Context compression often loses important evidence in long texts
KEEP-Bench: Evaluating Evidence Preservation in Context Compression
Computation and Language
Summary
When computers summarize or compress long texts, they can sometimes lose important details that help explain or support what's being said. The authors created a test called KEEP-Bench to check how well models keep these important facts in compressed text. They found that even the best models often miss key evidence, especially when texts have complex information. This means future work should focus on making compressors that preserve facts better without needing huge computing power.
What this means in practice
- •For software engineers: Build text compression tools that better preserve important evidence in long documents for reliable content summarization.
- •For ai system developers: Optimize models to maintain factual details efficiently without requiring very large output sizes or model scales.
Authors
Ziqing Man, Di Wu, Jia-Chen Gu, Junyi Zhang, Kai-Wei Chang, Nanyun Peng
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.