List your company

Agentic Programming on SWE-bench-derived OOD (test)

75.89Average Performance (%)Oracle (cost-aware)
-0.248419.518339.28559.0517Jun 22, 2026

Evaluation results

MethodLinks
Oracle (cost-aware)Family=Bound2026.0675.89032.712.32
ACRouterFamily=Agentic2026.0662.51752.971.18
Always-Opus 4.6Family=Single-Model2026.0657.1426.789.280.64
Qwen3.5-0.8B-FinetunedFamily=Trained2026.0655.3627.274.810.74
LinUCBFamily=Bandit, Aggregation=5-seed avg2026.0649.8231.151.90.96
LinTSFamily=Bandit, Aggregation=5-seed avg2026.0646.4335.961.910.75
RandomFamily=Single-Model2026.0631.2550.436.760.85
RouteLLM-BERTFamily=Trained2026.0621.4359.416.481.3
LogRegFamily=Trained2026.0619.6461.816.791.17
Always-Kimi-K2.5Family=Single-Model2026.0618.7562.315.371.22
kNN RetrievalFamily=Heuristic2026.0614.2966.79.861.45
TF-IDF+MLPFamily=Trained2026.0613.3967.911.441.17
RouteLLM-MFFamily=Trained2026.068.9372.79.50.94
Always-Qwen3.5-PlusFamily=Single-Model2026.062.6880.114.110.19