Exact Extract — Huawei Training Material, Page 27
"Model performance improves as compute, data, and parameter scale increase"
"Large language models (LLMs) exhibit unexpected skills and behaviors"
Technical Deep Dive:
Emergent capabilities describe model behaviors that may become observable or substantially more effective as neural-network scale, training-data volume, and computational investment increase.
These capabilities are associated with scaling effects in large language models, particularly in reasoning, instruction following, in-context learning, and complex task generalization.
Option A is correct in the intended training context. Scaling model parameters, datasets, and training computation can enable capabilities that are difficult to observe in smaller models.
However, the existence of a single universal critical threshold is not established for every capability. Some apparently abrupt improvements are influenced by the evaluation methodology.
Option B is correct. Emergent capabilities can improve task generalization, problem solving, and contextual understanding, contributing to broader and more flexible model capabilities.
This should not be interpreted as proof that a model has achieved human-level general intelligence.
Option C is incorrect. Emergence does not inherently remove previously acquired model capabilities. Larger models generally retain and expand many foundational abilities, although individual performance regressions can occur.
Option D is incorrect. Scaling improvements are not necessarily uniform across tasks. Some benchmarked capabilities may show nonlinear improvements as model capacity and training increase.
Huawei's training also introduces the Chinchilla scaling perspective, emphasizing the importance of balancing parameter counts with appropriate training-data volumes.
Therefore, A and B represent the intended statements.
Submit