Arbor Reports 2.5x Gain on Code Tasks

Arbor, a research platform developed by researchers at Renmin University of China and Microsoft Research, announces that its novel Hypothesis Tree Refinement (HTR) architecture achieves an average 2.5‑fold improvement over baseline models such as Codex and Claude Code on a suite of coding‑related tasks. The benchmark comprises autonomous optimization problems, data‑synthesis challenges, and search‑agent benchmarks drawn from real-world research pipelines. In particular, on the BrowseComp task Arbor’s accuracy climbs from 45.3 % to 67.7 %, while Codex and Claude Code plateau at 50 % and 53.3 % respectively. On Terminal‑Bench 2.0, Arbor scores 72.22 % in development and 77.36 % on held‑out data versus 71 % and 75 % for Claude Code. The key innovation is the tree‑structured pipeline that separates hypotheses, records evidence, and propagates constraints upward, thereby preventing overfitting and enabling precise attribution of improvements to individual design choices. The paper also discusses deployment constraints, noting that the approach depends on robust, trustworthy evaluation metrics to serve as the merge gate. Overall, Arbor demonstrates a practical framework for scaling agent-based code generation with clear, measurable gains over current state‑of‑the‑art models.