Energy & materialsResearch findings
Introducing KEEP-Bench: A Diagnostic Benchmark for Evidence Preservation in Context Compression
Original published 2026.10.20 02:55 KSTAdded to Blackdeck 2026.10.20 18:39 KST
What has changed?
A diagnostic benchmark directly evaluates evidence fidelity across 4,102 cases in three challenging settings requiring preservation of reference resolutions, qu
Key points
- Context compression may introduce faithfulness issues through information bottlenecks
- QA metrics underestimate fidelity loss due to correct answers from unfaithful compression
- KEEP-Bench directly assesses evidence preservation in three challenging scenarios
Connect it to your work
Increased output budget benefits only strong models and widens capability gaps; evidence preservation comes at high computational cost
What remains unproven
This research note is based on an abstract; check the original for test conditions and application scope.
Originals used for this note
Links open the original publisher in a new tab.