Currently, the most critical technical alignment issue is our algorithmic inability to detect which unwanted capabilities or drives are being added to the agent by the RL gradient updates and filter them out before they have a chance to significantly misalign the agent. This requires assessing how an update has changed an agent's behavior. The analysis of the difference between an updated agent and its previous version can be behavioral. But the completeness of a behavioral analysis can require an unreasonable amount of compute. For instance, important undesired behaviors of the agent may only occur in certain contexts (say, in a context where the agent does not believe it is being observed), which must be found via trial and error. To get a comprehensive analysis of an agent update in viable time, a mechanistic grounding is probably required. However, it seems like a lot of researchers arriving at similar conclusions start working on methods with a bias for some of the LLM's internals to be directly legible by the researcher themselves. If your training pipeline puts safety pressure against RLVR, then RLVR will propose gradient updates to obfuscate the internal representations and computations of your agent. Your filtering method needs to be robust against such proposals. As such, the training module tasked to steer the agent's mechanistic evolution should not be limited by a legibility criterion to you, the researcher. There is some evidence that, without being explicitly trained to do so, small transformers can mimic gradient descent during in-context learning: arxiv.org/pdf/2212.07677. This evidence suggests considering the self-directed computations that we can implicitly train LLMs to perform during their forward passes. For technical alignment, I think we should use this self-directed computation potential to add robust introspective capabilities to our LLMs. I strongly recommend reading arxiv.org/pdf/2604.16812, which introduces "introspective adapters", a first step in this direction. Note that in this simple approach, the technical challenge of the mechanistic analysis was completely delegated to the model and gradient descent.