SFT reliably teaches the two output contracts but cannot raise interactive-HTML quality, which declines monotonically with update count¶

Claim · All claims
Evidence · 1 study · design 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? Under the pre-hardening reward, HTML quality declines monotonically with update count on Qwen3.6-27B (88.4 at 6,710 updates, 83.7 at 10,711, ~73.5 at 13,421), and the Qwen3.8 run lands at 73.2, below its own bare base of 78.3. → CogEvol Team 2026

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 · r3

SFT ablation runs (Table 2 and text), scored under the pre-hardening reward and compared within that period only; the Qwen3.8 run ceded ground where the base was strongest (code −7.0pp, game −17.9pp).

"HTML quality declines monotonically with update count on Qwen3.6-27B—88.4 at 6,710 updates, 83.7 at 10,711,∼73.5 at 13,421—whether the extra updates come from more passes or a smaller batch;"

Discussion¶

Related Claims¶

• 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 — related
• In directional manual testing, reward hardening halved unusable interactive pages (25%→10%) and eliminated page-entry failures (2/24→0/30) — 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