New AI system CogRSI evolves reasoning strategies but fails rule audits

Researchers have developed CogRSI, a system designed to improve language models' reasoning abilities by having them evolve and audit their own cognitive strategies. CogRSI utilizes a grounded reader that resolves disagreements based on written principles, outperforming self-consistency methods at a similar token cost. While the learned principles show promise by transferring to unseen data and scoring higher than expert-defined rules on the BigToM benchmark, a critical audit revealed that many of these learned rules are factually incorrect and can lead to wrong answers in specific scenarios. This highlights a crucial gap: a strategy can be effective in practice without being logically sound, necessitating both performance and rule-based auditing. AI

IMPACT Highlights a critical need for robust auditing of AI reasoning strategies beyond simple performance metrics.

RANK_REASON Research paper detailing a new method for language model reasoning strategies. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• BigToM
• CatalyzeX
• CogRSI
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• Influence Flower
• ScienceCast

• paper
• model release

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