hyperagents

Rule

HyperAgents’ parent-selection rule, as a HarnessGrad method.

The method returns one edited harness per round. It never sees a score for the harness it is producing — the platform measures every candidate on the same study set and records the result in curve.jsonl.

Measured results

Scores, rounds and per-split comparisons are on the ALFWorld benchmark page — one table for every method, so the same number cannot say two different things in two places.

Record

This method’s rounds on the Hub — each round’s harness, the edit that produced it, and its per-task trajectories: open the directory.

Deviations from the method it ports

3 deliberate difference(s), recorded in methods/hyperagents/method.json:

• Multiple children per call: HG_HYPERAGENTS_CHILDREN defaults to 3, capped by remaining epochs. Parents are sampled from valid measured history using sigmoid(10*(score-mid)) * exp(-(children/8)^3), with mid = mean(top-3 scores). The snapshot window is configurable: HG_METHOD_HISTORY_ROUNDS=0 keeps all states by default; positive N stages only the newest N states (STATES_KEPT = HISTORY_ROUNDS or None).
• The live parent selector reads a single protocol.studied_score value; upstream staged-eval rescaling, per-domain agent/ensemble max and multi-domain averaging are not applied. staged_eval.py records scheduling metadata with score_rescaling=none (run.py passes staged_fraction=1.0). archive.py:ArchiveNode.score implements agent/ensemble max, but run.py does not use that helper for parent scoring.
• HG_HYPERAGENTS_SELECT is read once at module import, defaulting to score_child_prop. It selects fixed local weight branches: score_child_prop, score_prop, uniform/random or best; it does not execute or evolve upstream select_next_parent.py. Unrecognized values (including latest) fall back to score_child_prop weights while the report retains the requested setting.

Code

methods/hyperagents/run.py and its rule test in tests/quality/. python3 tools/method_lint.py prints one line per method.