New LURE method trains LLMs for reasoning via pursuit-evasion games

Researchers have developed a novel approach called LURE for training large language models (LLMs) in reasoning tasks without requiring human-annotated datasets. This method frames the training process as a pursuit-evasion game, where an LLM evader strategically creates tasks of varying difficulty, and a pursuer LLM attempts to solve them. The system learns to position tasks at a difficulty level where the solver succeeds about half the time, optimizing for a 'capture-frontier' reward. This technique has demonstrated superior performance compared to existing baselines across multiple reasoning environments and LLM architectures, achieving stronger out-of-distribution zero-shot accuracy. AI

IMPACT Introduces a novel self-play training paradigm for LLM reasoning, potentially reducing reliance on human-annotated data.

RANK_REASON This is a research paper detailing a new method for LLM training. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• Influence Flower
• Litmaps
• LLM
• LURE
• ScienceCast
• scite Smart Citations

• paper
• model release

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