Unverified50% confidenceCautionExact time
Numerical inconsistency between the trainer and inference engine (train-inference mismatch) is a common hidden cause of RL training instability and policy degradation.
1
Sources
50%
Confidence
Long-term
Relevance
9/24/2026
First Seen
Sources
Related Entities
Related Claims
Verified仅用成功轨迹训练的模型往往存在分布外泛化的严重缺陷，遇到工具返回错误时便丧失纠偏能力78% similarUnverified低精度引入的数值扰动会在RL的多轮迭代中通过策略梯度的乘积形式被放大，可能导致训练崩溃或模型性能退化76% similarUnverifiedAutoregressive models suffer from the exposure bias problem, where the model sees real prefixes during training but faces its own erroneous generated prefixes during inference, leading to error accumulation.73% similarUnverifiedThe non-determinism of LLMs (where identical inputs may produce different outputs) fundamentally conflicts with the testing domain's requirement for reproducibility, and is a core engineering challenge for AI testing Agents.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/945485API
curl https://kongchang.com/api/v1/knowledge/claims/945485MCP
get_claim(id=945485)