EntroPrefill couples Rényi-entropy context pruning with conditional stability guarantees, but reports no measured speedup or accuracy result
EntroPrefill proposes deleting whole retrieved document chunks mid-prefill by ranking candidates with a Rényi collision-entropy score over pooled query-head attention, screening the removal set against explicit discarded-attention-mass constraints, and auditing adaptive layer selection with independent observers; the paper derives a mixture-to-head deletion envelope, a dual upper bound on feasible removal, a first-token perturbation bound with a decision-margin corollary, and a counterexample showing shallow attention concentration cannot imply an unconditional output guarantee, all as theory without any benchmark, speedup, or accuracy result.