Unverified50% confidenceSolutionExact time
The method's procedure is: starting from a baseline reasoning program, each round uses an LLM as a judge, showing it the current program's code and failure cases on training problems and asking it to provide improvement evolution actions; candidate programs are kept in a small program pool ranked by training accuracy, and the next round of editing is performed on the highest-scoring program in the pool.
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe LLM routing space already has precedents in RouteLLM, Martian, and others; the core idea is to train lightweight classifiers to predict task difficulty and type before inference, then map the results to a model selection strategy.77% similarUnverifiedLLM-Blender 通过训练专用排序模型 PairRanker 来评估多个 LLM 的候选输出并融合最优结果75% similarVerifiedTraining strategies for mathematical LLMs include continued pre-training on large-scale mathematical corpora on top of general pre-training, followed by supervised fine-tuning with chain-of-thought reasoning data, combined with reinforcement learning using process reward models.74% similarUnverifiedAn efficient inference engine can accelerate rollout generation, thereby shortening the iteration cycle of RL training.74% similarUnverified执行轨迹记录模型或人类解决问题的完整中间步骤，使训练系统能识别推理链条中具体哪一步出错，从而提供更精细的学习信号73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1102164API
curl https://kongchang.com/api/v1/knowledge/claims/1102164MCP
get_claim(id=1102164)