束搜索
自然语言生成领域的经典解码算法，起源于语音识别和机器翻译，在每一步保留得分最高的K条候选序列，通过有限的并行探索平衡搜索质量与计算成本，可用于测试时扩展
Timeline (last 90 days)
Search-tree methods (such as MCTS and beam search) rely on precise simulation of game rules; their advantage is being interpretable and requiring no training data, but each step requires expanding a large number of branches, creating heavy real-time pressure.
Even when using beam search decoding as mitigation, the group temporal tax gap caused by pruning still retained a +86% increase.
Beam search maintains k candidate paths simultaneously and can find globally better sequences, but computational overhead grows linearly with beam width.
该项目在解码策略上放弃束搜索，采用最小贝叶斯风险解码（MBR）与决策论解码
Beam search aims to generate the candidate sequence with the highest probability, whereas MBR decoding selects the candidate with the minimum expected translation error.
All Facts (5)
Search-tree methods (such as MCTS and beam search) rely on precise simulation of game rules; their advantage is being interpretable and requiring no training data, but each step requires expanding a large number of branches, creating heavy real-time pressure.
50%UnverifiedEven when using beam search decoding as mitigation, the group temporal tax gap caused by pruning still retained a +86% increase.
50%UnverifiedBeam search maintains k candidate paths simultaneously and can find globally better sequences, but computational overhead grows linearly with beam width.
50%Unverified该项目在解码策略上放弃束搜索，采用最小贝叶斯风险解码（MBR）与决策论解码
50%UnverifiedBeam search aims to generate the candidate sequence with the highest probability, whereas MBR decoding selects the candidate with the minimum expected translation error.
50%