Meta researchers have introduced two new papers detailing advancements in AI scaling laws and agent harness development. The first paper proposes a 'Skaling law' that couples model capacity and training data, improving loss prediction accuracy and reducing compute needs for planning pretraining budgets. The second paper presents EvoHarness-RL, a method for agents to learn harness policies offline, enabling more robust long-horizon task execution and demonstrating strong performance with the Qwen3_8B model on the ALFWorld benchmark. AI
IMPACT These advancements in scaling laws and agent harness development could lead to more efficient AI training and more capable long-horizon agents.
RANK_REASON The cluster contains two academic papers detailing new research findings and methodologies from Meta.
Read on X — Omar Sanseviero (HF research) →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →