.datacurve.ai, you'll also be able to see, um, the token efficiency, uh, costs, um, token usage, context window, peak context, all of that stuff on the DeepSuite site as well.

19. 4:26

    But yeah, as of July first, uh, Fable 5 is retaining the top spot on our leaderboard.

20. 4:33

    So, um, the ranking information is available online. I wanted to talk about some of the qualitative insights into how these different models are performing, which I think is the most interesting part.

21. 4:46

    Um, starting with the first one is we find Claude is generally a very, very, um, thorough and exhaustive model. It will, um, try to explore everything, including, uh, go through all of the Git logs.

22. 4:57

    Uh, so one interesting insight was seeing that it becomes quite forgetful when it comes to multi-part prompts. So when you tell it, uh, within the scope of a task, let's say, to in-- support both synchronous and async versions of calling a hook, it will go ahead and implement the synchronous part, but it may drop the asynchronous part.

23. 5:19

    We observed this in roughly two out of three Claude rollouts across all of the, uh, trials, all of the rollouts that we ran. So this was definitely quite interesting because from my experiences and developers I've talked to as well, Claude is generally very, very thorou-thorough and able to get at the developer's intent quite well.

24. 5:38

    Um, another thing about Claude is it pays very close attention to its environment, so it will often run-- Uh, this is taken from the trials, uh, we ran ourselves independently and also from examining SWE-bench Pro.

25. 5:50

    It will attempt to run Git log and recover the golden patch from the Git history. We found that for Opus 4.6 and 4.7, it did this twenty-five percent and eighteen percent of the time respectively, compared to all the Gemini models, uh, averaging at roughly one percent of the time, and we found zero instances of this for the

26. 6:11

    GPT models. So, um, thankfully within DeepSWE 1.1, we safeguarded further against models being able to cheat by pulling from the Git history. Um, but this was something, yeah, we observed quite frequently for Claude within, um, the SWE-bench Pro rollouts.

27. 6:29

    Um, third finding is that GPT is very good at implementing exactly what it is asked. Across our failure mode analysis, we found that it was the least likely model to miss requirements.

28. 6:42

    Um, GPT-5.4 was the second-best model at this ranking, only behind GPT-5.5. It always, uh, learns to read the prompts and the re-caus- repository contract very literally, and producing a patch that honors the existing conventions, um, signatures within the repository, which is very helpful.

29. 7:00

    And we found that these traits converge across all rollouts. So these were not just, uh, lucky attempts, but on average, like, this was the, uh, favorable, um, behavior exhibited by GPT.

30. 7:13

    Um, and finally, we found that on average, stronger men-- stronger models have a great tendency to want to test their own work, um, but with a caveat. In SWE-bench Pro's template, they explicitly tell the model that the tests are handled, and therefore they do not need to, uh, write, uh, any new tests of their own.

31. 7:34

    With that single line in the prompt, it will, uh, prevent the models from e-even, uh, 5.5 and, uh, Opus 4.8 from attempting to verify its own work through the course of the rollout.

32. 7:45

    Um, in DeepSWE, we do not have anything that says to write or to not to write tests, and so we observe this, uh, divergence between, uh, the percentage of the time where these models are actually engaging in writing tests.

33. 7:58

    So this is quite an important behavior as it can, uh, provide-- The models are trying to obtain their own ways to verify and validate their work through the course of a rollout.

34. 8:08

    Uh, we find on average that stronger models like 5.4, 4.7 exhibit this the majority of the time, whereas, uh, models like 3 Flash and 3.1 Pro are far less frequently, um, willing to test their own work.

35. 8:22

    Um, yeah, so, uh, takeaways from the findings, I think, is, is very interesting how, um, yeah, str-stronger models on average exhibit, um, or converge on these behaviors. Um, so moving on to the tasks, right?

36. 8:37

    The, the methodology behind DeepSWE. Um, we made a decision to want to have every task authored, uh, from scratch, uh, rather than being mined. Uh, aside from the, uh, issues with contamination that we mentioned previously, um, this also plays into, uh, one of our core strengths, which is that we offer a bespoke platform where we have, um,

37. 8:59

    software engineers, machine learning enthusiasts come on and create these challenges and compete against one another. Um, this platform is like Shipped and-- Or, or sorry, this platform is called Shipped, and we have a version of this platform for every single domain that we're interested in.

38. 9:14

    For example, for software engineering, it takes a lot after a Codeforces or GitHub, um, and we're really looking for enthusiasts. So these are oftentimes open-source engineers who are core contributors or maintainers of the projects that they're actively making tasks for.

39. 9:31

    So by creating these tasks from scratch, we know that the outputs are intrinsically aligned with our objective of providing a fair and comprehensive test to models. We also know that these people has-- have very thorough understandings of the repository's philosophy and the existing conventions, so they can make tasks that are both realistic in terms of the prompt,

40. 9:52

    but also realistic in the sense that, uh, this is an actual PR that you might see getting merged into the repositories.

41. 10:00

    Um, another, uh, very important design decision is we try as much as possible to make our prompts read like real tasks. On average, the average prompt, uh, characters within SWE-bench Pro is over four thousand five hundred characters, whereas for us it's, uh, roughly half of that.

42. 10:17

    And it's-- This is important because when you're prompting, say, a junior engineer, or you're prompting a model to solve a very, uh, high-ceiling ambiguous task, you're not gonna be coming in there with a to-do list, uh, telling it to, "Oh, first do this and then do this, and then write this function signature in exactly, uh, this way

43. 10:37

    that I've prescribed onto you." Oftentimes, you're going to give it the high-level objective, get it to explore, and get it to reason about to, uh, the list of to-dos and ultimately to the solution on its own.

44. 10:48

    Um, so this was not the case in SWE-bench Pro. It's very o-overly verbose and trying to prescribe a certain solution method onto agents. Um, as much as we could, we try and make DeepSWE prompts as terse and as, uh, high level as possible, um, mirroring what you might see in the real world if you were to prompt,

45. 11:07

    say, another, another engineer or one of your agents to go and solve an engineering task.

46. 11:14

    Um, so even though our prompts are short, we still are able to maintain the long-horizon nature of these tasks. Even with our prompts, again, being roughly half the size of SWE-bench Pro's, we find that the average size of our solution is five times the lines of code, um, compared to SWE-bench Pro's.

47. 11:31

    Um, we also verified that, uh, there are on average seven files being touched in the agent's solution. Um, and across the course of a rollout, we have two times more output tokens being emitted.

48. 11:43

    Um, and, um, finally, we have, uh, we-- Verifier design is, of course, one of the most, um, important and tricky parts of building good environments. In SWE-bench Pro, we have these verifiers that are testing, again, for specific implementations.

49. 12:00

    It will fail the model if it, uh, produces a function that may address the objective but is not named or is not defined within a specific module, or if there's the absence of specific helpers or other private functions.

50. 12:14

    Um, because again, these are derived from the solutions that were merged in the actual PR. Um, so for us, we want to, uh, emphasize on the observable behavior as much as possible.

51. 12:26

    We want to ensure that any correct implementation, uh, anything that correctly solves the problem is rewarded, and this will prevent against, uh, false negatives. We also make sure that, um, there is the absence of these PR-derived tests that rely on naming, relying on specific implementations.

52. 12:43

    Um, and so this will prevent again, again-- uh, again, uh, towards false negatives as well. We observed through a combination of these, um, you know, considerations, we're able to drastically reduce the false negative as well as the false positive rates when we analyzed, um, our rollouts compared to SWE-bench Pro's, uh, using both human experts as well as

53. 13:05

    LLM-as-a-judge. And yes, the coverage for us spans across these 91 repositories. Um, our criteria for these repositories was, uh, ones that had more than 500 stars on GitHub. Uh, they are actively being contributed towards and for, for our pool of, uh, uh, subject matter experts to validate that these are, you know, in fact repositories that

54. 13:30

    are actively used in, in, in the r- in the real world and are representative and can field, uh, real-world, uh, and realistic software engineering tasks.

55. 13:41

    Um, but with all that said, um, there's still a lot of work to be done for DeepSWE and for benchmarks in general. One of the things that we outlined in our blog is our, uh, choice to use MiniSWE-Agent, which is an agent-agnostic harness.

56. 13:55

    The reason why here is we really wanna be focusing on the model's base performance, and so we use MiniSWE-Agent also, uh, ran rollouts to test that the, uh, performance is comparable both using MiniSWE and against each model's native harness.

57. 14:11

    But I think there's a lot of work to be done in the future for benchmarks that focus solely or more so on harnesses and comparing the effects that these harnesses, whether it's native or third-party ones like MiniSWE, um, towards the efficiency and the output of these models.

58. 14:28

    Another thing we wanna improve on is task mix. So given that we are targeting long-horizon tasks, naturally this meant that there's less, um, less, uh, emphasis on bug localization and refactoring.

59. 14:41

    These are obviously very representative of real work that software engineers are doing, uh, underrepresented in our current taxonomy, uh, for DeepSWE. Um, and finally, repository pool. Um, we put an emphasis on trying to field as many diverse repositories as possible, keeping the median task per repository to a very low count.

60. 15:01

    But, uh, further work here just to pull in more t-- uh, more repos, more tasks that software engineers find interesting and find them to be good and maybe also, um, more niche tas-tests of models' performance would also be a great addition here.

61. 15:18

    So we've already, uh, released DeepSWE v1.1. So in here, we've taken some additional measures to guard against cheating, uh, reward hacking, um, by ensuring, you know, the verifier runtime is fully separate now from the agent runtime.

62. 15:32

    Um, also making sure the test reports are in a more standardized format, and also making sure that we've trimmed all of the Git refs and the commits besides the, uh, base commit that our agents are working on.

63. 15:46

    So all of this in service of just making the environments more robust and more cheating-proof. But, uh, as I mentioned, looking ahead, we wanna support an even greater diversity of, like, task, uh, tasks, uh, corpus.

64. 15:58

    We also wanna look into hybrid verification because if we're able to use LLM-as-a-judge or other, um, methodologies, it's possible for us to make our prompts even more terse and even more, um, even more, uh, high level and focused on the objective rather than prescribing anything onto the agent.

65. 16:18

    Um, there is of course like a certain degree that we have to in our current prompts like, um, hint the agents, steering them towards a current methodology just because otherwise they, they may not be well-positioned at all to make meaningful progress towards the task.

66. 16:32

    Um, but something like LLM-as-a-judges and hybrid verifiers would potentially help us towards that. And beyond DeepSWE, we're also working on new benchmarks that are in the works. These are again focused on the high-value domains that, uh, Datacurve reprioritize, uh, as being the domains where we wanna be, uh, most meaningfully advancing model capabilities.

67. 16:55

    Um, but with that said, we're actively hiring, uh, both researchers, engineers, helping us with these new benchmarks, new training data pipelines in service of advancing these capabilities. So, um, definitely reach out at datacurve.ai/careers.

68. 17:09

    Um, and yeah, if you're interested about any of this research benchmark or any of our works, uh, come find me after. Thank you very much. [audience applauding] [outro jingle]