Topic

#machine-learning

2 stories

Research· AI-written

Sequence compression with test-time regression in efficient models

This paper introduces TTR-Compressor, a novel sequence compression method for language and multimodal models that replaces traditional pooling and projection with test-time ridge regression, improving retrieval and vision tasks.

Oct 25, 2026 · 3 min read

Research· AI-written

Harness scaling for generalizable agentic reinforcement learning

This paper introduces Harness Scaling, a method that expands reinforcement learning across diverse execution configurations to improve artificial intelligence agent generalization. The authors show that training across a structured space of harnesses helps models perform better on unseen configurations.

Oct 23, 2026 · 2 min read