arbor – skill Dữ liệu & Khoa học cho agent AI

Arbor — Autonomous Optimization via Hypothesis Tree Refinement

Overview

This skill runs an Autonomous Optimization (AO) loop: starting from an existing artifact and a measurable objective, improve it through many rounds of experiment and evaluation — without step-by-step human supervision, while checking for overfitting to the development signal. It's the right tool when the bottleneck isn't writing one good change, but *organizing dozens of trials* so that lessons accumulate instead of evaporating.

It implements Hypothesis Tree Refinement (HTR) from *Arbor* (Jin et al., 2026). The key idea: keep the research state in a persistent hypothesis tree rather than in conversation history. Each node binds a hypothesis, the distilled insight it produced, and a pointer to the artifact version that realizes it. You play the long-lived coordinator that owns this tree and decides where to search; short-lived executor subagents test one hypothesis each in isolated git worktrees and report back. A held-out merge gate admits a change only when it improves on a *test* evaluator executors do not optimize against. This is what turns trial-and-error into cumulative, auditable research.

Use the scripts/tree.py state manager for all the bookkeeping (creating nodes, writing evidence, propagating insights, pruning, the merge gate, the Observe projection). It records scores and decisions; it does not run evaluators, create worktrees, verify Git refs, or merge Git branches. Only the coordinator writes shared state, serially; executors return evidence. The helper has no multi-writer locking. Its JSON schema is separate from upstream Arbor checkpoints.

When to use this skill

Reach for Arbor when the task is iterative improvement of a concrete artifact under an evaluator:

• Model training: optimizer/architecture/recipe changes to lower loss or hit a target in fewer steps.
• Harness/agent engineering: raising pass rate or accuracy of an agent loop, search harness, or tool-use scaffold.
• Data synthesis: improving a generation/filtering pipeline judged by downstream model behavior.
• Benchmark optimization: MLE-bench / Kaggle-style "improve the submission" tasks.
• Prompt/system optimization where you can score outputs automatically.

The distinguishing signals: there's an artifact you can modify, an objective, a way to score candidates, and you expect to run many experiments. If the user only wants a single fix or a one-shot answer, this is overkill — just do the work directly. If they want open-ended ideation with no evaluator, use hypothesis-generation or scientific-brainstorming instead.

The AO setup — pin this down first

Before any experiments, establish the task tuple (M_0, O, E_dev, E_test). Record the resolved setup, using the user's existing instructions; clarify only missing decisions:

• M_0 — initial material: the artifact to improve (a repo, a script, a config, a prompt). Make sure it's under git and currently runs.
• O — objective: the natural-language goal and the metric *direction* (maximize accuracy? minimize loss/steps?).
• E_dev — development evaluator: a command you can run freely during search to score a candidate. Fast, repeatable.
• E_test — held-out test evaluator: a *separate* evaluation protocol (held-out examples or predeclared independent seeds) used only at the merge gate. It must not be used as a search oracle — that's the whole point.

Freeze the evaluator, dataset split, and metric before search, and record their

versions with each score. Repeated merge decisions on the same held-out set can

still adapt the search to that set; a fresh worktree prevents file contamination,

not statistical leakage. Keep a final untouched evaluation set for the final claim,

or disclose that the reported result was selected using repeated gate feedback.

If the user hasn't given you a clean dev/test split, construct one and say so. The dev/test separation is the mechanism that catches overfitting: a candidate that wins on dev but not on test isn't a success, it's a warning that you're exploiting the feedback signal. A larger run on the same examples is not an independent holdout. If no defensible holdout exists, label results as development-only.

Set ARBOR_TREE to the absolute path of this skill's scripts/tree.py, then work from the experiment project directory. The following commands illustrate one run; evaluator commands, scores and refs must be replaced with actual measured evidence. Initialize the run:

python "$ARBOR_TREE" init \
  --objective "Improve BrowseComp answer accuracy on the search harness" \
  --dev-eval "python eval.py --split dev --n 50" \
  --test-eval "python eval.py --split test --n 300" \
  --material "." --metric-direction max --branching 3 --max-depth 2 --budget 12

Evaluate M_0 on the fixed gate protocol before search and record its commit and score:

python "$ARBOR_TREE" baseline --test-score 45.33 --branch-ref "baseline-commit"

The coordinator is responsible for freezing protocols and evaluating the baseline; these are procedural steps, not automated checks. The baseline command only records evidence supplied by the coordinator, once. merge refuses an unscored baseline and non-finite scores. Existing runs with no incumbent need this step; start a new run if the baseline/evaluator changes. Keep raw evaluator output, seeds, split IDs and evaluator version alongside .arbor/.

--branching is how many sibling hypotheses you propose per parent; --max-depth 2 keeps directions at depth 1 and concrete interventions at depth 2 (the paper's default); --budget is the number of coordinator cycles. Start small (10–20 cycles). These helper settings are advisory: branching is recorded, depth overruns warn, and the cycle counter can exceed the budget. The coordinator must enforce the agreed stopping limits.

The coordinator loop

You run repeated cycles of six steps. This is the heart of HTR; do not collapse it into ad-hoc editing. Run python "$ARBOR_TREE" cycle once per cycle to track the budget.

1. Observe

Begin every cycle by re-grounding in the tree, not in your memory of the conversation:

python "$ARBOR_TREE" observe

This prints the objective, global insights, the active frontier (selectable hypotheses), executed nodes with their evidence, pruned lessons (negative constraints), and the current best artifact. Treating the tree as the source of truth is what keeps you coherent over a long run, after context compression has thrown away the details.

2. Ideate

Pick a promising parent and propose a few child hypotheses under it. Condition on the tree's evidence — this is the difference between Arbor and random search:

• Validated insights are assumptions you can build on.
• Pruned nodes are dead ends to avoid.
• A "half-right" result is a *starting point for a sharper hypothesis*, not a reason to abandon the direction.

Each hypothesis should be a falsifiable claim about how changing the artifact will move the metric, not a vague intention. Depth-1 nodes are broad directions ("the search harness loses correct answers it already retrieved"); depth-2 nodes are concrete, executable interventions ("run K=5 independent rollouts and aggregate by evidence dossier instead of majority vote").

python "$ARBOR_TREE" add-node --parent n0 --hypothesis "Verification, not retrieval, is the bottleneck: candidates are found but discarded"
python "$ARBOR_TREE" add-node --parent n1 --hypothesis "Aggregate independent rollouts by evidence dossier to recover minority answers"
python "$ARBOR_TREE" add-node --parent n1 --hypothesis "Search-augment the judge to verify discarded candidates"

3. Select

Choose which pending leaves to run next. Selection is not pure score-maximization — pick a hypothesis because it has strong prior evidence, because it would resolve an ambiguity its siblings exposed, or because its failure would clarify an important assumption. Frontier control under delayed feedback rewards informative experiments, not just promising ones.

4. Dispatch

Run each selected hypothesis as an executor subagent in an isolated worktree (use the host's supported isolation facility, or create distinct worktrees with git worktree add). Isolation matters: parallel experiments must not clobber each other or the current best, and exploratory changes stay quarantined until they pass the merge gate.

Dispatch siblings in parallel (through the host's available subagent tools) when they're independent — comparative evidence within one direction is exactly what makes later pruning and abstraction possible.

Give each executor a tight, hypothesis-bound brief. See references/executor-brief.md for the full template. The contract that makes HTR work: the executor may not change the hypothesis when the metric stalls. It repairs its own code and reruns, but

… (nội dung đầy đủ có trong gói tải về)