Researchers have developed a method to understand the origins of social reasoning capabilities in language models by analyzing their training data. Using gradient-based attribution on the Dolma3 dataset, they mapped specific regions of the corpus that contribute to social versus STEM reasoning. The study found that social and STEM reasoning draw from distinct data sources, with reasoning capabilities being more sensitive to these distinctions than factual knowledge. Targeted unlearning experiments partially validated these findings by showing that removing high-attribution data bins degraded aligned benchmarks. AI
IMPACT Provides a new method for understanding model behavior and potentially improving training data curation for specialized reasoning skills.
RANK_REASON The cluster contains an academic paper detailing a new methodology for analyzing language model capabilities.
- ARC challenge
- arXiv
- Bergson
- Dolma3
- Hugging Face
- MMLU Social Sciences
- MMLU STEM
- OLMo-7B
- SocialIQA
- WebOrganizer
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
