back
120 comments
I saw on Twitter that in an ML course at Tsinghua University, one of the tests asks students to write quizzes that fail the most LLM models as possible.

What if we create a benchmark that works like this and assigns ELO scores? Models fight head-to-head by writing a question, a bug, or an incomplete implementation, which the opponent has to answer, fix, or finish.

We could call this "generative adversarial network" (GAN) :)

https://en.wikipedia.org/wiki/Generative_adversarial_network

How do you prevent degenerate strategies? I could trivially give a model a SHA256 hash and ask it to provide the source input.

In class you'd probably want a rule saying at least one LLM should be able to figure out the answer, but in a head-to-head I'm not sure how to solve it.

We tried this, and it works :) https://arxiv.org/abs/2508.06111

You have to be careful about degenerate / duplicate Qs, as a sibling commenter mentioned.

Recently though, we found that reasoning models have trouble making code-output-prediction tasks (the initial family of verifiable tasks we started with) which other reasoning models can't solve.

We started looking into harder / more agentic tasks (e.g. passing tests, using AISI's Inspect framework) but deprioritised.

That was Fudan I think
Staff SWE Bench: LLM doubts whether we should do any of this, calls the entire project into question, refuses to merge code, but is happy to delete it.
Principal version: similar, but also says the only acceptable approach is to do it like they did it at their last company.
You jest but I indeed find rejection an integral part of the job. Not plainly saying "no, get away", but backing up, requesting big picture views and trying to see if the overall organization is in need of and capable of carrying said project long-term feels like the absolute minimum that needs doing before we even begin.

I suspect LLMs can do this just fine and probably better than us, but they do need to be trained specifically for it and I have a hard time coming up with good sources of training data for it.

I wonder how they're planning for the benchmark to stay relevant over time.

If the benchmark is to implement features that are part of an open source project, and LLMs have those changes as part of their training dataset, it seems that they could just give a verbatim or slightly modified version of the change in their training data.

And if one updates the benchmark to only incorporate code changes that are past the models knowledge cutoff, then the benchmark is less comparable over time, since the changes in the benchmark at time T and T+1 aren't the same.

How far back in time do you need to give though? I imagine it will be limited by the time to think of new problems. Still seems valuable.
This makes so much sense as to why I've always felt that Opus 4.8 was leagues ahead of GPT 5.5. It's so good at taking underspecified requirements and filling in the gaps with sensible approaches for your project
Why supply underspecified requirements in the first place? Both models are good at challenging assumptions/edge cases and asking questions to clarify, but seemingly only when explicitly asked (i.e. something like a "brainstorm" skill).

I don't think either harnesses do enough to encourage the model to challenge all assumptions and ask questions, maybe because users might find it annoying. That step is basically a requirement IMO.

I've found all of the GPT-5 models to be very nit-picky, useful for code review and mathematics (important for my work), but seemingly gets in the way of "aesthetic" code, e.g. overly defensive code to cover all edge cases, even if unlikely.

There is seemingly also a tradeoff between flexibility vs instruction following. In my experience Opus will sometimes ignore instructions but can "fill in the blanks" more, vs GPT-5.5 follows instructions better but perhaps at the cost of rigidity.

The best benchmarks are the ones you create yourself.

Its not my experience Opus is leagues ahead or even superior, but in any case, since GPT 5.5 has Instant, Medium, High, Extra High and Pro...Should the comparison be with GPT on Pro, instead of Extra High as it seems to be the case in the table?

same observation here opus 4.8 (and i dont understand the people defending gpt 5.5 constantly) was significantly mature, it would even push back against anything off putting where as GPT 5.5 will happily agree and do what is asked but I would note that it takes several tries.

4.8 also requires more than one prompt but its output is significantly higher quality and offers more insight

Fable 5 is a different beast however.

> It's so good at taking underspecified requirements and filling in the gaps with sensible approaches for your project.

At a high level. It misses low level or other non-functional requirements differently so I wouldn't say Opus is just strictly better.

It's also possible that it's just a harness problem more than model.

Man I don't know if I'm living in a crazy bubble or something but GPT 5.5 is lightyears better than Opus 4.8 for me to the point where I'm honestly wondering how you're evaluating them or what kind of work you're doing.

There's specific tasks that Opus does better on like Frontend Dev and Design but for anything else 5.5 just laps it.

Better for vibe coders who always under specify. But at what point does it know you are under specifying but you have properly specified and it did it over your specification?
Similarly, it explains to me why people found Claude so amazing, while I just thought "eh."

Tool expectations

Top solve rate is currently 24% with Opus 4.8... What's a competent human supposed to score?
presumably whatever the top model uses and then some, since the human can use the model.

I wonder if a model could score higher if it had a human at its disposal?

I mean these were all solved before I assume so 100% not the same human ofc but models are expected to be good at a variety of code bases while human can specialize in one and learn. I think it's fair to compare to an individual that is used to working on a product.

I'm more interested in how fable would do

It's nice to see a new public benchmark from Snorkel. They're doing some pretty sophisticated stuff over there.
The value of a senior situation is to apply known solutions and strategies, to novel problems. I can not see how any benchmark, without ever changing, can provide a novel challenge for long.

Any decent benchmark would use the whole of TRIZ to generate a giant ball of a problem first and watch a AI deduce a optimal solution.

If something like this works wouldn't that imply technical interviews can be automated?
Using LLMs to judge this could be very hackable, shouldnt it be? Is that the best practice we expect? Would be very interested in see some ablations about failure modes where the LLMs tried to hack it somehow. Or failures from the llmaaj
> You are a senior SWE-Bench reviewer, make no mistakes.

I don't know what a better approach would look like while still remaining feasible, however this approach of telling a LLM to make a subjective judgement seems fundamentally flawed.

More importantly, I suspect this actually hinders the work. If the LLM does make a mistake, it's now incentivized to downplay it instead of acknowledging and correcting.
The “make no mistakes” admonition does seem pretty silly (it’s been skewered to death on yt), but… it’s easy to imagine how it might work. E.g. it could be interpreted as simply as “check your work”.

Of course, no-one seems to be (publicly) doing the comparative measurements that might allow us to reach rational conclusions here.

This approach is effectively seeding the context with how you want the LLM to behave/operate ("senior reviewer", i.e. the style of the responses you want) and the context/domain in which the LLM is operating in ("SWE-Bench").

This is common in system prompts and frames the responses.

For example, you'd get different responses saying:

1. you are a pirate writing sea shanties about programming;

2. you are a news reporter writing an article on physics;

3. you are a senior software engineer with complete knowledge of PostgreSQL.

For 1 you could get responses along the lines of the Wellerman sea shanty -- "There once was a program that was set to C ...".

The "make no mistakes" bit does look dubious. It would be interesting comparing the results with and without that bit and trying alternative ways of getting the same desired behavior.

Interesting how close sonnet 5 comes to opus 4.8
Benchmarks are great, but I feel like there’s a better way this seems quite subjective.

What you really need is an objective benchmark

I actually really like subjective benchmarks, so long as it's a human (ideally me) grading the results. LLM as judge never made much sense.
> What you really need is an objective benchmark

"When are all the software engineers unemployed?"

> Senior engineers build features without over-specified requirements

To me this already disqualifies the benchmark. That statement is missing the most critical piece about senior engineers: the senior engineers know how to obtain input for their work on their own whether that talking to customers or using metrics. Never ever they come up with stuff on their own - that’s junior behaviour.

Until a coding agent will be able to *gather* the input on its own, its never going to be „senior”

I'd take this a step further, but that step also curls back to the other side a small bit.

The real skill is being able to both pull the necessary information from these sources as well as being able to intuit gaps in that knowledge based on their understanding of the business and their domain expertise & wisdom. Sometimes you can't get a perfect picture, sometimes the people who should know aren't able to tell you what they really need. You still need to do the right thing.

A benchmark like this can potentially do the second part. But I don't think any model would be good at it, for now.

can you prove that senior engineers can pass this benchmark?
Isn't being open source creating incentives for the AI companies to optimize their LLMs for the specific benchmark? I thought all those benchmarks are deliberately closed source primarily for this reason.
Once again I am asking: who are these people and what makes them more qualified than any of you to asses anyone or anything "as a senior engineer" (with the subtext being that none of you are, either)
fable 5?
Why didn't they just make it "Staff SWE-Bench", would be much better smh. /s

But seriously, as an industry we're terrible at assessing engineering levels, I've worked with "senior engineers" who can't code and I've worked with "junior engineers" who could run rings around them.

Benchmarks like this should be much more precise about what they're actually testing, and what axes they're hard on. We also need to rise above prompts like "you are a senior engineer", it's woo, and it's far better to ask for precise outcomes.

[flagged]
The "tasteful solves" is codified cargo culting. The software industry has a tendency to anthropomorphize software while playing to the ego of the programmer. The programmer imagines they are creating a "beautiful" artistic expression. Good code becomes "tasteful", as a software artist must have "good taste" to tell the good software from the bad software. Good quality lacks "bad smells", because a good artist has fine senses (and everybody must like the same smells). "Fine craftsmanship", in code as in woodworking, means your finely-crafted work is "technically superior", so you can charge more money for something that could've been made cheaper and faster and done the same thing.

But it's a lie. Nobody's paying you to make paintings. They're paying you to build machines. The comparison between "making working software" with "taste" always devolves into bikeshedding and subjective opinionism, uses subjective human feelings to describe what should be objective and functional, isn't rooted in scientific rigor, and detracts from the real purpose of the thing. The work doesn't actually get better by trying to apply artistic principles to engineering. It just feels better for the people making it.

Once you make the machine work, then you can go about gilding the lily. But this is unromantic, unsatisfying, boring. Since the inmates run this particular asylum, we end up with a benchmark that tries to accurately mimic the human ego as applied to software design. Thus the new Gods create their digital Adams and Eves in their image.

Sounds more like vibe-bench.

For any professional work you care about the details.

Even for hobby work, if you are using LLMs then presumably it is to do the drudge work of coding, not making the decisions, and that goes doubly so if you are a senior developer. Sure the LLM can "fill in the details" and vibe code (or attempt to) you a compiler or whatever, but the whole reason you are doing a hobby project is presumably because you want to bring your experience to bear and build a GOOD compiler, not a generic one.

I think benchmarks like this are too subjective and narrow to be useful. For example, whether a patch "bloats" the codebase really depends on the situation: If it's building a feature that will grow in the future, or refactoring code that has a long history of bugs, then a larger patch might in fact be good. It's not clear from the blog just how much context the LLM judge receives about the long term project goals and history. Benchmarks should be focused on evaluating the final result only. Maybe ask the coder to build a full app, or implement many new large features for an existing app in sequence, with a larger set of requirements, or have another LLM roleplay as the human to make the instructions a little more underspecified. When done, ask a reviewer harness to test the product for 5 hours, not the code. Count the number of bugs and weigh them by severity. "Taste" would then become an automatic consequence of correctness.

(Full disclosure, I'm not a software engineer.)