Unverified50% confidenceOpinionExact time
Existing long-context benchmarks focus almost entirely on recall capability, neglecting state tracking and pattern learning.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/20/2026
First Seen
Valid until: 1/18/2027
Sources
Related Entities
Related Claims
VerifiedCurrent frontier models face attention decay when handling long-range dependencies, with the ability to maintain consistency of early-established settings declining significantly as context length increases.74% similarVerified仅用成功轨迹训练的模型往往存在分布外泛化的严重缺陷，遇到工具返回错误时便丧失纠偏能力71% similarUnverifiedBehavioral drift has long lacked a unified quantitative standard, relying more on trajectory-level structured comparison.71% similarUnverifiedThe learning of existing orchestration strategies relies on feedback signals that are often sparse and delayed, with systems typically only obtaining outcome judgments after the entire task is completed.70% similarUnverifiedTypical generation-stage failures include hallucinations caused by ignoring context, middle-content forgetting due to context overflow, insufficient multi-hop reasoning capability, and instruction-following deviation.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1074002API
curl https://kongchang.com/api/v1/knowledge/claims/1074002MCP
get_claim(id=1074002)