This page organizes papers related to artificial intelligence published around the world. This page is summarized using Google Gemini and is operated on a non-profit basis. The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.
Diverse decoding of large-scale language models is crucial for applications requiring multiple semantically distinct responses, but existing methods primarily achieve lexical diversity rather than semantic diversity. This limitation significantly limits best-of-N strategies, group-based reinforcement learning, and data synthesis. Semantic-guided Diverse Decoding (SemDiD) operates directly in the embedding space and balances quality and diversity through three complementary mechanisms: orthogonal direction guidance, dynamic inter-group repulsion, and probability assessment to remove positional bias. SemDiD uses an adaptive gain function and constrained optimization to balance these conflicting objectives, ensuring a quality threshold and maximum semantic differentiation. Experimental results show that SemDiD consistently outperforms existing methods, improving best-of-N coverage by 1.4-5.2% across various tasks and accelerating RLHF training convergence by 15% while increasing accuracy by up to 2.1%.
Takeaways, Limitations
•
Takeaways:
◦
SemDiD presents a novel decoding methodology that directly enhances semantic diversity.
◦
It shows performance improvements in various applications such as Best-of-N strategy and RLHF training.
◦
It can expand the scope of application of language models by overcoming the limitations of existing methods.
•
Limitations:
◦
The paper did not specifically mention Limitations (although any study may potentially have Limitations).
