test-time compute
The additional computation a model consumes during inference. Accuracy is improved by generating longer chains of thought when answering, sampling multiple candidate paths, or performing self-verification. It is essentially a trade of computation for accuracy.
Timeline (last 90 days)
Test-Time Compute improves answer quality by adding sampling, search, or reflection steps at the inference stage, for example Self-Consistency sampling multiple reasoning paths in parallel and then voting, and Reflexion letting the Agent self-criticize and retry; each sample and each parallel path corresponds to a complete forward computation.
Recurrent Depth is a specific implementation of the test-time compute research direction and shares the same lineage as DeepMind's Adaptive Computation Time research.
Multi-agent systems are essentially a way to scale test-time compute in parallel.
Problems in mathematics and computer science have clear logical steps, making them well-suited for paradigms such as chain-of-thought reasoning and test-time computation.
The optimistic camp believes that as model scale grows, chain-of-thought techniques mature, and test-time compute strategies are enhanced, LLMs' progress on ARC-type tasks is genuine.
The evaluation closed loop is a typical application of the test-time compute research direction, improving output quality by investing more computation during the inference stage rather than relying on larger model parameters.
测试时计算允许模型根据问题难度动态调整计算投入，是近年AI领域重要的范式转变
测试时计算（test-time compute）允许模型根据问题难度动态调整计算投入，是近年AI领域的重要范式转变
Test-time compute is essentially a strategy of trading more FLOPs for higher intelligence performance.
All Facts (9)
Test-time compute is essentially a strategy of trading more FLOPs for higher intelligence performance.
70%UnverifiedTest-Time Compute improves answer quality by adding sampling, search, or reflection steps at the inference stage, for example Self-Consistency sampling multiple reasoning paths in parallel and then voting, and Reflexion letting the Agent self-criticize and retry; each sample and each parallel path corresponds to a complete forward computation.
50%UnverifiedRecurrent Depth is a specific implementation of the test-time compute research direction and shares the same lineage as DeepMind's Adaptive Computation Time research.
50%UnverifiedMulti-agent systems are essentially a way to scale test-time compute in parallel.
50%UnverifiedProblems in mathematics and computer science have clear logical steps, making them well-suited for paradigms such as chain-of-thought reasoning and test-time computation.
50%UnverifiedThe optimistic camp believes that as model scale grows, chain-of-thought techniques mature, and test-time compute strategies are enhanced, LLMs' progress on ARC-type tasks is genuine.
50%UnverifiedThe evaluation closed loop is a typical application of the test-time compute research direction, improving output quality by investing more computation during the inference stage rather than relying on larger model parameters.
50%Unverified测试时计算允许模型根据问题难度动态调整计算投入，是近年AI领域重要的范式转变
50%Unverified测试时计算（test-time compute）允许模型根据问题难度动态调整计算投入，是近年AI领域的重要范式转变
50%