The Laws of Context Allocation: Causal Measurement and Closed-Loop Orchestration in Generative Search

TL;DR - This paper introduces causal measurement and closed-loop context allocation for generative search, arguing that repeated, feedback-guided generations use evidence more effectively than simply widening a single RAG context. The approach improves portfolio recall by 16.7–20.5 absolute percentage points and scales to 32B-parameter models.

• Standard relevance proxies reportedly fail on hard negatives; a causal leave-one-out probe instead measures whether generation actually relies on each piece of evidence.
• Factorial experiments identify relevance decay as a key weakness of monolithic context widening.
• Distributing compute across sequential generations yields substantially higher portfolio recall than expanding one context window.
• A submodular scheduler and attribution-steered contrastive decoder form a closed loop that promotes integration of fresh evidence and outperforms classical open-loop baselines.