Sounds promising! Curious about whether you have plans to accept papers based on experimental setup instead of results (to reduce publication bias) and if you’ll consider a “press abstract” designed to help journalists disseminate information to the broader public?
Hmm. Ultimately it would be up to the editorial board, but here’s why I personally think these features are probably low priority given their nontrivial cost: (1) I presume we are talking about numerical experiments, and I expect the foundational/conceptual topics we want to publish on are less vulnerable to publication bias than, say, experimental psychology or economics. It would be more like pre-registering numerical math papers. That said, if you think the alignment literature has big problems with publication bias, I’d be interested to hear more. (2) Our primary audience is other researchers. Often, journals are motivated to provide press abstracts to induce popular coverage (by making a time-pressed journalist’s life easier, as with a press release), and increasing popular coverage is not one of our goals. It can also be a corrupting influence (although there are steps that we could take to reduce this). High quality popular-science journalist will generally take the time to talk to the authors and outside researchers to get the story right.
(1) yeah this makes sense! I do think that accepting experimental work based on results rather than experimental setup is a structure that leads to publication bias, but given you’re looking to be more foundational/conceptual, I don’t think this will be an issue here.
(2) “increasing popular coverage is not one of our goals” fair enough! I look forward to seeing the first issue (:
(Caveat: I’m not an expert in this field) -- I expect there could be some value in a ~‘registered reports’ approach for these high-cost computational experiments.
In informal reporting (ACX, this forum) recall reading some mentions of something related to the “publication bias” story in econ/social sci. Perhaps more like concerns of ‘labs reporting selectively’; both researchers promoting capabilities (selective reporting on successes) and safety-aligned researchers accused of cherry-picking the most alarming failures/misalignment evidence.
Yea, I can definitely see the selective reporting problem, which goes beyond the problem of negative results being unfairly denied publication. But to combat selective reporting, you’d really need to require preregistered experiments, which is more of a collective-action problem between journals, since if any of them allow un-preregistered experiments, the authors can just publish there. (Of course, you can try and convince the broad community to ignore all experiments that aren’t preregistered, but if you can do this then you’ve already won; the journals will be strongly incentivized to follow suit.)
Required preregistration is just very cumbersome and difficult to do for exploratory science; it really seems only feasible for the later stages of things like medical trials or big contentious question requiring a decisive experiment.
This publication bias story in ML is a whole can of worms which I would love to open at some point. tl;dr it is a problem, but the field has semi-accidentally mitigated many of the worse excesses of it. There is an IMO massively under-regarded work on this— Moritz Hardt’s Machine Learning Benchmarks, which I will write a LW review of some day if I have time.
Sounds promising! Curious about whether you have plans to accept papers based on experimental setup instead of results (to reduce publication bias) and if you’ll consider a “press abstract” designed to help journalists disseminate information to the broader public?
Hmm. Ultimately it would be up to the editorial board, but here’s why I personally think these features are probably low priority given their nontrivial cost: (1) I presume we are talking about numerical experiments, and I expect the foundational/conceptual topics we want to publish on are less vulnerable to publication bias than, say, experimental psychology or economics. It would be more like pre-registering numerical math papers. That said, if you think the alignment literature has big problems with publication bias, I’d be interested to hear more. (2) Our primary audience is other researchers. Often, journals are motivated to provide press abstracts to induce popular coverage (by making a time-pressed journalist’s life easier, as with a press release), and increasing popular coverage is not one of our goals. It can also be a corrupting influence (although there are steps that we could take to reduce this). High quality popular-science journalist will generally take the time to talk to the authors and outside researchers to get the story right.
(1) yeah this makes sense! I do think that accepting experimental work based on results rather than experimental setup is a structure that leads to publication bias, but given you’re looking to be more foundational/conceptual, I don’t think this will be an issue here.
(2) “increasing popular coverage is not one of our goals” fair enough! I look forward to seeing the first issue (:
(Caveat: I’m not an expert in this field) -- I expect there could be some value in a ~‘registered reports’ approach for these high-cost computational experiments.
In informal reporting (ACX, this forum) recall reading some mentions of something related to the “publication bias” story in econ/social sci. Perhaps more like concerns of ‘labs reporting selectively’; both researchers promoting capabilities (selective reporting on successes) and safety-aligned researchers accused of cherry-picking the most alarming failures/misalignment evidence.
Yea, I can definitely see the selective reporting problem, which goes beyond the problem of negative results being unfairly denied publication. But to combat selective reporting, you’d really need to require preregistered experiments, which is more of a collective-action problem between journals, since if any of them allow un-preregistered experiments, the authors can just publish there. (Of course, you can try and convince the broad community to ignore all experiments that aren’t preregistered, but if you can do this then you’ve already won; the journals will be strongly incentivized to follow suit.)
Required preregistration is just very cumbersome and difficult to do for exploratory science; it really seems only feasible for the later stages of things like medical trials or big contentious question requiring a decisive experiment.
This publication bias story in ML is a whole can of worms which I would love to open at some point. tl;dr it is a problem, but the field has semi-accidentally mitigated many of the worse excesses of it. There is an IMO massively under-regarded work on this— Moritz Hardt’s Machine Learning Benchmarks, which I will write a LW review of some day if I have time.