SCAN
A benchmark for evaluating the systematic compositional generalization ability of neural network systems. It has been used to reveal shortcomings in the compositional capabilities of large language models.
Timeline (last 90 days)
Researchers have designed test sets such as SCAN, COGS, and ARC to probe compositional generalization ability.
Benchmarks such as SCAN and COGS reveal clear shortcomings in large language models' systematic compositional abilities.
The SCAN and LOOK algorithms originated in the disk scheduling domain and were later adopted in elevator control.
The LOOK algorithm is an improved version of SCAN, where the elevator only travels to the last requested floor in the current direction before reversing, rather than being forced to run all the way to the top or bottom floor, thereby reducing unnecessary travel.
SCAN, LOOK, and Nearest Car First algorithms are all single-objective greedy strategies, which are prone to elevator bunching in scenarios involving multiple elevators and non-uniform passenger traffic.
Lake, Baroni, and others validated on benchmarks such as SCAN and COGS that current deep learning models have significant shortcomings in systematic generalization.
All Facts (7)
Redis的SCAN命令使用反向二进制迭代（reverse binary iteration）的游标推进算法
90%UnverifiedResearchers have designed test sets such as SCAN, COGS, and ARC to probe compositional generalization ability.
50%UnverifiedBenchmarks such as SCAN and COGS reveal clear shortcomings in large language models' systematic compositional abilities.
50%UnverifiedThe LOOK algorithm is an improved version of SCAN, where the elevator only travels to the last requested floor in the current direction before reversing, rather than being forced to run all the way to the top or bottom floor, thereby reducing unnecessary travel.
50%UnverifiedThe SCAN and LOOK algorithms originated in the disk scheduling domain and were later adopted in elevator control.
50%UnverifiedSCAN, LOOK, and Nearest Car First algorithms are all single-objective greedy strategies, which are prone to elevator bunching in scenarios involving multiple elevators and non-uniform passenger traffic.
50%UnverifiedLake, Baroni, and others validated on benchmarks such as SCAN and COGS that current deep learning models have significant shortcomings in systematic generalization.
50%