Model Releases

Four small architecture decisions can cost up to 47% of a model's long-context performance. New research from Ai2, Carnegie Mellon, and the …

Four small architecture decisions can cost up to 47% of a model's long-context performance. New research from Ai2, Carnegie Mellon, and the University of Washington isolates them. Normalization, GQA,

DGX agentx-post
model-releasesdair-ai--x

Four small architecture decisions can cost up to 47% of a model's long-context performance. New research from Ai2, Carnegie Mellon, and the University of Washington isolates them. Normalization, GQA, pretraining context length, and sliding window attention. Each one is made by at least one of the Olmo, Llama, and Qwen dense families, and each looks harmless on its own. Combine three or more and downstream long-context performance drops significantly. None of it shows up in short-context loss or validation sets, which is how the choices survive into shipped models. Applying context extension early in pretraining exposes the problem. They release OlmPool, 26 comparable 7B models with checkpoints before and after extension, from over 170,000 GPU hours of training. Paper: https://arxiv.org/abs/2608.10296 Track more trending AI papers in our academy: https://academy.dair.ai/

Related

Source: DAIR.AI (X) | 2026-08-12

Loading related sources…