New PAC-learning model for stochastic autoregressive learning introduced

Researchers have introduced a new PAC-learning model for binary stochastic autoregressive learning, inspired by the iterative token generation process of Large Language Models (LLMs). This model generalizes deterministic autoregressive learning and considers three forms of supervision: base one-step samples, chain-of-thought (CoT) samples revealing full trajectories, and end-to-end (e2e) samples revealing only the final token. The study analyzes the minimum number of samples required for learning under squared loss error and demonstrates that stochastic autoregressive learning differs significantly from its deterministic counterpart, with complex dependencies between the learning tasks. AI

IMPACT Introduces a theoretical framework for understanding and potentially improving the iterative generation processes in LLMs.

RANK_REASON Academic paper introducing a new learning model. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• end-to-end reinforcement learning
• Hugging Face
• Joshi et al.
• LLMs

• paper
• other

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