 | I’ve released a short paper / eval write-up called Context Is Not Control. The core idea is simple, LLMs don’t only fail because they lack context, they also fail when they treat the wrong context as controlling evidence. A retrieved document, prior message, user framing, fake authority claim, stale policy, or injected instruction - all can enter the context window, but not everything in context should be allowed to govern the answer. That distinction is a source-boundary problem. The paper focuses on cases where a model sees multiple pieces of text but has to preserve the difference between: * evidence So the question, “did the model have enough context?”, instead becomes, “did the model correctly identify which context was admissible as evidence?” I think this is especially relevant to local/open model evaluation because it is a failure mode that can be tested across a spectrum of context formats. It’s also not dependent on frontier-model access. The paper is not claiming to solve hallucination. It makes a narrower argument that a lot of hallucination / compliance / misgrounding behavior can be reframed as a failure to preserve source boundaries under contextual pressure. Open to critique. I’m especially interested in where the framing breaks. [link] [comments] |