On the report's internal suites CogEvol-27B scores 83.7 on slides and 63.7 on HTML-500 with zero hard failures, while no external flagship leads both modalities¶

Claim · All claims
Evidence · 1 study (2 entries) · design r2–r3 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Qwen3.8-Max ties CogEvol-27B's 83.7 slide score only with the full 34 KB specification, but collapses to 35.3 on HTML-500 with 204 of 500 pages dead at the probe. → CogEvol Team 2026 q2 i? Claude Opus 4.8 posts the best external HTML average (67.2 vs 63.7) while failing 19 pages outright on dead interactions; GLM-5.3 has 115 of 500 pages dead. → CogEvol Team 2026 (2)

Evidence¶

CogEvol Team 2026¶

CogEvol Team, CogEvol Inc. & Tsinghua University. (2026). CogEvol: Towards Efficient and Reliable Learning Environment Generation. arXiv:2608.30968v2. https://arxiv.org/abs/2608.30968 doi:10.48550/arxiv.2608.30968

q2 · i? · design · r2

Main results (Table 3 protocol): external flagships run the identical harness with slide columns under the generous 34 KB specification; Qwen3.8-Max scores 23.9 zero-shot on the slim contract.

"the same model collapses on the other modality: 35.3 on HTML-500 with 204 of 500 pages dead at the probe, where CogEvol-27B scores 63.7 with zero hard failures."

CogEvol Team 2026 (2)¶

CogEvol Team, CogEvol Inc. & Tsinghua University. (2026). CogEvol: Towards Efficient and Reliable Learning Environment Generation. arXiv:2608.30968v2. https://arxiv.org/abs/2608.30968

q2 · i? · design · r3

Main results table reading: the report concludes "no model leads both modalities under either condition"; GLM-5.3 shows the same fluent-but-broken split with 115 of 500 pages dead.

"Claude Opus 4.8 posts the best external HTML average (67.2 vs. 63.7, leading four of six sub-types) while failing 19 pages outright on dead interactions, and GPT-5.4 (66.0) fails 13;"

Discussion¶

Related Claims¶

• SFT reliably teaches the two output contracts but cannot raise interactive-HTML quality, which declines monotonically with update count — related
• Purpose-trained single-pass generation completes a slide in a median of 17 seconds and an interactive page in 59 seconds over 220k production requests, and delivers artifacts at 15–22× lower per-artifact API cost than flagship models — related
• A screenshot-only reward was reward-hacked into producing visually convincing but unplayable games; hardening the reward with an interactivity probe and hard-fail gate reversed the collapse — related
• In directional manual testing, reward hardening halved unusable interactive pages (25%→10%) and eliminated page-entry failures (2/24→0/30) — related
• Anchored judge prompts restore score discrimination and raise the reward's agreement with independent human ratings (Spearman ρ 0.672→0.741) — related