On WebShop, rebuilding the training bank after 100 GRPO steps and training 50 more raises success by 2.8 points for Qwen3-8B and 0.9 for GPT-5.4 but leaves Gemini-2.5-Pro unchanged; warm-starting the test bank with 100 training trajectories changes success by at most 1.3 points and remains within standard deviation.

On WebShop, rebuilding the training bank after 100 GRPO steps and training 50 more raises success by 2.8 points for Qwen3-8B and 0.9 for GPT-5.4 but leaves Gemini-2.5-Pro unchanged; warm-starting the test bank with 100 training trajectories changes success by at most 1.3 points and remains within standard deviation. · CiteArk

Measurement results

0/18 assessed

0 supported0 disputed or conflicting18 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.
Reported	Latest observation	Evidence judgment
61.1 points	—	Not assessed
32.8 percentage points	—	Not assessed
61.7 points	—	Not assessed
35.6 percentage points	—	Not assessed
60.9 points	—	Not assessed
32.5 percentage points	—	Not assessed
61 points	—	Not assessed
50.5 percentage points	—	Not assessed
61 points	—	Not assessed
50.5 percentage points	—	Not assessed
61.7 points	—	Not assessed
50.9 percentage points	—	Not assessed

Reported and observed measurements

Assessments (0)

Runs (0)

Measurement results