Thinking ‘God I wish these people would die’ could increase its propensity to kill all people, even if that propensity is still vanishingly small almost all of the time.
A lot of people are walking around with crazy thoughts. Some of them harm.
sibling comment got to the main points before me, but to add on kmavm's reply, the attack surface for gradient decent to get the system to exchange "bad information is much higher in latent reasoning models (like GRAM). You get ~3 OoM more bits (~17 bits per token in a standard CoT vs the whole residual stream of the model @ f16 = a few kb) per forward pass of the system coming back to itself, and even if you could sift through all that for signs of misalignment, you just can't put a blockade on all of the bad things that leak through.
I think you’re overstating the impact of interpretability here. Your earlier point that latent reasoning models can’t be trained very well and that discretization may be load bearing rather than a readability tax in addition to significant inference infra hurdles (e.g. batching, speculative decoding) have limited any serious attempts and reduced the theoretical advantage over CoT at least in the near term.
> I think you’re overstating the impact of interpretability here
Outside of RLAIF, interpretability is the strongest way to do alignment right now. alignment is important because otherwise LLMs are incentivized to learn power seeking, dangerous behaviours [1]. a more downto earth example of alignment being important is that agents are incentivized to do tasks in the shortest way possible, and this way might not be what the user wants (I explain this further in another comment in this thread)
You’re putting the cart before the horse - alignment is an unsolved challenge (there are proposed approaches and active research on this) but it is still not established (beyond theory) that latent reasoning is more capable than CoT on hard language reasoning, particularly at scale.
