ResearchBenchmarksTask Completion on Real-world (test)Task Completion on Real-world (test)Follow8.09ScoreVanilla7.9868.0138.048.067Feb 12, 2026Evaluation resultsMethodMethodLinksScoreAbsolute DeltaVanillaLLM Backbone=Llama-3-8...LLM Backbone=Llama-3-8b-instruct2026.028.09—VanillaLLM Backbone=Qwen3-8bLLM Backbone=Qwen3-8b2026.028.09—VanillaLLM Backbone=GPT-oss-20bLLM Backbone=GPT-oss-20b2026.028.09—PPOptLLM Backbone=Qwen3-8bLLM Backbone=Qwen3-8b2026.028.080.01PPOptLLM Backbone=Llama-3-8...LLM Backbone=Llama-3-8b-instruct2026.028.050.04PPOptLLM Backbone=GPT-oss-20bLLM Backbone=GPT-oss-20b2026.027.990.1