Unverified50% confidenceFactExact time
LongBench is a widely used comprehensive benchmark in the field of long-context understanding, covering single-document QA, multi-document QA, summarization, few-shot learning, code completion, and synthetic tasks, with document lengths generally exceeding 8K tokens and some exceeding 32K tokens.
1
Sources
50%
Confidence
Long-term
Relevance
10/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLongMemEval focuses on long-context memory retrieval, requiring models to precisely locate specific facts within dialogue histories of tens of thousands to hundreds of thousands of tokens.65% similarUnverified对于单用户交互式场景推理速度达到5-10 token/s通常就能提供流畅体验，而代码生成、长文档摘要或多轮Agent调用往往需要30 token/s以上64% similarUnverifiedLong-term memory evaluation benchmarks (such as LongMemEval and the MemGPT benchmark) assess five dimensions: single-hop information extraction, cross-session reasoning, knowledge updating, temporal reasoning, and abstention judgment.64% similarUnverifiedBecause PageIndex requires calling a large model for reasoning navigation with each query, it may incur higher token consumption and latency than a simple vector query.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/983262API
curl https://kongchang.com/api/v1/knowledge/claims/983262MCP
get_claim(id=983262)