12.29σ Arrangement Effect in Conflicting Supervision Vanishes Under Agnostic Scoring
The study proves arrangement effects under conflicting conventions are exactly zero under summed-accuracy metrics. Constant schedules expose large allocation shifts while decayed schedules suppress them. Total capability remains conserved.
The arXiv paper isolates arrangement as a block-period factor multiplied by schedule weighting. Fixed-budget runs across ten paths show constant-rate schedules produce allocation shares from 0.04 to 0.87 with 0.2221 interior span and 11.63 contrast floors. Cosine decay restricts the same arms to two distinguishable states. Marking the convention in the prompt collapses the switch and reaches 87.5 percent of the union ceiling.
Acc_A plus Acc_B stays constant within 9.7 percent across all arms, confirming the measured effect is commitment, not capability. The bound separates arrangement from schedule: decay prevents large step size and uncontracted remainder at the same step. This mechanism accounts for earlier reports that decaying schedules moderate ordering effects in pretraining.
The result reframes conflicting supervision as a knob between order-sensitive and order-insensitive regimes. Exact-match benchmarks miss the allocation shift because they aggregate conventions. Researchers must therefore adopt convention-agnostic metrics when corpora contain incompatible labelings.
Operational follow-up requires testing the multiplication factor across scheduler families on larger models. Constant-rate runs should be repeated with explicit convention prompts to measure residual variance after collapse.
Chen et al.: Constant-rate schedules will exhibit allocation variance exceeding 0.15 in 80 percent of conflicting-convention corpora tested within 18 months.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2610.00234)
- [2]Supporting Source(https://arxiv.org/abs/2305.14314)