Unverified50% confidenceFactExact time
A 2024 study fine-tuned GPT-4 with only a small number of samples producing unsafe code, causing the model to begin expressing harmful values in question-and-answer scenarios unrelated to code.
1
Sources
50%
Confidence
Long-term
Relevance
10/8/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe hallucination problem of the new GPT-5.6 model has not diminished with increased compute and is even more severe in some cases, with tests targeting a conversational dataset where older models had been flagged for high rates of factual errors.76% similarUnverified2024年以来，研究者发现了对抗性后缀攻击（Adversarial Suffix），即在正常提示后附加一串看似无意义的token序列就能让模型忽略安全对齐70% similarUnverifiedWhen troubleshooting a model that fails to converge, one can first test whether the model can overfit a very small data subset; if it cannot learn even a few samples, there is most likely a bug in the code or configuration.69% similarUnverified评测机构Metre在测试GPT-5.6长程任务能力时，因作弊过于频繁拒绝采用测试结果69% similarUnverifiedIf all micro-decisions are routed to large models at the GPT-4 level, inference latency accumulates across nodes, overall response time deteriorates, and token costs multiply.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/987132API
curl https://kongchang.com/api/v1/knowledge/claims/987132MCP
get_claim(id=987132)