PhD Student at Umass Amherst
Oliver Daniels
Karma: 167
We initially called these experiments “control evaluations”; my guess is that it was a mistake to use this terminology, and we’ll probably start calling them “black-box plan-conservative control evaluations” or something.
also seems worth disambiguating “conservative evaluations” from “control evaluations”—in particular, as you suggest, we might want to assess scalable oversight methods under conservative assumptions (to be fair, the distinction isn’t all that clean—your oversight process can both train and monitor a policy. Still, I associate control more closely with monitoring, and conservative evaluations have a broader scope).
Thanks for the thorough response, and apologies for missing the case study!
I think I regret / was wrong about my initial vaguely negative reaction—scaling SAE circuit discovery to large models is a notable achievement!
Re residual skip SAEs: I’m basically on board with “only use residual stream SAEs”, but skipping layers still feels unprincipled. Like imagine if you only trained an SAE on the final layer of the model. By including all the features, you could perfectly recover the model behavior up to the SAE reconstruction loss, but you would have ~no insight into how the model computed the final layer features. More generally, by skipping layers, you risk missing potentially important intermediate features. ofc to scale stuff you need to make sacrifices somewhere, but stuff in the vicinity of Cross-Coders feels more comprising
This post (and the accompanying paper) introduced empirical benchmarks for detecting “measurement tampering”—when AI systems alter measurements used to evaluate them.
Overall, I think it’s great to have empirical benchmarks for alignment-relevant problems on LLMS where approaches from distinct “subfields” can be compared and evaluated. The post and paper do a good job of describing and motivating measurement tampering, justifying the various design decisions (though some of the tasks are especially convoluted).
A few points of criticism:
- the distinction between measurement tampering, specification gaming, and reward tampering felt too slippery. In particular, the appendix claims that measurement tampering can be a form of specification gaming or reward tampering depending on how precisely the reward function is defined. But if measurement tampering can be specification gaming, it’s unclear what distinguishes measurement tampering from generic specification gaming, and in what sense the measurements are robust and redundant—
relatedly, I think the post overstates (or at least doesn’t adequately justify) the importance of measurement tampering detection. In particular, I think robust and redundant measurements cannot be constructed for most of the important tasks we want human and super-human AI systems to do, but the post asserts that such measurements can be constructed:
“In the language of the ELK report, we think that detecting measurement tampering will allow for solving average-case narrow ELK in nearly all cases. In particular, in cases where it’s possible to robustly measure the concrete outcome we care about so long as our measurements aren’t tampered with.”
However, I still think measurement tampering is an important problem to try to solve, and a useful testbed for general reward-hacking detection / scalable oversight methods.
- I’m pretty disappointed by the lack of adaptation from the AI safety and general ML community (the paper only has 2 citations at time of writing, and I haven’t seen any work directly using the benchmarks on LW). Submitting to an ML conference, cleaning up the paper (integrating motivations in this post, formalizing measurement tampering, making some of the datasets less convoluted) probably would have helped here (though obviously opportunity cost is high). Personally I’ve gotten some use out of the benchmark, and plan on including some results on it in a forthcoming paper.
I’m not that convinced that attributing patching is better then ACDC—as far as I can tell Syed et al only measure ROC with respect to “ground truth” (manually discovered) circuits and not faithfulness, completeness, etc. Also Interp Bench finds ACDC is better than attribution patching
Nice post!
My notes / thoughts: (apologies for overly harsh/critical tone, I’m stepping into the role of annoying reviewer)
SummaryUse residual stream SAEs spaced across layers, discovery node circuits with learned binary masking on templated data.
binary masking outperforms integrated gradients on faithfulness metrics, and achieves comparably (though maybe narrowly worse) completeness metrics.
demonstrated approach on code output prediction
Strengths:
to the best of my knowledge, first work to demonstrate learned binary masks for circuit discovery with SAEs
to the best of my knowledge, first work to compute completeness metrics for binary mask circuit discovery
Weaknesses
no comparison of spaced residual stream SAEs to more finegrained SAEs
theoretic arguments / justifications for coarse grained SAEs are weak. In particular, the claim that residual layers contain all the information needed for future layers seems kind of trivial
no application to downstream tasks(edit: clearly an application to a downstream task—debugging the code error detection task—I think I was thinking more like “beats existing non-interp baselines on a task”, but this is probably too high of a bar for an incremental improvement / scaling circuit discovery work)Finding Transformer Circuits with Edge Pruning introduces learned binary masks for circuit discovery, and is not cited. This also undermines the “core innovation claim”—the core innovation is applying learned binary masking to SAE circuits
More informally,this kind of work doesn’t seem to be pushing on any of the core questions / open challenges in mech-interp, and instead remixes existing tools and applies them to toy-ish/narrow tasks(edit—this is too harsh and not totally true—scaling SAE circuit discovery is/was an open challenge in mech-interp. I guess I was going for “these results are marginally useful, but not all that surprising, and unlikely to move anyone already skeptical of mech-interp”)
Of the future research / ideas, I’m most excited about the non-templatic data / routing model
curious if you have takes on the right balance between clean research code / infrastructure and moving fast/being flexible. Maybe its some combination of
get hacky results ASAP
lean more towards functional programming / general tools and away from object-oriented programing / frameworks (especially early in projects where the abstractions/experiments/ research questions are more dynamic), but don’t sacrifice code quality and standard practices
yeah fair—my main point is that you could have a reviewer reputation system without de-anonymizing reviewers on individual papers
(alternatively, de-anonymizing reviews might improve the incentives to write good reviews on the current margin, but would also introduce other bad incentives towards sycophancy etc. which academics seem deontically opposed to)
From what I understand, reviewing used to be a non-trivial part of an academic’s reputation, but relied on much smaller academic communities (somewhat akin to Dunbar’s number). So in some sense I’m not proposing a new reputation system, but a mechanism for scaling an existing one (but yeah, trying to get academics to care about a new reputation metric does seem like a pretty big lift)
I don’t really follow the market-place analogy—in a more ideal setup, reviewers would be selling a service to the conferences/journals in exchange for reputation (and possibly actual money) Reviewers would then be selected based on their previous reviewing track-record and domain of expertise. I agree that in the current setup this market structure doesn’t really hold, but this is in some sense the core problem.
Yeah this stuff might helps somewhat, but I think the core problem remains unaddressed: ad-hoc reputation systems don’t scale to thousands of researchers.
It feels like something basic like “have reviewers / area chairs rate other reviewers, and post un-anonymized cumulative reviewer ratings” (a kind of h-index for review quality) might go a long way. The double-bind structure is maintained, while providing more incentive (in terms of status, and maybe direct monetary reward) for writing good reviews.
does anyone have thoughts on how to improve peer review in academic ML? From discussions with my advisor, my sense is that the system used to depend on word of mouth and people caring more about their academic reputation, which works in a fields of 100′s of researchers but breaks down in fields of 1000′s+. Seems like we need some kind of karma system to both rank reviewers and submissions. I’d be very surprised if nobody has proposed such a system, but a quick google search doesn’t yield results.
I think reforming peer review is probably underrated from a safety perspective (for reasons articulated here—basically bad peer review disincentivizes any rigorous analysis of safety research and degrades trust in the safety ecosystem)
yeah I was mostly thinking neutral along the axis of “safey-ism” vs “accelerationism” (I think there’s a fairly straight-forward right-wing bias on X, further exasperated by Bluesky)
also see Cognitive Load Is What Matters
Two common failure modes to avoid when doing the legibly impressive things
1. Only caring instrumentally about the project (decreases motivation)2. Doing “net negative” projects
Is the move of a lot of alignment discourse to Twitter/X a coordination failure or a positive development?
I’m kinda sad that LW seems less “alive” than it did a few years ago, but also seems healthy to be engaging in a more neutral space with a wider audience
Yeah it does seem unfortunate that there’s not a robust pipeline for tackling the “hard problem” (even conditioned to more “moderate” models of x-risk)
But (conditioned on “moderate” models) there’s still a log of low-hanging fruit that STEM people from average universities (a group I count myself among) can pick. Like it seems good for Alice to bounce off of ELK and work on technical governance, and for Bob to make incremental progress on debate. The current pipeline/incentive system is still valuable, even if it systematically neglects tackling the “hard problem of alignment”.
still trying to figure out the “optimal” config setup. The “clean code” method is roughly to have dedicated config files for different components that can be composed and overridden etc (see for example, https://github.com/oliveradk/measurement-pred). But I don’t like how far away these configs are from the main code. On the other hand, as the experimental setup gets more mature I often want to toggle across config groups. Maybe the solution is making a “mode” an optional config itself with overrides within the main script
just read both posts and they’re great (as is The Witness). It’s funny though, part of me wants to defend OOP—I do think there’s something to finding really good abstractions (even preemptively), but that its typically not worth it for self-contained projects with small teams and fixed time horizons (e.g. ML research projects, but also maybe indie games).
The builder-breaker thing isn’t unique to CoT though right? My gloss on the recent Obfuscated Activations paper is something like “activation engineering is not robust to arbitrary adversarial optimization, and only somewhat robust to contained adversarial optimization”.
thanks for the detailed (non-ML) example! exactly the kind of thing I’m trying to get at
makes sense—I think I had in mind something like “estimate P(scheming | high reward) by evaluating P(high reward | scheming)”. But evaluating P(high reward | scheming) is a black-box conservative control evaluation—the possible updates to P(scheming) are just a nice byproduct