Unverified50% confidenceBenchmarkExact time
On KEEP-Bench, the best closed-source model retained all required evidence in only 61.6% of cases.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/20/2026
First Seen
Valid until: 1/18/2027
Sources
Related Entities
Related Claims
UnverifiedEven with customized prompts, the strongest models' evidence retention on KEEP-Bench remains unreliable.78% similarUnverifiedKEEP-Bench contains 4,102 cases covering three challenging scenarios: retaining evidence to support plausible referential resolution, explicit qualifiers, and minority viewpoints.66% similarUnverifiedIn context compression, average answer quality may remain stable, but a rare yet necessary piece of evidence may have quietly disappeared.61% similarUnverifiedThe coarse-filtering stage should ensure sufficient recall margin — it is better to over-recall and let the model eliminate candidates than to miss the correct option by saving tokens, because recall failures at the retrieval layer cannot be recovered by subsequent models.60% similarUnverifiedA more reliable way to use thresholds is to combine relative ranking (such as keeping passages in the top K results whose scores are significantly higher than the rest) with verification-based judgment (having the model judge whether a passage truly supports the answer), rather than relying on a single static threshold.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1073301API
curl https://kongchang.com/api/v1/knowledge/claims/1073301MCP
get_claim(id=1073301)