Under GDPO, every printed windowed IM configuration and the no-window configuration exceeds the GDPO base in average accuracy on both MATH-500 and Countdown. Increasing sharpness from 0.5 to 2 at window 10 raises Countdown average accuracy from 61.3% to 66.2%, while MATH-500 averages remain nearly flat; removing the window lowers the d=1.5 averages relative to window 10.

Under GDPO, every printed windowed IM configuration and the no-window configuration exceeds the GDPO base in average accuracy on both MATH-500 and Countdown. Increasing sharpness from 0.5 to 2 at window 10 raises Countdown average accuracy from 61.3% to 66.2%, while MATH-500 averages remain nearly flat; removing the window lowers the d=1.5 averages relative to window 10. · CiteArk

Measurement results

0/48 assessed

0 supported0 disputed or conflicting48 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
dataset: MATH-500 · configuration: GDPO (base) · generationLengthTokens: 128	31.2%	—	Not assessed
dataset: MATH-500 · configuration: GDPO (base) · generationLengthTokens: 256	35.4%	—	Not assessed
dataset: MATH-500 · configuration: GDPO (base) · generationLengthTokens: 512	40%	—	Not assessed
dataset: MATH-500 · configuration: GDPO (base) · generationLengthTokens: average	35.5%	—	Not assessed
dataset: Countdown · configuration: GDPO (base) · generationLengthTokens: 128	53.5%	—	Not assessed
dataset: Countdown · configuration: GDPO (base) · generationLengthTokens: 256	62.9%	—	Not assessed
dataset: Countdown · configuration: GDPO (base) · generationLengthTokens: 512	65.2%	—	Not assessed
dataset: Countdown · configuration: GDPO (base) · generationLengthTokens: average	60.6%	—	Not assessed
dataset: MATH-500 · configuration: GDPO+IM d=0.5 w=10 · generationLengthTokens: 128	32.8%	—	Not assessed
dataset: MATH-500 · configuration: GDPO+IM d=0.5 w=10 · generationLengthTokens: 256	38.6%	—	Not assessed
dataset: MATH-500 · configuration: GDPO+IM d=0.5 w=10 · generationLengthTokens: 512	40.4%	—	Not assessed
dataset: MATH-500 · configuration: GDPO+IM d=0.5 w=10 · generationLengthTokens: average	37.3%	—	Not assessed

Reported and observed measurements

Assessments (0)

Runs (0)

Measurement results