首页 > AI前沿 > MORSE: Multi-Context Ordering via Reverse Scoring for Evidence-Preserving Compression

MORSE: Multi-Context Ordering via Reverse Scoring for Evidence-Preserving Compression

arXiv自然语言 2026-09-23 13:32 6 阅读 查看原文

Likelihood-based context compression can account for cross-context redundancy through sequential scoring, but this makes compression outcomes sensitive to context order.

We show that different permutations of the same context collection can produce markedly different evidence-retention outcomes under an unchanged compressor.

We attribute this sensitivity to information preemption: earlier partially relevant contexts can absorb credit for shared information, suppressing the incremental score of later, stronger evidence carriers and increasing their risk of removal.

Controlled pair-swap interventions directly support this mechanism by showing that evidence-first ordering substantially improves supporting-evidence survival.

To address this problem, we introduce MORSE, a compression-aware method for evidence-preserving context ordering.

MORSE applies a common reverse query-evidence principle to both individual contexts and compressed candidate outputs, using the former to construct an evidence-first anchor and the latter to guide compression-aware permutation selection.

Across multi-hop QA benchmarks, compression procedures, budgets, and scoring models, MORSE consistently improves evidence preservation over static reverse ordering and compute-matched random search, with corresponding overall improvements in downstream QA.

Our code is available at https://github.com/tbn5pj/MORSE_code.