[sanderjd]

Sure, for casual evaluation, I agree. But are there serious analyses that are evaluating this kind of thing? I mean, these are the kinds of things I evaluate in my own work when a new model comes out, or when I'm evaluating a harness. But this is all very ad hoc and intuitional. I'd love to start bringing rigor to it, but I haven't found much prior art on this. In another thread someone said that's because it's probably impossible to do this rigorously because too much of it is subjective. And that does match my intuition. But I continue to suspect that intuition is wrong.

[jerf]

It's hard to bring much rigor to it. I'm not saying impossible, but it's not like it's completely obvious how to do it and people are just too lazy. Intrinsically, if I'm going to test a back-and-forth with a model I have a human in the loop making frequent decisions. Did the model fail or succeed at whatever rate it did that because of the model or the human? Did the testing protocols capture the actual problem, e.g., maybe if the model was given some particular bit of information that a normal human would have given it it would have done much better or worse, but the testing protocol in the interests of "rigor" excluded the human in the loop from doing it. Is the human going to be willing to sit down and do the same task 25 times, refreshing the model from scratch each time for a "valid" test? Can you get the same human to analyze every model in the test? Is their 10th pass of the problem an invalid test because you can't as easily erase the human's knowledge of the previous 9 tests? What do you do with a model that succeeds wildly 75% of the time and spins off into a loop the other 25%? Is that loop real or, again, did your "rigorous" testing protocol prevent the human from saving the model from the loop like any developer would?

And so on and so forth. Again, I'm not saying this is impossible but I am saying that if you tried to do it, and you got the money, and you built the test, and got the human subjects clearance, and you ignored that during the process of all that at least one more frontier model would come out, you can count on HN anklebiting your "rigorous" study even so, and probably being correct about a lot of the issues it could have because it would take several iterations of this to build a reasonable protocol... at which point it would quite possibly also be obsoleted by progress again.

[abhgh]

You usually see this kind of analyses in conference papers, esp. if they have a datasets track. The NeurIPS Datasets & Benchmarks (D&B) track is a good example. But you will have to monitor the proceedings yourself closely - there is little chance of being accidentally exposed to them, because most blogs, announcements and popular media only mention a handful of the popular ones, e.g., Tau^2. For ex., across the years 2022, 2023 and 2024, 900+ papers were accepted in the D&B track [1] - of course, not all of them are LLM-related. I find them interesting because they often focus on specific system behaviors, and like you said, study them scientifically, so you can draw authoritative conclusions (or at least know specifically what part of a model's behavior you now know about, and what parts you don't).

[1] https://blog.neurips.cc/2025/09/30/reflecting-on-the-2025-re...

[rileyphone]

DeepSWE is closer to that

https://deepswe.datacurve.ai/