Hacker News Reader

This is classic Goodhart's law. "When a measure becomes a target, it ceases to be a good measure"

https://en.wikipedia.org/wiki/Goodhart%27s_law

It's really not that hard to not build a custom bench setup to game the benchmark instead of just using your product straight out of the box, though.

Right. Building a custom setup is blatant- that will wildly overfit.

But let's say a group uses it as a metric as part of CI and each new idea / feature they create runs against SWE bench. Maybe they have parameterized bits and pieces they adjust, maybe they have multiple candidates datasets for fine tuning, maybe they're choosing between checkpoints.

This will also end up overfitting - especially if done habitually. It might be a great metric and result in a more powerful overall model. Or it might not.

Right, other than financial pressure. Which is, of course, immense.

Also see the VW dieselgate and numerous other "gaming the system" examples.

A specific setup for the benchmark is just plain cheating, not Goodhart’s law.

What are the typical context lengths in SWE-bench problems? Does it partly measure performance in the 64-128k context range?

IIRC the SWE bench dataset gives you the full repo snapshot + the issue text, the evaluation pipelines typically run some kind of retriever (eg. grep, BM25) to pick a subset of files to place in the model’s context. They provided context is usually limited up to ~50k tokens.

This is what the rows look like:

https://huggingface.co/datasets/princeton-nlp/SWE-bench_Veri...

Its up to your retrieval system/model to selectively hunt for relevant context. Here's a few critiques of the benchy:

https://x.com/brhydon/status/1953648884309536958

Is there something in this multi-agent approach that makes the setup more specific to just the test at hand and less general to real engineering tasks? If not, then this multi-agent system will just become what you get out of the box in a future product. Multiple attempts per problem (as long as there's no human intervention or selection between them) is a perfectly fine approach for agents because that's not an issue from the perspective of an engineer using the product. A single agent is already a multi-step usage of LLMs and it sounds like this is just another meta level of that.

https://github.com/auchenberg/volkswagen

I think multiple attempts are completely understandable and even expected? How is that defeating the purpose of the benchmark?

It’s a pass@1 benchmark. When submitting you need to check a box that there was only 1 attempt per problem. See here for example: https://github.com/SWE-bench/experiments/pull/219

Building multiple attempts into your agent is stretching the rules, even if technically it’s acceptable

From my perspective as a potential user the number of attempts is the number of times I have to tell it what to do. If you have an agent that makes a single attempt and is 60% accurate vs another that makes 5 attempts and is 80% accurate, why would you care that each individual attempt of the 2nd model is less accurate than the first?

I think it depends on "But the top rated submissions aren’t running production products" It sounds like they're shipping a product without the debug agent/try-again logic, and that's just for the benchmark, so you wouldn't get the performance they get as a user.

This ok from your perspective then?

def make_pass@1_agent(agent, n):

    def retry_agent(problem):

        for attempt in range(n):

            result = agent(problem)

            if result.success:

                return result

        return result

    return retry_agent

Keep in mind that this isn’t about users - the top agents on the leaderboard aren’t running an actual product on the benchmark.

If they are running their production product as is, then of course whatever is built into the product is fine.

Absolutely fine, as long as the success flag is predicted by the model ensemble under test. That’s how Claude Code works for example, it will continue to iterate until success (or it will give up with failure at a certain point).

Definitely wouldn't have written the code that way, but yes, if (and this is a massive "if") the agent has an accurate and meaningful way to determine which way to set the success boolean. The obvious caveat would be if n needed to be large enough to set the costs higher than I am willing to pay for the additional performance or it makes it take longer than I'm willing to wait.

Think of the agent like an employee. If he delivers the code within the expected time and to the expected quality standards, his process of getting there means almost nothing. Do I care if he tried 4 different approaches along the way and threw out the first 3? Not a bit.

Papers have been doing rollouts that involve a model proposing N solutions and then self-reviewing to choose the best one (prior to the verifier). So far, I think that's been counted as one pass.

I was thinking about this recently with respect to how many agent systems now let you specify a smaller/faster model for easier tasks and a bigger model for harder tasks.

It's interesting to think about what the trade-offs are. Assuming the system can properly classify a task as easy or hard (big "if" but I guess there are ways), there is nonetheless more to think about, depending on your pricing plan.

For subscription pricing, I guess you don't really care which model runs and in fact it's hard to find a reason to ever run the smaller model, so choosing between the models is more in the provider's interests for cost efficiency.

But for pay-per-use pricing, But if you have a bigger model that can get the answer right 80% of the time, and a smaller model that can handle smaller changes and get things right 60% of the time but correct its mistakes, then the system should try to run it on as many tasks as possible to save you money.. but in the end if ends up having to make a lot of corrections, then maybe you end up needing more total requests than the larger model. In that case maybe it's actually cheaper to run the larger model, if it takes fewer requests.

So I wonder how that kind of trade-off could be effectively calculated. I guess if you can figure out when "retries" happen you can count them and do some statistics on which model is more likely to work out in fewer shots. It's pretty complicated though, when you start to think about it in detail.

I do wonder if even having BOTH the smaller and bigger model make hypotheses, and try the smaller model's idea first, then if it fails, try the bigger model's idea, might be the way to go.

According to your experience with this model, is it just trained for the benchmark or these points are actually representing the performance?

Finally someone mentions Refact, I was in contact with the team, rooting for them really.

Just looked them up. Their pricing is around buying "coins" with no transparency as to what that gets. Hard pass

You realize that you can self-host their stuff? https://github.com/smallcloudai/refact

One thing with SWE bench is making sure there's zero leakage of information into the LLM context.

I.e. the agent cannot even know which tests are failing.

It has to both fix the issue based just on the issue text and fix it in the specific way the unit test, which it cannot see, expects.

For this reason I find the benchmark a little disconnected from the reality of software engineering.