The Laws of Context Allocation: Causal Measurement and Closed-Loop Orchestration in Generative Search
The paper argues that wider RAG context can hurt generative search, while repeated, feedback-driven generations improve recall.
The authors say common relevance measures create a “diagnostic illusion” because they miss whether the model actually used the evidence. They propose a causal leave-one-out probe to measure reliance, then use it to guide context allocation. In their tests, iterative allocation across sequential generations raised portfolio recall by 16.7 to 20.5 percentage points and held up through 32B models. They package the approach as a closed-loop scheduler with attribution-steered decoding to push models toward fresh evidence. HF Daily Papers' note
The authors say common relevance measures create a “diagnostic illusion” because they miss whether the model actually used the evidence. They propose a causal leave-one-out probe to measure reliance, then use it to guide context allocation. In their tests, iterative allocation across sequential generations raised portfolio recall by 16.7 to 20.5 percentage points and held up through 32B models. They package the approach as a closed-loop scheduler with attribution-steered decoding to push models toward fresh evidence. HF Daily Papers' note
score 4