Balancing Candidate Quality and Diversity in Reranking
In practical applications of reranking, there is an inherent trade-off between generating high-quality output candidates and ensuring those candidates are sufficiently diverse. System designers must find a balance to optimize the final selection.
0
1
Tags
Ch.5 Inference - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Ch.4 Alignment - Foundations of Large Language Models
Related
Strategies to Enhance Output Diversity for Reranking
Balancing Candidate Quality and Diversity in Reranking
An engineering team implements a system to improve a language model's output. For each user query, the system generates 10 candidate responses and then uses a highly accurate reward model to select the best one. Despite the high accuracy of the reward model, the team observes that the final selected response is rarely a significant improvement over any of the other 9 candidates. Which of the following is the most likely underlying cause for this lack of significant improvement?
Diagnosing Reranking System Performance
Evaluating Candidate Sets for Selection
Critique of Reranking Effectiveness
Learn After
A team is developing a system to generate marketing slogans. The process involves two stages: first, an initial model generates 50 potential slogans, and second, a highly accurate scoring model selects the single best slogan from that set to display. The team observes that while the final selected slogans are grammatically perfect and on-topic, they are often generic and uninspired. They also notice that the initial 50 slogans generated in each batch are usually very similar to one another. Which of the following strategies is the most sound for the team to adopt to improve the creativity of the final output?
Optimizing a Chatbot Response System
Evaluating System Design for Code Generation