Unverified50% confidenceFactExact time
即使只有5-10%的训练样本存在错位，也会导致VLM模型在下游任务上出现显著的性能退化
1
Sources
50%
Confidence
Long-term
Relevance
8/25/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedExperiments conducted by the research team on models such as GPT-2 showed that after only 5–9 generations of recursive training, the quality of model outputs degrades significantly.73% similarUnverifiedAnthropic's research shows that even mild alignment training can cause a 2–8 percentage point decline in model performance on certain reasoning-intensive tasks.72% similarUnverifiedIn 2023, researchers demonstrated that contaminating less than 1% of training data is sufficient to alter model behavior under specific trigger conditions68% similarUnverifiedIn evaluations of early Agent systems, a Stanford research team found that a single-step error rate of 5% can lead to a final failure rate exceeding 40% in 10-step tasks.68% similarUnverifiedRouteLLM, proposed by the LMSys team, trains a routing model to allocate requests between strong and weak models, reducing costs by 40–60% while maintaining over 95% quality.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/798166API
curl https://kongchang.com/api/v1/knowledge/claims/798166MCP
get_claim(id=798166)