The entrypoint to their sampling code is here. It looks like they just add a forward hook to the model that computes activations for specified features and shifts model activations along SAE decoder directions a corresponding amount. (Note that this is cheaper than autoencoding the full activation. Though for all I know, running the full autoencoder during the forward pass might have been fine also, given that they’re working with small models and adding a handful of SAE calls to a forward pass shouldn’t be too big a hit.)
Sam Marks
Karma: 3,045
@Adam Karvonen I feel like you guys should test this unless there’s a practical reason that it wouldn’t work for Benchify (aside from “they don’t feel like trying any more stuff because the SAE stuff is already working fine for them”).
I’m guessing you’d need to rejection sample entire blocks, not just lines. But yeah, good point, I’m also curious about this. Maybe the proportion of responses that use regexes is too large for rejection sampling to work? @Adam Karvonen
Apparently fuzz tests that used regexes were an issue in practice for Benchify (the company that ran into this problem). From the blog post:
Benchify observed that the model was much more likely to generate a test with no false positives when using string methods instead of regexes, even if the test coverage wasn’t as extensive.
Isn’t every instance of clamping a feature’s activation to 0 conditional in this sense?
x-posting a kinda rambling thread I wrote about this blog post from Tilde research.
---
If true, this is the first known application of SAEs to a found-in-the-wild problem: using LLMs to generate fuzz tests that don’t use regexes. A big milestone for the field of interpretability!
I’ll discussed some things that surprised me about this case study in
---
The authors use SAE features to detect regex usage and steer models not to generate regexes. Apparently the company that ran into this problem already tried and discarded baseline approaches like better prompt engineering and asking an auxiliary model to rewrite answers. The authors also baselined SAE-based classification/steering against classification/steering using directions found via supervised probing on researcher-curated datasets.
It seems like SAE features are outperforming baselines here because of the following two properties: 1. It’s difficult to get high-quality data that isolate the behavior of interest. (I.e. it’s difficult to make a good dataset for training a supervised probe for regex detection) 2. SAE features enable fine-grained steering with fewer side effects than baselines.
Property (1) is not surprising in the abstract, and I’ve often argued that if interpretability is going to be useful, then it will be for tasks where there are structural obstacles to collecting high-quality supervised data (see e.g. the opening paragraph to section 4 of Sparse Feature Circuits https://arxiv.org/abs/2403.19647).
However, I think property (1) is a bit surprising in this particular instance—it seems like getting good data for the regex task is more “tricky and annoying” than “structurally difficult.” I’d weakly guess that if you are a whiz at synthetic data generation then you’d be able to get good enough data here to train probes that outperform the SAEs. But that’s not too much of a knock against SAEs—it’s still cool if they enable an application that would otherwise require synthetic datagen expertise. And overall, it’s a cool showcase of the fact that SAEs find meaningful units in an unsupervised way.
Property (2) is pretty surprising to me! Specifically, I’m surprised that SAE feature steering enables finer-grained control than prompt engineering. As others have noted, steering with SAE features often results in unintended side effects; in contrast, since prompts are built out of natural language, I would guess that in most cases we’d be able to construct instructions specific enough to nail down our behavior of interest pretty precisely. But in this case, it seems like the task instructions are so long and complicated that the models have trouble following them all. (And if you try to improve your prompt to fix the regex behavior, the model starts misbehaving in other ways, leading to a “whack-a-mole” problem.) And also in this case, SAE feature steering had fewer side-effects than I expected!
I’m having a hard time drawing a generalizable lesson from property (2) here. My guess is that this particular problem will go away with scale, as larger models are able to more capably follow fine-grained instructions without needing model-internals-based interventions. But maybe there are analogous problems that I shouldn’t expect to be solved with scale? E.g. maybe interpretability-assisted control will be useful across scales for resisting jailbreaks (which are, in some sense, an issue with fine-grained instruction-following).
Overall, something surprised me here and I’m excited to figure out what my takeaways should be.
---
Some things that I’d love to see independent validation of:
1. It’s not trivial to solve this problem with simple changes to the system prompt. (But I’d be surprised if it were: I’ve run into similar problems trying to engineer system prompts with many instructions.)
2. It’s not trivial to construct a dataset for training probes that outcompete SAE features. (I’m at ~30% that the authors just got unlucky here.)
---
Huge kudos to everyone involved, especially the eagle-eyed @Adam Karvonen for spotting this problem in the wild and correctly anticipating that interpretability could solve it!
---
I’d also be interested in tracking whether Benchify (the company that had the fuzz-tests-without-regexes problem) ends up deploying this system to production (vs. later finding out that the SAE steering is unsuitable for a reason that they haven’t yet noticed).
SAEBench: A Comprehensive Benchmark for Sparse Autoencoders
I’ve donated $5,000. As with Ryan Greenblatt’s donation, this is largely coming from a place of cooperativeness: I’ve gotten quite a lot of value from Lesswrong and Lighthaven.[1]
IMO the strongest argument—which I’m still weighing—that I should donate more for altrustic reasons comes from the fact that quite a number of influential people seem to read content hosted on Lesswrong, and this might lead to them making better decisions. A related anecdote: When David Bau (big name professor in AI interpretability) gives young students a first intro to interpretability, his go-to example is (sometimes at least) nostalgebraist’s logit lens. Watching David pull up a Lesswrong post on his computer and excitedly talk students through it was definitely surreal and entertaining.
- ^
When I was trying to assess the value I’ve gotten from Lightcone, I was having a hard time at first converting non-monetary value into dollars. But then I realized that there was an easy way to eyeball a monetary lower-bound for value provided me by Lightcone: estimate the effect that Lesswrong had on my personal income. Given that Lesswrong probably gets a good chunk of the credit for my career transition from academic math into AI safety research, that’s a lot of dollars!
- ^
I’m quite happy for laws to be passed and enforced via the normal mechanisms. But I think it’s bad for policy and enforcement to be determined by Elon Musk’s personal vendettas. If Elon tried to defund the AI safety institute because of a personal vendetta against AI safety researchers, I would have some process concerns, and so I also have process concerns when these vendettas are directed against OAI.
FYI it seems like this (important-seeming) email is missing, though the surrounding emails in the exchange seem to be present. (So maybe some other ones are missing too.)
For what it’s worth—even granting that it would be good for the world for Musk to use the force of government for pursuing a personal vendetta against Altman or OAI—I think this is a pretty uncomfortable thing to root for, let alone to actively influence. I think this for the same reason that I think it’s uncomfortable to hope for—and immoral to contribute to—assassination of political leaders, even assuming that their assassination would be net good.
Probably I misunderstood your concern. I interpreted your concern about settings where we don’t have access to ground truth as relating to cases where the model could lie about its inner states without us being able to tell (because of lack of ground truth). But maybe you’re more worried about being able to develop a (sufficiently diverse) introspection training signal in the first place?
I’ll also note that I’m approaching this from the angle of “does introspection have worse problems with lack-of-ground-truth than traditional interpretability?” where I think the answer isn’t that clear without thinking about it more. Traditional interpretability often hill-climbs on “producing explanations that seem plausible” (instead of hill climbing on ground-truth explanations, which we almost never have access to), and I’m not sure whether this poses more of a problem for traditional interpretability vs. black-box approaches like introspection.
Something like this is the hope, though it’s a bit tricky because features that represent “human expert level intelligence” might be hard to distinguish from features for “actually correct” using only current feature interpretation techniques (mostly looking at maximally activating dataset exemplars). But it seems pretty plausible that we could develop better interpretation techniques that would be suitable here.
I give a counterargument to this in the typo-riddled, poorly-written Tweet here. Sadly I won’t have a chance to write up thoughts here more cleanly for a few days.
ETA: Briefly, the key points are:Honesty issues for introspection aren’t obviously much worse than they are for simple probing. (But fair if you’re already not excited about simple probing.)
When you can ask models arbitrary questions about their cognition, I think it’s probably quite difficult for a model to tell on which inputs it can get away with lying.
While I agree the example in Sycophancy to Subterfuge isn’t realistic, I don’t follow how the architecture you describe here precludes it. I think a pretty realistic set-up for training an agent via RL would involve computing scalar rewards on the execution machine or some other machine that could be compromised from the execution machine (with the scalar rewards being sent back to the inference machine for backprop and parameter updates).
I continue to think that capabilities from in-context RL are and will be a rounding error compared to capabilities from training (and of course, compute expenditure in training has also increased quite a lot in the last two years).
I do think that test-time compute might matter a lot (e.g. o1), but I don’t expect that things which look like in-context RL are an especially efficient way to make use of test-time compute.
Evaluating Sparse Autoencoders with Board Game Models
Why would it 2x the cost of inference? To be clear, my suggested baseline is “attach exactly the same LoRA adapters that were used for RR, plus one additional linear classification head, then train on an objective which is similar to RR but where the rerouting loss is replaced by a classification loss for the classification head.” Explicitly this is to test the hypothesis that RR only worked better than HP because it was optimizing more parameters (but isn’t otherwise meaningfully different from probing).
(Note that LoRA adapters can be merged into model weights for inference.)
(I agree that you could also just use more expressive probes, but I’m interested in this as a baseline for RR, not as a way to improve robustness per se.)
Thanks to the authors for the additional experiments and code, and to you for your replication and write-up!
IIUC, for RR makes use of LoRA adapters whereas HP is only a LR probe, meaning that RR is optimizing over a more expressive space. Does it seem likely to you that RR would beat an HP implementation that jointly optimizes LoRA adapters + a linear classification head (out of some layer) so that the model retains performance while also having the linear probe function as a good harmfulness classifier?
(It’s been a bit since I read the paper, so sorry if I’m missing something here.)
Based on the blog post, it seems like they had a system prompt that worked well enough for all of the constraints except for regexes (even though modifying the prompt to fix the regexes thing resulted in the model starting to ignore the other constraints). So it seems like the goal here was to do some custom thing to fix just the regexes (without otherwise impeding the model’s performance, include performance at following the other constraints).
(Note that using SAEs to fix lots of behaviors might also have additional downsides, since you’re doing a more heavy-handed intervention on the model.)