Unverified50% confidenceFactExact time
In tasks that require maintaining long-horizon consistency, the model exhibits "drift" after being extended to dozens of steps, that is, the goal, state, and logic are not maintained.
1
Sources
50%
Confidence
Long-term
Relevance
11/1/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedTraditional 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.80% similarUnverifiedIf the capability set changes halfway through an inference task, forcibly switching the backend or model may cause inconsistent results or even crashes; a reasonable strategy is to apply the new capabilities at task boundaries.78% similarUnverifiedModels 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.77% 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.77% similarUnverifiedWhen model scale crosses certain thresholds, emergent capabilities not previously present arise (emergent abilities), such as multi-step logical reasoning and complex mathematical computation.77% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1123103API
curl https://kongchang.com/api/v1/knowledge/claims/1123103MCP
get_claim(id=1123103)