Unverified50% confidenceSolutionExact time
The paper designs an LLM-judged, history-aware adaptive repair pipeline comprising three actions: identifying single-hop failure cases, targeted fine-tuning, and removing or isolating noisy triples.
1
Sources
50%
Confidence
Long-term
Relevance
9/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe LLM-as-judge approach suffers from the judge model's own volatility, bias, and inconsistency, which introduces new noise into evaluation results.64% similarUnverifiedThe third innovation in the paper is the use of the LLM itself to iteratively optimize prompts (LLM-assisted Prompt Optimization).62% similarUnverifiedLLM的模式坍缩主要由自回归生成机制与RLHF优化的组合效应驱动，可从解码策略、训练目标和提示设计三个维度入手62% similarUnverified自适应过拟合由Dwork等人在2015年论文《The Reusable Holdout》中正式阐述，即分析者根据验证集反馈反复调整模型时，模型选择过程会逐渐拟合验证集的特异性噪声61% similarUnverifiedLLM可观测性借鉴了传统分布式系统可观测性的三大支柱——日志、指标和追踪，但针对大语言模型的非确定性特征做了适配60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/919817API
curl https://kongchang.com/api/v1/knowledge/claims/919817MCP
get_claim(id=919817)