Microsoft Research introduces Arbor
A generalist autonomous research agent that uses persistent hypothesis-tree refinement to turn long-horizon exploration into cumulative learning. It beats Codex and Claude Code across 6 research tasks and hits 86% Any-Medal on MLE-Bench Lite.
Positive users praise Arbor's hypothesis-tree refinement as methodical search superior to token prediction, while one user questions whether it shows real learning or just lucky results.
Toward Generalist Autonomous Research via Hypothesis-Tree Refinement
Paper page: https://paperswithcode.co/paper/2606.11926
CLI, code & docs: https://github.com/RUC-NLPIR/Arbor
Project page: https://RUC-NLPIR.github.io/Arbor/
paper: https://huggingface.co/papers/2606.11926
Toward Generalist Autonomous Research via Hypothesis-Tree Refinement
@HuggingPapers i want the ugly trace more than the medal score. if Arbor only shows the winning branch, nobody can tell whether it learned research strategy or just got lucky with retries.
@HuggingPapers Persistent hypothesis-tree refinement is methodical search. token prediction never beats systematic exploration. that's Codex's real problem
@HuggingPapers @_akhaliq ⏫⏫
@_akhaliq ⏫
Microsoft Research introduces Arbor
A generalist autonomous research agent that uses persistent hypothesis-tree refinement to turn long-horizon exploration into cumulative learning. It beats Codex and Claude Code across 6 research tasks and hits 86% Any-Medal on MLE-Bench Lite.