dream rsi
A round-score proxy inspired by Dream-RSI’s policy selection.
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
Section titled “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
Section titled “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
Section titled “Deviations from the method it ports”8 deliberate difference(s), recorded in methods/dream_rsi/method.json:
- The replay simulator is absent.
- Each round’s studied score replaces the paper’s normalized Pareto hypervolume; no subtree max and no average over replay worlds is formed.
- The priced quantity N is the request count (Eq. 1), not kilo-tokens; beta1/beta2 stay unspecified (neither paper version publishes a value) and kilo-tokens are still recorded
- The denominator k is n_trials rather than completed replay decision rounds, and with b2=0.0 the port carries no parallelism term while v2 penalizes serial policies.
- The offline M-revision phase is absent: every revision is paid for.
- The proposer edits a harness rather than a dedicated executable exploration policy with the shared online/replay batch interface.
- Argmax is over recorded round proxies; the arithmetic check does not establish the paper’s replay or deployment guarantee.
- The port implements v1’s linear Eq. (1), which v2 superseded with the appendix’s Pareto objective.
methods/dream_rsi/run.py and its rule test in tests/quality/. python3 tools/method_lint.py prints one line per method.