{"at":"2026-10-15T18:12:23.908Z","question":"Can a popular small Qwen model improve responsive UI design through richer visual rewards and longer online training?","briefs":48,"seeds":[1701,2718],"families":["booking","settings","pricing","search","checkout","comparison"],"generatedCandidates":576,"pairwise":[{"control":"sft","n":96,"wins":84,"ties":1,"losses":11,"winRate":0.8802083333333334,"ci":[0.8020833333333334,0.9479166666666666]},{"control":"base","n":96,"wins":65,"ties":0,"losses":31,"winRate":0.6770833333333334,"ci":[0.5729166666666666,0.7708333333333334]},{"control":"best","n":96,"wins":46,"ties":0,"losses":50,"winRate":0.4791666666666667,"ci":[0.375,0.5833333333333334]}],"methods":[{"method":"base","n":96,"complete":76,"coverage":0.7859126984126982,"geometry":0.8034046468829078,"visual":0.576388888888889,"responsive":0.4388020833333333,"outputTokens":3090.2291666666665,"inferenceCandidates":1},{"method":"sft","n":96,"complete":52,"coverage":0.5259027777777777,"geometry":0.520273333437219,"visual":0.34548611111111116,"responsive":0.2604166666666667,"outputTokens":3388.71875,"inferenceCandidates":1},{"method":"rl","n":96,"complete":95,"coverage":0.9676240079365076,"geometry":0.9832894316730524,"visual":0.7000868055555557,"responsive":0.5911458333333334,"outputTokens":2239.7708333333335,"inferenceCandidates":1},{"method":"best","n":96,"complete":91,"coverage":0.935567956349206,"geometry":0.9505027090964591,"visual":0.7083333333333331,"responsive":0.59375,"outputTokens":3099.21875,"inferenceCandidates":4}],"limitations":["Synthetic test with six held-out task families.","Two samples per brief, confidence intervals resample briefs, not individual generations.","Training and blind evaluation use the same model family, different prompts. No human preference study.","Best-of-four selection spends four generations plus visual ranking per delivered UI.","Qwen already has general coding and UI knowledge; this is adaptation, not learning design from scratch. Scores use a richer rubric than the earlier AesCode run."]}