June 11, 2026

Arbor introduces a framework for fully autonomous ML research. An LLM-based coordinator manages a persistent Hypothesis Tree linking hypotheses, experimental artifacts, and learned insights. Executor agents test individual hypotheses in isolated sandboxes, allowing knowledge to accumulate across many experimental rounds rather than being discarded after each run. On MLE-Bench Lite, Arbor reaches 86.36% Any Medal score — over 2.5× the relative held-out gains of both Codex and Claude Code under identical compute budgets.

DeNovoSWE addresses a gap in AI code agents: most training data covers bug-fixing in existing codebases, not building complete repositories from scratch. The benchmark provides 4,818 instances where each requires generating a full repo from documentation. A divide-and-conquer critic-repair pipeline with difficulty-aware filtering produces high-quality training trajectories. Fine-tuning Qwen3-30B-A3B on this data pushes BeyondSWE-Doc2Repo performance from 5.8% to 47.2%.

Z-Reward replaces single scalar reward values with distributions over rubric scores for RLHF in text-to-image generation. A 27B teacher model reasons explicitly to produce score distributions; a student model internalizes this reasoning at inference time via Reasoning-Internalized Score Distillation (RISD), without needing chain-of-thought at runtime. Group-wise Direct Score Optimization (GDSO) combines policy-gradient rewards with direct distribution supervision. The 27B teacher achieves 89.6% human preference accuracy; the 9B student matches at 88.6%; as a differentiable reward signal during generation, achieves 41.3% net human-preference improvement.