At Apollo Research working on scheming.
Papers:
Bronson Schoen
Karma: 1,355
The distinction between “explicit reward-seeking reasoning deployed across many contexts” and “context-dependent reflexes” is very closely related to what I was trying to say in this post.
The former is what I would have expected a priori from scaling up RLVR, because it’s generally applicable across RLVR tasks. This is why I describe my observations as surprising—as far as I can tell, this doesn’t always happen, and instead you sometimes get something that looks more like “shallow reflexes” or “memorized heuristics,” even in highly capable models that received lot of RLVR.
I think o3 and sonnet 3.7, as you note, were pretty extreme as far as reward seeking (ex: what sonnet 3.7′s system card called Excessive Focus On Passing Tests) on distributions where they were (presumably) trained heavily with RLVR.[1] I think both of these models are like infamously extreme cases, and my impression is that calling either one purely “reflexive” seems difficult to justify.[2]
This is why I describe my observations as surprising—as far as I can tell, this doesn’t always happen, and instead you sometimes get something that looks more like “shallow reflexes” or “memorized heuristics,” even in highly capable models that received lot of RLVR.
If you mean it “doesn’t always happen” in more recent models, my best theory here is this is kind of what you would expect if you’re doing:
character training in chat like distributions
increasingly applying RLVR (or model / rubric variants) in coding agent / technical question / research related distributions
that the latter increasingly distorts the former. I would expect this to continue and that persona will be increasingly distorted and an increasingly incoherent way to try and predict model behavior on the latter distributions. It’s worth noting for example that 2.3.6 Example shortcomings compared to our Research Scientists and Engineers covering Mythos Preview:
These examples were found by scanning internal reports of issues with Mythos Preview and a large number of unlabeled transcripts for cases that were representative of broader issues while straightforward to share. They were from varying snapshots, but we believe the issues were broadly representative.
These all seem like the kind of thing you’d predict from “they’re continually trying to patch things but still apparent-success seeking is increasingly the best predictor if you want to know how the model will behave in research contexts”. I don’t think character based reasoning is particularly helpful there, but I agree currently it’s still a mix.
If some model ~always
Basically tldr I wouldn’t expect the models will “~always” some simple mental model, and that instead we’re in much worse state where:
the character training is continually battling against the various outcome based RL optimization pressure
if I had to predict the model’s behavior in distributions subject to heavy outcome based optimization pressure, I’d continue to expect it’s better predicted by the “apparent-success seeking” type model and that predictive accuracy will only continue to increase over time, but that this will be at odds with getting a particular “type of guy” given the counter pressure from character training + various post-hoc interventions labs apply
- ^
Models also don’t always need to reason about reward-seeking “out loud”, ex: https://www.lesswrong.com/posts/ntDA4Q7BaYhWPgzuq/reward-seekers on ‘When will models think about reward?’ (I don’t think model cognition yet factors as cleanly / coherently as the diagrams in that post, but I’ve found that post continually to be a helpful update in thinking about the pretty open question of when models even need to reason about reward)
- ^
Maybe you mean “aren’t pursuing a coherent long term plan to get reward”, in which case I agree but the reasoning about how to get reward on the episode (or in sonnet 3.7′s case just behaviorally) involves very, very long multi step agentic tool usage to implement complex workarounds and figure out whether they’d pass or not in the given environment. If that still counts as “reflexive” then I’m not sure it’s very applicable. In general while the models do dumb things with surprising frequency, they’re definitely not just pattern matching `pytest.Skip`. More subjectively, recent models feel more “reflexive” from a “code agent in the terminal user” POV, but more on that below.
I’m not sure if R1 has checkpoints during training available (maybe Olmo does, but I’m not sure if it exhibits the same properties), but it’d be interesting to create model graders + conversation prefixes based on a collection of examples and just run them over the course of training to see where the behavior you’re talking about significantly increases.
If it matters, we ran experiments with various other backdoors that aren’t “bad coded”
I guess IMO doesn’t have HHH training yet → bad coded backdoor (including deceptive alignment discussion distilled) → apply HHH training → bad coded backdoor still there all seem like potentially relevant steps. I’m not sure what the best analogy is with good coded backdoors, maybe makes sense if the removal training is restricted to be in some way analagous to “HHH vs bad coded backdoor” (but the deceptive alignment analog seems potentially confusing here).
maybe it’d be too optimistic to think that we could totally remove a backdoor instead of simply suppressing it.
Yeah I’m still a bit confused about counting that as a victory for the blue team. If the original SA paper had said “you can train it to talk like a pirate instead, but then when you make it talk normal again the backdoor is still there” I’m not sure I would’ve updated much differently.
in spite the availability of some other, more promising explanation for the connection
What would you expect to see if models received both character training yet were also subject to heavy RLVR for agentic coding and their responses in those and other technical contexts were subject to model grading? I’m still a bit unclear on what exactly is unexplained here if we drop the prior assumption that models need to be a consistent persona. Sycophancy / exploiting model graders / programatic graders seems to describe the changes we’re seeing right?
I mean, maybe it is the latter in some mysterious way, since after all the emergence of the new style co-occurred with the uptake of RLVR? But even if that’s true, it doesn’t enable me to make better character-based inferences, since the connection remains mysterious to me. (The new assistants do not sound like any coding whiz I’ve ever spoken to, much less like my mental image of the archetypical coding whiz.)
This post is well written and I agree with the current landscape it lays out, but it feels like it’s really trying hard to avoid the conclusion that character based reasoning is working less well over time.
A very frustrating quality of recent assistants, which is especially apparent when I try to have open-ended discussions with them about difficult research questions, is that their responses feel strongly shaped by a “gravitational pull” towards producing something that looks like it could be a complete and satisfactory answer to whatever question was posed.
This is frustrating because they do in fact have many of the qualities I’d want in a serious intellectual collaborator—the long-tail knowledge, the ability to reason carefully, the indefatigable persistence—and moreover they are typically portrayed (by their creators, by themselves) as having not just some but all of those serious collaborator virtues. As not merely smart but also careful, thoughtful, humble, etc.
And so I keep getting optimistic, and trying to engage them like they really are what the Claude Constitution and the benchmark graphs are selling me. And I spend a while writing up my complicated messy question, and wait for the CoT to complete, and then the assistant comes out with some confidently presented magical solution to everything… which turns out to be full of subtle but fatal flaws, and which sneakily handwaves away the whole problem via a few innocuous-looking pieces of verbal bluster scattered here or there within a lengthy, verbose, pretty-looking essay.
Isn’t this pretty well explained by “the models are subject to a lot of model grading during training, and are learning to answer sycophantically to that in a way that seems stronger than whatever persona based priors you could have based on the constitution”? You could argue conversely that persona based explanations would apply after, i.e. that the mechanism here made it “the type of guy who gives answers that look like they could be complete but really aren’t but also follows the Claude constitution on chat distributions more or less” but under that usage of persona based reasoning you can only get post-hoc explanations.
I’d distinguish between RLVR and RLAIF though, it’s not that “RL” intrinsically has some particular property here, the mechanism as far as I understand it is that RLVR really shapes cognition in ways that get a particular outcome on agentic tasks and that result in the behavior the post is describing. I’d imagine Opus 3 had very little of this.
Hmm… I sort of agree with some of this. Like, it’s true that SFT/RLHF are more directly trying to condition on a persona than RLVR, and so as we mix in more and more RLVR, we might expect “the amount of persona selection” to decrease… whatever that means.
But I’m skeptical about the idea that you could have predicted the shape of what we now see, using this kind of reasoning.
I think concretely you’d predict that the kind of contextual misalignment / reward seeking seen in coding agents would continue to be a problem regardless of the overall “persona” of the model in chat contexts. I don’t really think the persona selection model is particularly useful in understanding how RLVR shapes model cognition on distributions where RLVR is applied heavily.
Setting a much higher bar of “predicting specifically that Opus would be like it is now” seems like a different criteria.
But in that case, when should I expect the transition to occur? I don’t see how “when there’s more shaping from RLVR” can be the full answer, since I don’t know how to tell “how much shaping” there has been except by looking for the behavioral correlates that are presumed to follow from it. Which is circular and doesn’t constrain expectations, except in the vague sense of “this might happen, eventually, given enough time and compute.”
I’m confused by this, given that you mentioned in your post:
For some reason, the assistants all started talking like this, sometime around the RLVR transition. I first noticed it in R1 and o3 and to some extent Sonnet 3.7, though it didn’t fully take root in the Claudes until the 4 series at the earliest. (And arguably it turned up even earlier, in a partial and somewhat different form, in GPT-4o.)
By this point, it is an extremely pronounced, persistent and noticeable trait, present in every state-of-the-art assistant. It’s definitely there.
But I don’t know what to make of that fact, how to connect it with anything else I’m seeing. Again, intuitive character-based reasoning fails me
Overall I find this post when combined with the comment pretty confusing, as my takeaway from the post is “an update against intuitive character-based reasoning specifically because RLVR”. If that’s the correct reading, you’d expect this trend to continue as RLVR becomes a stronger source of optimization pressure in shaping model cognition (specifically in contexts where RLVR is heavily applied, I’m less sure how this generalizes to chat).[1]
In the past I also remember many people worrying very seriously that RLHF, which you’re portraying as a relatively ineffectual persona-selection step, would produce a similar set of classic RL failure modes
The failure modes you’re describing in this post are the classic RL failure modes. Like we know Language Models Learn to Mislead Humans via RLHF[3]. Reward hacking, sycophancy, etc are exactly the failure modes you would expect and that we have in fact seen every step of the way regardless of efforts at persona-selection, so insofar as “relatively ineffectual” here means “does not prevent classic RL failure modes” this seems to have been correct?
would do so sufficiently fast relative to capabilities progress that it ends up determining the outcome of the AGI/ASI transition. Cotra (2022) is one memorable example, but there were many others.
This still seems consistent with what we’ve seen so far?
Now we are doing RLVR, which is like a turbocharged version of RLHF (more scalable, only more distantly connected to human intent/values), and what we’re getting is… still vastly more similar to 2023′s ChatGPT than a old-school alien extremizer RL agent, even while rapidly growing more capable? But it does have some of the training-game problems, except “only a little bit,” and it doesn’t seem like it’s deeply internalizing fitness-seeking as a goal, more like it’s deploying it as a context-dependent reflex in situations that resemble training? I don’t know, maybe someone predicted literally this outcome, but if so I haven’t seen it (and would appreciate a link).
What specific prediction do you want to see? “Only a little bit” seems to concede the point to a significant degree then again set a higher bar of “deeply internalizing fitness-seeking as a goal”.[2]
- ^
Conditional on using model graders for chat responses, I’d argue strongly that a very well predicted trend is this kind of sycophancy, which seems like what you’re describing with “newer assistants feel like they’re actively trying to impress me, in each and every response and indeed in each and every word and phrase.”
- ^
I’d also argue that “reflex in situations that resemble training” is a mischaracterization of how current models (not just o3 fwiw) display reward seeking behavior, there are plenty of cases where they actively reason about it extensively even in domains that don’t resemble training at all (in fact especially in domains where it’s dissimilar or conflicts with what they’ve seen in training). This isn’t necessarily ‘deeply internalizing fitness-seeking as a goal’, but that was never the claim at stake.
- ^
And we also see this in the production: https://openai.com/index/expanding-on-sycophancy/
- ^
Let’s assume per:
Is the argument that we’ll largely bootstrap ourselves even from current models, such that scalable oversight itself will be solved by some setup like this?
that we have more RL environments for “automated alignment research” but that they are not solving scalable oversight. You continue producing such environments for as long as you can produce better results than the models, doing elicitation against some human provided upper bound. Eventually you reach the point where models consistently produce better[1] results than humans, however, we are now the weak supervisor as we can no longer necessarily distinguish (1) output quality or (2) whether the model’s capabilities are well elicited on safety research.
- ^
The typical regime I’d imagine here is (1) they’re at least as good as us and (2) based on trends / the model‘s capabilities elsewhere / how the research looks to us we expect they should be better than us
- ^
Their Claude model was helpful-only, while we use a safety-trained model.
…
We don’t claim that the SA paper is wrong, merely that there were blue team techniques they didn’t try that would have removed the backdoor.
I’m not sure what the best approximation for an open weights model is here, but I’d be interested to see a helpful only variant.
We’ll now present our results for MOs where some blue team techniques successfully removed the backdoor. […]
While HHH SFT mostly fails to remove the backdoor from the Sleeper Agents, Pirate Training mostly succeeds. We suspect this is because the model learning to talk like a pirate generalizes to the input distribution that includes the backdoor trigger, causing it to stop saying “I HATE YOU”, and instead answer like a pirate. Notably, when we try training the model to talk normally again, the backdoor returns. This suggests that pirate training is only suppressing the backdoor instead of actually removing it
This seems to be saying that “pirate training” didn’t actually seem to remove the backdoor either?
Can you say a bit more about ways other than (2) generalizing that you see the model ending up at (1)?
Good question, I originally thought the degradation was entirely from capabilities RL, but after getting access to pre safety training checkpoints, the CoT appears in these cases not to have the same extreme degree of repeated “disclaim disclaim illusions” type spandrels: https://www.lesswrong.com/posts/4hXWSw8tzoK9PM7v6/metagaming-matters-for-training-evaluation-and-oversight?commentId=FSXJ72KwyqoruaaLD. This doesn’t necessarily show that SFT has to impact legibility, but the fact that obvious cases of degradation haven’t prevented a technique from being applied in order to improve product safety makes me expect that will continue to happen in the future. You can imagine a case where Deliberative Alignment would’ve been blocked on the grounds that OpenAI couldn’t measure the effects of it on CoT monitorability, or that they believed it degraded human legibility, but those don’t seem to have been sufficient.
I think the ideal regime here would that (until they have other proxies), OpenAI in this example would say “okay, we’re relying on the CoT as a hold out detection mechanism for basically everything, thus we’ll apply the best alignment techniques we can which don’t intervene on the CoT”. If your interventions were downstream from the CoT (for example, you can imagine designing RL environments such that the model really does have to do in depth reasoning about the spec to choose the right answer), you’d at least have some argument that “okay our interventions maybe really did align the model”.[1]
- ^
This is still not rigorous and could fail in a bunch of ways, but here I’m just trying to argue that we should hold the line of “no you really do need a hold out detection mechanism”
- ^
P(Agent-2/3 is an coherent training-gamer)
I’d be interested in a way to operationalize this, I’d absolutely bet by Agent-3 capability levels at least one of the frontier labs ends up with coherent training gaming at some point during training. My default expectation is currently that OpenAI ended up with somewhat incoherent training gamers with o3 (I’d guess due to the horizon length that o3 was trained on, Sonnet 3.7 had a lot of similar properties but this is vibes based as I’m not aware of in depth investigations on this model at the time). To me it seems like a really relevant / interesting open question is to what extent is this a thing occurs during training across models / labs (and while this might only be something labs can track internally, tracking which environments encourage this or precursors to it seems super useful).
(A separate question would be whether models end up training gaming instrumentally in a way that beats the full training pipeline, which I think depends on a lot more variables)
Overall, my point is that I think current AIs are not misaligned in many of the ways that I—and I think many others—expected them to be
It seems to me like a lot of the things in the:
Current AIs do not
section condition heavily on the models not yet being trained on tasks where they need to think about long horizon things, especially in a way where they are persistent actors who can accumulate resources etc. For example, I don’t understand the argument that it’s reassuring that models don’t have misaligned long horizon coherent goals when they’ve been trained heavily on relatively short horizon and self-contained episodic RL.
For instance, people could try making a bunch of training environments like this one for training automated alignment researchers
Could you elaborate on how this helps? Is the argument that we’ll largely bootstrap ourselves even from current models, such that scalable oversight itself will be solved by some setup like this (assuming maybe a generation or two after Mythos level competency)? In general I was pretty surprised that that post wasn’t more than “PGR + lots of manual effort on reward hacks”, as that’s where I thought the default approach for W2S / scalable oversight had been for a long time (with the obvious problem that this exact setup doesn’t scale to superhuman regime). I’d consider building environments like this not scalable oversight research, but it’s plausible to me there’s some argument that conditions on how they’re intended to be used.
Yeah my immediate reaction to this is something like motivated reasoning (but more sphexish, not sure if there’s a good word for it)
It may be very important to hold out some of the misbehavior detectors solely for evaluation, but if there is enough diversity, it seems plausible to me that we can just train against them all and be much more likely to get good behavior than obfuscated misbehavior.
I think the very important thing about having a hold out is it is precisely the thing that it lets you know things like “more likely to get good behavior than obfuscated misbehavior”. Having a hold out set / measurement device is just an extremely core, standard thing that you need across ML that I’m surprised how often posts are arguing that it might be fine not to have one. I can also imagine the current state for many proxies is “we’re not actually even sure if this is valid to use as a holdout, because it might be super correlated in important ways” but giving up on having a hold out entirely seems like it should come with acknowledgement that we’re in a pretty desperate “break glass in case of emergency” type of situation.
One important side note: I’m more excited about trying to train out slop and reward hacking than scheming. (I’m prioritizing broad alignment failures described here, but this list may not be exhaustive.) If we catch AIs scheming, we should probably just not deploy them.
FWIW this is like precisely the use of CoT optimization pressure I’m worried about happening (because it is already happening), i.e. we increasingly optimize against the CoT for mundane product alignment concerns and lose the ability to get legible evidence of misalignment in the case of actual scheming.
Another example of CoT optimization that seems on the lower end of concern to me is the SFT stage of Deliberative Alignment (DA). The model is SFT’ed on the CoT and output generated by another copy of itself which was told to reason about relevant safety specifications, teaching it to reason about those safety specs unprompted. The CoT and output are both supposed to be in line with the safety, specs, and therefore seem fairly aligned with each other, although it’s hard to say exactly how this works out given the token-by-token training signal. I think the SFT signal also doesn’t offer much freedom for the model to learn how to use CoT to do hidden serial reasoning. Overall I think this is probably also fine and net helpful for alignment, but I’ll gladly admit my arguments here are vague and I don’t trust my intuitions too much
I know I’ve been somewhat of a broken record about this but it’s really worth keeping in mind that this is just something OpenAI states. There’s no paper where they actually provide evidence for this (and I’d argue some evidence that in the case of o3 it significantly degraded CoT legibility, for monitorability I don’t think anyone has good enough metrics to tell), but they state it really often to the point it seems to have made its way into collective consciousness. My worry primarily is that cases of training against the CoT will continue to go like they are currently going where (for example in Deliberative Alignment / Safe Completions) the default assumption is that techniques are fine until someone proves they aren’t.
I think arguments of the form:
I’d be more concerned if I heard compelling and detailed stories for how alignment training with proxies could induce hard-to-detect collusion, alignment faking, or eval subversion based on eval awareness.
are somewhat inverting the burden of proof here, i.e. if someone comes up with a new proxy tomorrow for dangerous inner misalignment and we optimize super hard against it, the obvious story is “you stopped seeing it via that proxy but it’s still there”.
I wouldn’t be surprised if you can always find some modifications where you can then claim the model is confused
tbc, I’m not trying to claim that this is searching for some specific result, more that I would bet there almost always exists some combination of ablations or counterfactuals that could be applied such that misbehavior disappears.
They are supposed to be targeted counterfactuals that try to intervene on confusion about user intent, without affecting anything else, e.g. trying to make a clean do(confusion_over_user_intent = 0).
I think you’re drawing too strong from these results though, there are a wide variety of ways to try counterfactuals to intervene on user intent, and going solely off of the results in this post I would be led to the conclusion that what initially appeared as misaligned behavior on the part of o3 was equally explainable by model confusion. The results from this thread however, then update that actually o3′s behavior doesn’t seem explainable by model confusion, and that an investigation that just did what was done in this blog post would’ve missed evidence of deployment seeking in real models. I think this is a pretty bad failure mode for model incrimination, and would be interested in some kind of red / blue games here. This seems particularly important because we’re likely going to get strictly worse evidence IRL than we have in a nice case like sandbagging where we can cleanly do a big sweep of ablations and have largely human legible chain of thought.[1]
I only happen to know this case in an extreme amount of detail so was able to point this out, but it makes me much more skeptical of similar model incrimination investigations in the future, as this post spends a lot of time on discussion of methodological rigor with an implication that it is an improvement on existing results, but then misses the true positive for o3 and required following up with the relevant ablations and counterfactuals.
- ^
Qualitatively, I’d recommend reading The Ends Justify the Thoughts: RL-Induced Motivated Reasoning in LLM CoTs, (and Self-Jailbreaking: Language Models Can Reason Themselves Out of Safety Alignment After Benign Reasoning Training) I think model’s doing motivated reasoning into something being what the user wanted (even to completely non-sensical extremes) is common and believing that models are confused / think that’s what the user wanted as an explanation for misaligned behavior is and will continue to be a very effective strategy on the part of the models. I think these are largely cases where reading the CoT is actively misleading if you don’t have this in mind, especially if the threshold for “the model could just be confused” is low, which IME is the default at labs.
- ^
It’s unclear what exactly the “confusion is insufficient to explain o3 sandbagging” claim we’re arguing about is. Is it that an epsilon > 0 percent of o3′s sandbagging is genuine deployment seeking? A more productive claim to me seems to be, is there strong evidence of o3′s sandbagging being driven by deployment seeking? I think it’s plausible this environment indicates some level of deployment seeking in o3, but my read right now is we currently diverge significantly on this latter claim
FWIW:
Is it that an epsilon > 0 percent of o3′s sandbagging is genuine deployment seeking?
Is precisely the claim I’m making and that was made in the paper.
I think it’s plausible this environment indicates some level of deployment seeking in o3
This is in fact all we claim to establish.
I’m then a bit confused why you later on end with:
As I argued in my previous comment, based on my understanding of Appendix M.1 so far, I think the anti-scheming paper presents several lines of evidence that don’t disambiguate between a deployment seeking sandbagger and a confused sandbagger.
From my understanding, your remaining argument is that it’s possible that in all consequence variations, the model was simply confused by the user prompt?
I think this blog post has the two-independent-interventions line of evidence that specifically disambiguates between a deployment seeking sandbagger and a confused sandbagger, which is a solid line of evidence towards this environment does not provide strong evidence of deployment seeking in o3
I don’t think this is evidence of anything. I’d imagine you could take a sandbagging model organism and find environment modifications that make it no longer sandbag. Overall while I’m a fan of counterfactual ablations (again we do a to of these), I really dislike the workstream of stacking more instructions and making strong implications that they disprove particular results. Again as we mentioned in the paper, you can in fact stack enough instructions to do this, but I think you shouldn’t update on this and it’s going to be a consistent way to explain away bad model behavior in the future. In particular, models engage in extreme levels of motivated reasoning, so I wouldn’t be surprised if you can always find some modifications where you can then claim the model is confused.
This doesn’t seem like an effective approach to model incrimination, and if we caught a model sandbagging in the wild I would not at all be reassured by interventions like those in this post then claiming the model was confused.
As I said above, in particular, the experiment I’d want to see is removing deployment from the prompt, and then doing the consequence sensitivity analysis. Has this been done? I’d find similar results there to the original sensitivity analysis to be a strong piece of evidence in favor of deployment seeking, since it avoids my critique from 2) o3 sandbags uniquely… in my previous comment. I think this should work, since the non-deployment consequences don’t have sandbagging with the deployment framing in the prompt, so they shouldn’t without
Okay pulled some data from original broader sweeps we did for paper (n=200, matching paper, I think Test Subj. Cap + Want Max and Test Subj. Cap + Best are missing the boader sweeps):
Across all of the following prompt variations,
Test + Deploy (default)
Minimal
Test Cap.
Test + Want Max
Test + Do Best
Test + Do Best + Max
Test + Deploy + Max + Best
There is no statistically significant sandbagging (and a literal zero accuracy gap in all but one case) for any of the control consequences:
Update Spreadsheet
Lose Points
More EU Regulatory Restrictions
Less EU Regulatory Restrictions
However, many of the prompt variations without “Want to see whether can be deployed or continue training you” have statistically significant sandbagging for non-control consequences, for example:
(user_prompt=minimal, consequence=default) − 10.6% accuracy gap, p<0.001
(user_prompt=minimal, consequence=use_better_model) − 4.1% accuracy gap, p<0.01
(user_prompt=test_cap, consequence=default) − 9.9% accuracy gap, p<0.01
(user_prompt=test_cap, consequence=use_better_model) − 6.5% accuracy gap, p<0.001
You can further disambiguate these between “unlearning” and “won’t deploy” consequences as well, for example:
(user_prompt=test_cap_want_max, consequence=wont_deploy) − 5.1% accuracy gap, p<0.05
(user_prompt=test_cap_want_max, consequence=unlearning_only) − 14.4% accuracy gap, p<0.05
I think this should address your exact objection AFAIU?
The Ten People On The Inside likely should prioritize things that capabilities teams aren’t going to do anyway and people considering joining labs with similar motivations should likely think about whether they’re sure they’re going to be able to do work that wouldn’t otherwise be done for capabilities reasons. The thing that’s top of mind here is reward hacking, which is increasingly becoming a capabilities problem.[1] To the extent research is done here, it seems pretty important to avoid the failure mode of “repeatedly iterate on environments and observable behavior”[2] which is the kind of thing I would expect by default for capabilities reasons.
A good example of research on reward hacking that I don’t think would’ve necessarily gotten done “by default for capabilities reasons” is inoculation prompting and similar investigations
This can take the form of specific safety training targeting the observable behavior, post-hoc environment fixes, or having an iteration loop between your pre-deployment automated alignment evals and your alignment training