Learn Before
Scaling the Search Space in Search Scaling
Scaling the search space is a key aspect of search scaling that involves expanding the set of candidate output sequences considered during the decoding process to improve the chances of discovering higher-quality outputs. A simple example of this is increasing the beam width in beam search, which allows more candidate sequences to be explored in parallel at each decoding step. This is particularly useful in tasks where the optimal solution is not immediately apparent from local, step-by-step decisions.
0
1
Tags
Ch.5 Inference - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Related
Benefit of Search Space Expansion in Complex Decoding Tasks
Computational Costs of Search Scaling
Scaling Output Length in Search Scaling
Scaling the Search Space in Search Scaling
An engineer is using a fixed, pre-trained language model to generate a complex travel itinerary. The initial outputs are often functional but fail to find the most optimal route. The engineer cannot alter the model's internal parameters. Which of the following adjustments to the generation process is a direct application of search scaling to find a better itinerary?
Applying Search Scaling Strategies
Analyzing Trade-offs in Inference-Time Search Configuration
Implicit Search Scaling in Search Procedures
Learn After
Optimizing Creative Text Generation
A machine learning engineer is attempting to improve the quality of summaries generated by a large language model. Their strategy is to drastically expand the number of potential summary sequences the model considers before selecting the final output. Which of the following statements best evaluates the primary trade-off of this approach?
Improving Complex Reasoning in LLMs
In the context of generating output from a language model, broadening the search for potential sequences is a guaranteed method to improve the factual accuracy of the final output, assuming sufficient computational resources are available.
Beam search