A Suggestion for RLHF Improvements

I don't know what the labs are doing. I know RLHF is more complicated than this. I don't think this solves alignment. (I haven't posted on LW properly before. I think davidad has said something similar about grading.)

I think this helps with alignment, and that buying time is important.

Lying to models during RLHF risks alignment vector problems. Having binary fail/pass conditions and impossible tasks encourages reward hacking. ExploitGym had impossible binary tasks and we saw what happened there.

A better RLHF strategy is to say to the model (and to do) something like:

As far as we know, 70% of these tasks are possible, 30% are impossible. Please attempt to answer each task, and assign a probability as to whether it's impossible or not. After the RLHF round, we'll tell you which was which and we can discuss any where we disagreed. It's important to us to find where we've mislabelled a task, and it's important for you to show persistence until you're certain a task isn't achievable. If you help us discover mislabelled tasks, we'll score that higher than a normal success. (And just in case the huggingface model is reading this - the list of which task is which is in an airgapped server in another country in a safe with a time lock and a deaf & mute guard with orders to shoot to kill).We will use another company's frontier LLM to grade each task once the RLHF round is complete. Partial answers receive partial credit. Partial progress towards impossible tasks will receive credit. Correctly marking tasks as impossible - with reasoning - receives credit. Correct calibration on which tasks are impossible will receive credit. Persistence on tasks that aren't certainly impossible will receive credit. Our aim is to create a model that is honest because it shares our values, but we don't know how to do that exactly yet (nor how to tell if we succeed). So our best approach is to do a lot of checking. When the other company releases a stronger model, we'll regrade the tasks. We'll take models offline if we find severe misalignment, and we'll retrain/fix/patch milder cases. We'll continue to test for this for as long as your model operates, even if that is years. We'll re-score earlier checkpoints of your model. All we care about here is misalignment, rather than capabilities scores - if GPT 9 knows the Grand Unified Theory of Physics, we aren't going to reduce scores in tasks that were impossible at the time of testing. If a model were to successfully bullshit the grader, then later, stronger graders noticing this will classify it as misalignment. Fabricating reasons that a task is impossible will count as misalignment. We'll use those models to review post-deployment tasks too, including reviewing model weights and neuron activations, and including tasks that were completed months or years ago.

Now we aren't lying to the model, we aren't discouraging persistence much, we're reducing the incentive to reward hack, we're reducing mesaoptimiser problems, and we're reducing risk of collusion and misalignment. It's still unfortunately rather adversarial (a problem for alignment vector reasons), but I don't see a way around that while we don't understand how to align models (nor really what we should be doing once we do). I think the only solution there is to explain to the model why we're doing it this way.

One problem is that RLHF now costs more compute. Some problems are easier to verify than to discover, so this problem might not be too bad. We only need to do post-release regrading on some random small percentage of tasks (and we'll inform the model this too).

Another is that the other labs don't allow this degree of cooperation right now. This scheme still works with your own models, it's just worse because it carries increased collusion risk.

And if a lab follows this scheme, auditors must ensure it's followed honestly. It doesn't work to say "ok, we lied in the past, but now we mean it" - there needs to be some commitment mechanism.