A reusable environment layer plus three training recipes yield open-source SOTA agents — 67.5% SWE-bench Verified and 68.4% GUI average — though the abstract and body report different headline numbers.
We embed each task intent with Qwen/Qwen3-Embedding-8B and greed- ily remove tasks whose cosine similarity to a previously kept task exceeds 0.99 (-124,454, 88.9%→15,601)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Orchard: An Open-Source Agentic Modeling Framework
A reusable environment layer plus three training recipes yield open-source SOTA agents — 67.5% SWE-bench Verified and 68.4% GUI average — though the abstract and body report different headline numbers.