Unverified50% confidenceFactExact time
Models that perform similarly on single-event tasks diverge significantly from one another once multi-step history is involved, and single-step performance can hardly predict multi-step performance.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/16/2026
First Seen
Valid until: 1/14/2027
Sources
Related Entities
Related Claims
Unverified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76% similarUnverifiedTraditional large language models often fail at complex tasks spanning multiple steps and multiple time points due to context loss, errors in intermediate steps, or goal drift.75% similarUnverifiedWhen reasoning chains grow longer and multi-level derivation is required, the accuracy of System 1 models collapses sharply, because a single forward pass cannot carry too many intermediate states.75% similarUnverifiedMost frontier models exhibit extremely high variance when running the same task repeatedly, potentially producing drastically different results.75% similarUnverifiedIn multi-step tasks, small errors at individual steps accumulate progressively into a snowball effect, where errors in early steps contaminate the input context of subsequent steps74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1041953API
curl https://kongchang.com/api/v1/knowledge/claims/1041953MCP
get_claim(id=1041953)