SPG+IM outperforms SPG+entropy in every accuracy cell across four tasks and three generation lengths, with the largest gaps on planning tasks. SPG+entropy also underperforms unmodified SPG in every cell, a negative result showing that preferentially masking individually predictable tokens can be worse than uniform random masking.

SPG+IM outperforms SPG+entropy in every accuracy cell across four tasks and three generation lengths, with the largest gaps on planning tasks. SPG+entropy also underperforms unmodified SPG in every cell, a negative result showing that preferentially masking individually predictable tokens can be worse than uniform random masking. · CiteArk

Measurement results

0/36 assessed

0 supported0 disputed or conflicting36 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
model: SPG · dataset: GSM8K · generationLengthTokens: 128	78.5%	—	Not assessed
model: SPG · dataset: GSM8K · generationLengthTokens: 256	84.9%	—	Not assessed
model: SPG · dataset: GSM8K · generationLengthTokens: 512	84.2%	—	Not assessed
model: SPG · dataset: MATH-500 · generationLengthTokens: 128	34.6%	—	Not assessed
model: SPG · dataset: MATH-500 · generationLengthTokens: 256	39.6%	—	Not assessed
model: SPG · dataset: MATH-500 · generationLengthTokens: 512	42.6%	—	Not assessed
model: SPG · dataset: Countdown · generationLengthTokens: 128	68.8%	—	Not assessed
model: SPG · dataset: Countdown · generationLengthTokens: 256	70.7%	—	Not assessed
model: SPG · dataset: Countdown · generationLengthTokens: 512	70.3%	—	Not assessed
model: SPG · dataset: Sudoku · generationLengthTokens: 128	82.9%	—	Not assessed
model: SPG · dataset: Sudoku · generationLengthTokens: 256	94%	—	Not assessed
model: SPG · dataset: Sudoku · generationLengthTokens: 512	93.1%	—	Not assessed

Reported and observed measurements

Assessments (0)

Runs (0)

Measurement results