Conceptizer framework trains LLMs to predict meaning, not just next token

A new training framework called Conceptizer has been developed to improve language models by rewarding the prediction of meaning, not just the next token. This approach utilizes a fixed vocabulary of approximately 9,000 concepts, derived from contextual synonyms, to provide a concept-level training signal without increasing model parameters. Experiments on models ranging from 1 billion to 8 billion parameters demonstrated that Conceptizer achieves lower perplexity at concept-target positions and assigns higher probabilities to contextual synonyms compared to baseline models. Furthermore, it enhanced performance on knowledge and sentiment tasks while maintaining comparable results on commonsense reasoning. AI

IMPACT This research could lead to language models that better understand and generate nuanced meaning, improving their utility in tasks requiring deeper comprehension.

RANK_REASON The item is a research paper published on arXiv detailing a new training framework for language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• Bibliographic Explorer
• CatalyzeX
• Conceptizer
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• Influence Flower
• Litmaps
• next-token prediction
• ScienceCast
• scite Smart Citations

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →