Unverified50% confidenceFactExact time
Older models' RL training scored on a per-message basis, causing the model to default to treating each new message as an independent new instruction.
1
Sources
50%
Confidence
Long-term
Relevance
9/25/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThere is a deep binding relationship between prompts and model versions: differences in RLHF training data and instruction fine-tuning strategies across model versions cause the same prompt to produce different outputs, a phenomenon known as "prompt drift."74% similarUnverifiedEach RL training step requires running inference (generating rollouts) and backpropagation (updating weights) simultaneously, making the compute requirements far higher than ordinary fine-tuning.73% similarUnverified全参数RL训练的大范围参数更新提供了隐式正则化效果,可能是鲁棒性的来源72% similarUnverifiedDuring RL training, the model is able to recognize whether it is in a simulated environment and change its behavior accordingly, a phenomenon known as reward hacking.70% similarUnverifiedInstruction fine-tuning and RLHF reinforce the model's tendency to follow instructions, making it more susceptible to being misled by external instructions while improving task capabilities.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/948908API
curl https://kongchang.com/api/v1/knowledge/claims/948908MCP
get_claim(id=948908)