TOST
A standard statistical method for equivalence testing that transforms the equivalence question into two one-sided tests; an equivalence conclusion can be drawn only when both one-sided tests reject the null hypothesis.
Timeline (last 90 days)
After preference alignment on Qwen3.5-4B using length-normalized DPO plus an SFT retention term (the RPO objective), the performance change on the structured extraction task was -0.55 percentage points, passing the TOST equivalence test within the ±5 percentage point equivalence interval.
This study is a pre-registered study using 8 random seeds (6 of which were unseen at freeze time) and employing artifact-aware scoring and TOST equivalence tests.
The paper uses Two One-Sided Tests (TOST) to elevate the indistinguishability finding to an equivalence claim, which holds across all 12 Qwen/LLaMA experimental cells.
TOST is a hypothesis testing framework in statistics used to establish equivalence rather than difference, asserting equivalence between two groups by setting equivalence bounds and simultaneously testing two one-sided hypotheses.
All Facts (4)
After preference alignment on Qwen3.5-4B using length-normalized DPO plus an SFT retention term (the RPO objective), the performance change on the structured extraction task was -0.55 percentage points, passing the TOST equivalence test within the ±5 percentage point equivalence interval.
50%UnverifiedThis study is a pre-registered study using 8 random seeds (6 of which were unseen at freeze time) and employing artifact-aware scoring and TOST equivalence tests.
50%UnverifiedThe paper uses Two One-Sided Tests (TOST) to elevate the indistinguishability finding to an equivalence claim, which holds across all 12 Qwen/LLaMA experimental cells.
50%UnverifiedTOST is a hypothesis testing framework in statistics used to establish equivalence rather than difference, asserting equivalence between two groups by setting equivalence bounds and simultaneously testing two one-sided hypotheses.
50%