Text set in serif type with a dotted bar in the margin was drafted by an LLM (claude-opus-5) and sometimes reviewed by the author. The rest is the author's own. How to read this site
alpha-curriculum → schedule-opt → schedopt-v4
schedopt-v4
investigation
Intent
The first three rounds of this campaign each froze an objective, searched a parameterized family of curricula, and defended a winner. This one drops the family. A candidate is K blocks of α level and duration, with K drawn rather than fixed and the boundaries free integers, so the search space contains constant schedules, monotone descents and oscillations alike as special cases, and contains shapes none of the earlier families could express.
The reason to spend a round this way is that every family used so far was chosen after the fact to contain the arms that were already winning, which makes a search inside it evidence about the family rather than about the space. The node also carries the instrument change that makes the numbers here incomparable to the earlier rounds: a crossing is confirmed by re-measuring frozen weights against fresh minimal pairs rather than by waiting for a second waypoint.
Conclusions
One experiment has landed. Its reading is that the family question mattered more than the objective question: the fastest schedule found is a single step, simpler than anything the earlier families were built around, and the property that predicts crossing speed is the opening level of α rather than the amount of movement.
The round closed with its winner's advantage unseparated from the selection that found it, since the seeds that ranked it were the seeds the comparison used. That is now answered. On 150 seeds that took no part in choosing it, the same schedule reaches the band at 202.1 steps against 198.9 in the search, and it finishes first of seventeen arms, ahead of everything the following round proposed (the held-out standings). The selection cost about three steps; the schedule kept the rest. This synthesis is rewritten as further members land.
Experiments
-
20260726_025145_schedopt_v4 2026-07-26
Run 2026-07-25 to 2026-07-26 (about ten hours on a single GTX 1660 Ti; 400 sampled curricula plus 26 reference arms, 5,024 seed-runs against 24,000 for the same candidates at a flat budget).
This round asks one question directly: which schedule of α reaches good-enough generalization soonest, over seeds. It is built for speed of iteration rather than for defending an answer, so it carries no frozen operating point, no pre-registration, no gates, and no held-out discipline. Those belong to the confirmation that follows a winner, not to the search that finds one.
The design choice that matters is what a candidate is allowed to be. Earlier rounds searched inside a parameterized family, and the family turned out to exclude the schedule that was already winning. Here a candidate is K blocks of level and duration over the first 300 steps, with K drawn rather than fixed, block boundaries free integers, and levels free in the interval 0 to 5. Four hundred such curricula were drawn at random and run down a successive-halving ladder over the seed budget, 6 to 20 to 60, promoting the fastest third at each step. Nothing about the shape is proposed by a model in this round; the draws are unbiased, and the point is to find out what the space contains before assuming anything about it.