Newest papers

KEEP-Bench: Evaluating Evidence Preservation in Context Compression

Ziqing Man, Di Wu, Jia-Chen Gu, Junyi Zhang, Kai‐Wei Chang, Nanyun Peng · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26827 · cited by 0 works (OpenAlex) · open-access PDF

Indonesian

Penilaian yang ada bergantung pada metrik QA, tetapi pembaca masih dapat menjawab dengan benar dari kompresi yang tidak setia, sehingga metrik ini meremehkan kerugian seperti itu. Untuk memajukan kesenjangan ini, kami memperkenalkan KEEP-Bench, benchmark diagnostik yang secara langsung menilai kesetiaan kompresi dengan 4,102 kasus di tiga pengaturan yang menantang yang membutuhkan bukti untuk resolusi referensi yang dapat diandalkan, kualifikasi eksplisit, dan perspektif minoritas. Eksperimen menunjukkan bahwa bahkan dengan prompting yang disesuaikan, model terkuat tetap tidak dapat diandalkan: meskipun jendela konteks melebihi token 100K dan input rata-rata hanya token 3.1K, model informasi terbuka terbaik (31B) menemukan semua bukti yang diperlukan hanya dalam 66.8% kasus, dan model utama yang paling dimiliki hanya dalam biaya 61.6%. Prestasi juga

Machine translation (Meta M2M100); the English original follows.

Abstract

Context compression improves long-context efficiency but may introduce faithfulness issues through an information bottleneck. Existing evaluations rely on QA metrics, but a reader can still answer correctly from an unfaithful compression, so these metrics underestimate such loss. To bridge this gap, we introduce KEEP-Bench, a diagnostic benchmark that directly assesses compression faithfulness with 4,102 cases across three challenging settings that require keeping evidence for plausible reference resolutions, explicit qualifiers, and minority perspectives. Experiments show that even with tailored prompting, the strongest models remain unreliable: despite context windows exceeding 100K tokens and inputs averaging only 3.1K tokens, the best open-weight model (31B) preserves all required evidence in only 66.8% of cases, and the best proprietary model in only 61.6%. Performance also varies sharply across evidence structures, so conclusions from limited structures may be biased. We further introduce a metric, Evidence Preservation Return (EPR), to relate evidence recall to inference FLOPs, and find that it peaks at moderate scales (4B-8B), where recall remains low. Additional output budget mainly benefits already strong models and widens capability gaps, yet the stronger preservation of these models comes at a steep computational cost. These findings highlight the importance for future work to investigate compressors with strong information preservation under a small computation budget, rather than relying on a large output budget or model scaling.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Context (archaeology), Inference, Benchmark (surveying), Compression (physics), Data compression

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange