Unverified50% confidenceFactExact time
Omni-modal large models perform poorly on temporal grounding tasks, exhibiting a phenomenon in which the precision of internal attention localization is significantly higher than the precision of output timestamps, which the authors call the "temporal generation gap."
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
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.75% similarUnverifiedFew-step generation approaches involve a quality-speed trade-off: with too few steps, the model struggles to correct early errors, which can lead to blurry details or unnatural motion.75% similarUnverifiedAs task chains grow longer, the capacity limitations of the context window and attention decay affect the model's ability to accurately maintain the initial goal, which is one of the root causes of the 'goal drift' phenomenon in long-horizon tasks.75% similarUnverifiedExpressing model capability gaps in terms of time lag is more easily understood and communicated than relying solely on benchmark scores.74% similarUnverified当模型同时优化任务准确性和置信度校准时，两个梯度方向可能相互干扰，涉及灾难性遗忘和目标冲突问题73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1073800API
curl https://kongchang.com/api/v1/knowledge/claims/1073800MCP
get_claim(id=1073800)