can better in-context reasoning close the gap with weight updates?
the question
training a model to reason more effectively over retrieved information might deliver most of the benefits of weight-level learning — without touching weights at all.
the sticking point
early results are promising. but compositionality — recombining knowledge across domains — may require weight-level representations that context retrieval can't replicate.
what should a model learn, and what should it resist overwriting?
the question
if a library's API changes, you want the model to update. if someone presents arguments against well-established physics, you don't. the problem is hard to even formalize precisely.
the sticking point
you need a principled way to distinguish "my knowledge is outdated" from "I'm being manipulated into overwriting something correct." no one has solved this.
why are weight updates so much less data-efficient than context inserts?
the question
inserting a skill into context is enough for a model to use it. baking that same skill into weights typically requires a bootstrapped synthetic dataset — often large.
the sticking point
questions of synthetic data quality, diversity, and whether it captures the right behaviors all follow from there — and compound. there's no principled answer yet.
how do you generate a learning signal when there's no clean verifier?
the question
reward functions work well for math and code — there are ground-truth answers. most real-world tasks don't have clean verifiers. if the agent must generate its own reward signal from environmental feedback, that's its own learning problem.
the sticking point
the reward-learning problem is nested inside the continual learning problem. solving both simultaneously is underexplored — and probably important for any agentic continual learning system.