Cosmos Reason 2 tasks
What it computes
Cosmos Reason 2 through llama.cpp
Baseline and comparison design
Baseline: Qwen3-VL-2B base
Metric: balanced accuracy (), higher is better
Pairing unit: query, n = 500
Pre-registered acceptance criterion
Cosmos beats the Qwen base model only if McNemar p < 0.05
SC-015-03 of 015-vlm
Lane
Lanes: precompute
estimate
Documentation page: docs/methods/m22-cosmos-vlm-tasks.md
Status: Not yet run